Earlier in the week, I was at an AI × Fintech conference, surrounded by product leads, engineers, and builders debating how to push boundaries and where the next wave of AI innovation will come from. Later that week, I was at a finance symposium & gala dinner, in a room full of investment bankers, PE professionals, LPs, and venture capitalists discussing leadership, mentorship, and the future of the industry.

At the AI conference, I found myself explaining what private credit actually is, how capital gets raised and deployed, and where the finance industry’s pain points lie – all based on my years on Wall Street. At the finance symposium, I was explaining the capabilities and limits of AI, drawing from my work through The Bull and The Bot: what today’s models can actually do, how workflows could evolve, and where “human in the loop” is still needed.

So in one room, I was translating finance to the AI crowd. In the other, I was translating AI to finance.

The contrast between the two was striking. The AI crowd understood the technology deeply and engaged thoughtfully with finance, but often from a data-first perspective rather than the lived workflow reality that drives decisions inside firms. The finance crowd, meanwhile, understood markets intuitively and was curious about AI, but hadn’t yet found the mindset or structure to practically apply it across workflows. Many felt that the technology wasn’t quite there yet, or that the costs simply didn’t seem to outweigh the benefits.

And that made me think. While finance firms are interested and investing in AI, the measure for success still keeps coming back to one question: are we actually seeing ROI? As one finance attendee put it, “I get that people are more efficient using AI, but where are the ROI numbers in dollar form?”

The ROI Paradox

Now, ‘what’s the dollar return on my investment?’ is the most natural question for any finance professional to ask, because it’s how we’re wired. But the more I engaged in conversations across these two conferences, the more it became clear to me that ROI in AI isn’t missing -it’s just being measured in the wrong units right now.

Because the reality is, some returns show up first as edge: in how fast you learn, how well you decide, or how effectively your teams adapt.

What that means is, not every AI return can, or should, be measured in dollars.

Instead, I think about ROI in three levels:

1. Dollar value: purely financial - revenue up, cost down.

o This is what leaders fixate on (“show me the dollars”).

o This is lagging ROI, not leading ROI.

2. Numerical value: quantifiable but not dollarized.

o Productivity metrics, time saved, number of deals screened, user adoption, etc.

o This can be measured numerically - but doesn’t translate neatly into P&L yet.

o It’s “ROI you can count, but not cash.”

3. Intangible value: cultural, behavioral, reputational.

o Curiosity, collaboration, experimentation, power-user influence.

o These are the compounding drivers of all future ROI, but they’re invisible on balance sheets.

Early-stage ROI in AI is numerical and intangible, not financial. It’s measured in learning velocity, decision leverage, and cultural momentum – things that don’t have price tags yet but eventually shape the bottom line.

And this idea is echoed by something Lloyd Blankfein (former CEO of Goldman Sachs) said during his keynote at the AI x Fintech conference:

“In trading, if you want to win a bid in a market, the person who’s one millisecond faster – whose machines are half a block closer to the exchange – wins 100% of the time. So how could you not invest in the technology that gets you closer? The same applies to people exercising judgment over investment decisions. You want the best tools and resources to at least be able to have an arms race and try to stay ahead.”

Firms keep asking for ROI in dollar form, but what Lloyd was describing is the ROI of edge. It’s not about whether AI has “paid off” yet – it’s about whether you’re compounding your learning faster than the competition. The edge doesn’t have to be big, it just has to exist. And over time, those milliseconds of advantage – in knowledge, workflow, or insight – become the gap between staying relevant and falling behind. Edge matters, no matter how minute it seems at the present.

And that’s the paradox of AI ROI in this moment: firms want to see the dollars now, when the real returns show up in edges that only compound and become obvious later.

Invisible Alpha

Lets take the idea of ‘Decision Leverage’ as example.

Every established firm already has decades of proprietary data – investment memos, analyses, deal notes – all of which are a latent form of alpha that no human team can fully digest.

AI becomes the tool that translates that buried knowledge into leverage and active intelligence. That’s a massive form of ROI that doesn’t appear on P&L. Rather, it appears in sharper decisions. By surfacing patterns humans couldn’t possibly see, AI fills in blind spots we didn’t know existed. The edge AI offers isn’t in replacing judgment but is in amplifying it.

For example, lets say a team uses AI and can now review 10 deals instead of 5, and five years later they outperform because one of those extra 5 turned into a winning investment. That’s not visible in quarterly ROI reviews because how do you quantify a better decision you never would’ve made without AI? You can’t measure the value of a deal you never reviewed, because you wouldn’t have even known it existed. And to that end, you can’t P&L the AI-enabled insight that prevented a bad deal or surfaced a good one. So the real value of AI here is counterfactual: its invisible in the short term, but defining over time.

Thus “Decision Leverage” is intangible in quality but numerically observable in behavior. You may see it in the data trail of better decisions being made faster (e.g. more opportunities screened, higher-quality ICs, fewer reversals, faster approval cycles, etc.). But you can’t yet price it in dollar terms because the causal chain to P&L is too long or complex.

ROI in AI isn’t just about cost reduction or revenue gains – it’s about the numerical and intangible values that include things like decision leverage. Are teams accessing more of the proprietary data they own to enhance their insights? Are they making decisions faster, with greater confidence? Are they compounding insight faster than competitors? Those are the real return metrics, but they just don’t fit neatly into a spreadsheet. So while leaders may be measuring for visible gains, the real ROI right now sits in invisible advantage.

Translating Leadership for the AI Era

But to see that invisible advantage, leadership itself has to evolve.

At the finance symposium, Kewsong Lee (Former CEO of The Carlyle Group) shared a growth framework that stuck with me: Think Big. Move Quick. Perform Better. Sitting in the audience, I couldn’t help but think about how directly it applies to the kind of leadership needed to guide organizations through AI adoption today.

Thinking big means expanding perspective: lifting your focus from the immediate efficiencies of AI to the broader transformation it enables. It’s the shift from asking “Where can we cut cost?” to “How do we make better decisions?” or “How do we create new kinds of value?” For AI, thinking big means seeing the entire system – data, workflows, people – as part of a continuous learning loop. If you keep measuring AI through the same dollar-cost ROI lens, and you’ll keep drawing the same conclusion: that the returns aren’t good enough. When in reality, it may be the lens that’s too narrow.

Moving quick is about rhythm and readiness. True leadership isn’t just defined by the big, high-stakes decisions, but also by the ability to make smaller ones frequently and appropriately. And that’s exactly what successful AI adoption looks like. You don’t wait for perfect certainty; you test, learn, and adjust in motion. Organizational adoption doesn’t happen through strategy decks, it happens through iteration.

And the real signal that this is happening? Power users.

The focus right now shouldn’t be on measuring AI returns in dollar form. Instead, it should be on tracking how many power users you’re creating inside a firm. Power users are the ones compounding learning inside your organization. They experiment, share prompts, and make adoption contagious. They’re the internal translators turning curiosity into capability. The intangible value they create – a forward-looking, curious, agile, creative culture – is the early ROI every firm should be optimizing for. That’s the real multiplier effect.

Performing better is where conviction and resilience come in. There’s no perfect playbook for integrating AI into established systems, and perhaps there shouldn’t be. The leaders who tend to stand out are the ones who form clear opinions about where AI truly adds value, and who have the conviction to test those opinions in the open.

It’s less about reckless risk-taking and more about ownership of perspective. Because in a landscape moving this fast, describing the trend isn’t enough; what separates strong leadership is the willingness to interpret it — and to stand by that interpretation, test it, and learn from it.

Resilience plays a part, too. Not in the motivational-poster sense, but in the operational one. The cadence of experimentation will always include failures, and performing better often comes down to recovering faster, learning from each iteration, and refining conviction through data and experience.

Conclusion

The question isn’t whether AI can prove its ROI. It’s whether we can learn to translate it.

The firms that will win aren’t the ones waiting for perfect dollar metrics – they’re the ones investing through the translation phase, building edge in how they learn, decide, and lead.

The signals are already there: rising numbers of power users, faster learning loops, and enhancements in decision leverage, learning velocity, and cultural momentum. These are the numerical and intangible metrics that define early-stage ROI. They may not fit neatly into a spreadsheet, but they shape the ones that will.

Think big. Move quick. Perform better. Because when curiosity compounds, so does adoption. And when adoption compounds, ROI follows.

Share

Over the last few weeks, I used Make.com to build a multi-agent workflow that creates, logs, and emails me daily Substack Notes ideas. The system runs across five agents, each with a specific role.

Before we dive deeper, here’s a quick visual of the workflow’s architecture as reference:

When I started building this Notes generator, I quickly realized that the challenge wasn’t so much in creating the agents themselves. Rather, it was in orchestration: deciding how much each agent should handle, how they should interact, and how often they should run. And it was that last piece – cadence – that shaped not just how well the system worked, but also how much it cost to run the workflow.

Two of the earliest agents I built illustrate this idea clearly:

Agent 1: Style Analyzer. It ingests a Master Text File (“MTF”) of all my long-form Substack posts to date and makes API calls to GPT-5 to generate a style guide that captures my voice, thought process, and content focus. That style guide then feeds into Agent 3, which also uses GPT-5 through API calls to synthesize the inputs and generate new ideas.

Agent 2: News Sourcer. It runs two HTTP modules inside Make.com to pull fresh AI and finance headlines via NewsAPI calls, organizes them, and passes them to Agent 3 as inputs. Agent 2 runs every day, since news changes daily and I wanted the workflow to produce Notes options that refer to real-time occurrences.

At first, it seemed logical – and architecturally simple – to run both agents daily, since Agent 3 depends on inputs from both Agent 1 (style guide) and Agent 2 (news). But I quickly realized that Agent 1 and Agent 2 each work with very different types of data - and that difference created the first real design-cost tradeoff.

The Problem: Misaligned Cadence = Wasted Cost

Agent 2’s inputs change constantly, so a daily trigger made sense. However, Agent 1’s source file, the MTF, only updates when I publish a new Substack post, which is usually two or three times a month.

Because I chose to run Agent 1 on GPT-5 latest, every run meant making API calls to a premium model. But, unless I was publishing new posts everyday, running Agent 1 daily like Agent 2 meant repeatedly paying GPT-5 rates to re-analyze the same archive over and over again. In other words, I’d be incurring costs without adding any value most of the time.

The Solution: Align Cadence With Data Freshness

The fix was to pull Agent 1 out of the daily loop and give it its own scenario. Instead of running every day like the rest of the workflow, it now runs twice a month — once in the middle and once at the end. Each run processes the full archive of my work (including any new Substack posts since the last run), generates an updated style guide for Agent 3, and saves it automatically to Dropbox. Then, Agent 3 simply refers to the latest style guide in Dropbox whenever it runs. This design change kept the workflow comprehensive but eliminated the wasted expense of daily Agent 1 runs.

The cost impact showed up clearly in my usage dashboard. On a baseline day (September 13, shown below as example) with the daily workflow running without Agent 1, the system processes about ~6,000 tokens at a cost of $0.04, driven by Agent 3’s GPT-5 calls.

Tokens:

Cost:

On a day when Agent 1 also runs (September 14), usage spikes to ~22,000 tokens and the cost doubles to $0.08

Tokens:

Cost:

Running Agent 1 nearly quadruples token usage for that day and adds about four cents to the workflow’s total cost.

When you extend that out over a month, the picture becomes clearer:

For Agent 1 alone:

• Daily cadence: $0.04 × 30 = $1.20

• Bi-monthly cadence: $0.04 × 2 = $0.08

• → ~15x savings on Agent 1’s cost by implementing bi-monthly cadence

For the total workflow:

• Daily cadence: $0.08 × 30 = $2.40

• Bi-monthly cadence: ($0.04 × 28) + ($0.08 × 2) = $1.28

• → ~50% savings overall by implementing bi-monthly cadence

In dollar terms, the difference may feel small today. But these runs are based on the current size of the MTF, which will only grow as I keep publishing long-form posts. The larger the MTF gets, the more tokens Agent 1 will need to process — and the more expensive those runs will become.

This is why the small architectural shift mattered. It highlighted a broader principle: to optimize cost, execution frequency should match how often the underlying data actually changes.

The Contrast: Zero Cost = Architectural Freedom

While Agent 1 showed me how cadence and cost are tightly linked, Agent 2 highlighted the flip side: when resources are free, you can afford to optimize for quality rather than efficiency.

In my workflow, Agent 2 uses HTTP modules in Make.com to call the NewsAPI, and at my workflow’s scale, those calls are free. That meant I didn’t have to design around cost, but around completeness.

At first, I had Agent 2 use a single HTTP module to fetch a mixed set of 10 articles, which often skewed the results: some days it returned all AI headlines, other days all finance.

Because the API calls carried no cost, the better solution was to split the request into two HTTP modules (two separate NewsAPI calls), both operated by Agent 2: one for AI news, one for finance. Running them in parallel guaranteed balanced coverage every day, giving Agent 3 a stronger input set to work from.

And so the design philosophy Agent 2 taught me was simple: when cost isn’t a constraint, use that freedom to optimize for quality.

Beyond My Workflow: The Cost Lesson

While in my workflow the savings only added up to a few cents, at enterprise scale the same cadence principle determines whether costs stay contained or spiral out of control.

Agent 1 showed me the discipline side of the equation: when resources are expensive, cadence becomes the lever to control spend. Agent 2 highlighted the opposite: when resources are free, cadence gives you the freedom to optimize for quality and breadth. Together, they frame the spectrum of choices every AI agent workflow designer faces.

The same dynamic applies when evaluating AI solution vendors or building internal tools. A vendor might cut costs by using a cheaper model, little reasoning, caching results or reducing refresh frequency. An internal team might over-optimize for efficiency and risk stale outputs, or ignore costs and build infrastructure that won’t scale.

The point isn’t that these trade-offs always happen, but that they could. And they’re easy to miss if you only ask whether the system “works.”

That’s the real lesson. Cost isn’t just a budget line, it’s a design constraint that shapes orchestration. Sometimes it forces restraint, sometimes it grants freedom. The key is knowing which situation you’re in and architecting accordingly.

Share

No, I don’t. But I do know how to use AI really well – well enough to build systems that adapt to me and solve my needs.

Over the past few weeks, I built a multi-agent workflow that emails me five Substack Notes inspirations every morning at 8 a.m., tailored to my style and focus areas. I built this system for two reasons: 1) to give myself a daily reminder to post consistently, and 2) to get fresh perspectives from my AI agents that can either support or challenge my own thoughts on various issues. Sometimes I'll adapt and explore one of its suggested angles or use an article it surfaces as a starting point. Other times, the ideas simply nudge me in a different direction, and I end up writing a note from scratch. In every case, this workflow helps me move faster and with more consistency.

My workflow orchestrates five different agents:

• Agent 1 analyzes my long-form writing style and content (Substack Posts),

• Agent 2 pulls fresh AI and finance news every day,

• Agent 3 combines the output from the two to generate idea inspirations,

• Agent 4 packages the options and sends me an email each morning,

• and Agent 5 logs the option & action I decide to pursue in a spreadsheet.

Designing and wiring these agents together took plenty of trial, error, and iteration. After weeks of building, debugging, and refining, I now have a fully functioning workflow that’s become an indispensable part of my everyday routine.

Here’s a quick demo of how the workflow actually works in practice. I’ve set it so that the workflow is auto-triggered every morning at 8 am, but for this video I manually triggered it mid-day so you can see how it works in real time.

First, What is an Agentic Workflow?

When it comes to modern-day AI, for many people, especially non-technical beginners, AI simply means a smart chatbot like ChatGPT that spits out answers to questions.

In reality, AI has many subfields: machine learning, computer vision, natural language processing, robotics, and knowledge representation and reasoning, among others. These are the research domains that drive progress.

But when you move from the research lab to the real world, it’s more useful to think in terms of functional categories. These categories cut across subfields and describe how we actually experience AI day-to-day.

Among these categories, two dominate today: deterministic automation and generative AI.

• Deterministic automation refers to rule-driven workflows that execute steps (if x → then y), fetch data, log outputs, and enforce rules. Think of it as the rails and infrastructure that keep things moving.

• Generative AI refers to systems that create new text, images, audio, or code. Think of it as the creative engine that makes new content possible. These are powerful but probabilistic, so they need guardrails.

The most interesting frontier is the convergence of the two, and in its most advanced form (as we think of it today), becomes agentic AI. These are systems at the intersection of creation+guardrails+execution that can plan/route actions, use tools, and leverage memory/state to adapt over time, typically with a human-in-the-loop for oversight.

So, What Did I Actually Build?

My multi-agent workflow relies on five agents: two handle generative work (style analysis and idea synthesis) and three handle deterministic automation (news sourcing, email delivery, and choice logging). Together, they form a reliable daily system that ships five tailored ideas at 8 a.m. without fail. I call this agentic-lite: a working proof-of-concept that shows how generative AI and deterministic automation can be combined into a structured workflow that runs on schedule, produces consistent outputs, and builds the foundation for learning over time.

Pairing generative AI with deterministic automation only works when roles and handoffs are explicit: who creates, who packages, who remembers, and when the human steps in. Each of my agents has one specific job, and I, as the human-in-the-loop, review, refine, or reject the outputs as the last checkpoint.

Why go multi-agent instead of asking one agent to do everything? Because splitting work across multiple agents is often more reliable and cost-effective than forcing one giant agent to do it all.

When you use multiple AI agents, you can break down complex problems into focused units of work or knowledge, and each one becomes easier to handle. Microsoft frames the advantages of using multi-agents as specialization, scalability, maintainability, and optimization. In practice, I’ve seen all four in my own system: each agent stays focused on one task, I can add or swap agents without touching the rest, debugging is localized and simple, and I can choose the cheapest method or most efficient model for each job.

The principle is similar to how organizations are structured. You would never expect one person to handle product development, engineering, finance, marketing, sales, and customer success all at once. Teams are divided by function, and there is usually a central leader (often a product manager) who aligns them on vision, timelines, and process so the product actually ships.

My system works the same way: the agents each own a task, the workflow provides the rails, and I serve as the central coordinator who reviews the output, performs quality control and ensures the system’s results match my needs. That oversight also generates training data that will eventually feed back into Agent 3 for smarter idea generation.

Takeaways

Over the next few posts, I’ll share deep dives on each of my agents, the orchestration mechanics, design-tradeoffs, failure modes, and my Phase 2 plans for implementing automated feedback loops. This will help the system will learn over time, as I work with Agent 5 and Agent 3 to make the workflow more adaptive and agentic.

For now, here are three key takeaways to keep in mind as you learn and work with agent workflows.

1. You Don’t Need Code, You Need Clarity

Most people assume AI proficiency requires coding. It doesn’t. Knowing how to code is useful – it gives you finer control and can expand what’s possible. But it’s only a fraction of the AI story.

What it really requires is clarity. Clarity about what “good” looks like, how models behave, where the human should step in, and what you want the system to remember or learn.

Thus the foundation of working with AI is threefold:

1. Understanding AI’s capabilities and limits

2. Shaping a clear vision for how it should work for you

3. Moving that vision into execution with the no-code tools now available

Unlocking AI’s potential doesn’t require coding – it’s accessible to anyone with vision and determination.

2. Systems Can Now Adapt to You – So Make Them

For most of history, the majority of us adjusted ourselves to fit into someone else’s system. At work, that meant learning an established workflow, memorizing steps, and trying to optimize within a predefined structure. In tech, it meant downloading an app someone else made and figuring out how to use it the way it was designed.

AI flips that dynamic. Now, if the way you work is different, you can architect the system around yourself: describe the workflow, set the rules, and let the machine run it.

Exploring new AI apps is valuable: it helps you see what’s possible and teaches you how to handle AI’s quirks, capabilities, and limits. But the real gains don’t come from app-hopping. They come from designing your own workflows, using AI tools to stitch together agents that work the way you want. By building agent workflows that fit you, your preferences become features and your process becomes a product. Customizability is AI’s greatest strength.

3. Orchestration Matters More Than Agent Design

When people think about multi-agent workflows, they usually focus on the agent itself: which model to use, what prompt to write, what personality to assign. But the real value lies in the spaces between agents.

Orchestration is about how agents connect: sequencing, handoffs, error handling, where the human-in-the-loop sits, and how the system learns and adapts over time. That’s what I found mattered most in my own build. The challenge wasn’t so much about making a single agent smarter. Rather, it was in appropriately designing the spaces where they meet.

And Microsoft’s agent orchestration framework supports this point. Whether it is sequential pipelines for reliability, concurrent inputs for parallel perspectives, group chat for debate, dynamic handoffs for routing, or manager-led planning for open-ended problems, every pattern underscores the same lesson: collaboration architecture defines success.

Think of agent design as writing a job description. Orchestration is designing how the team works together. Without it, agents drop information, loop endlessly, or talk past each other. With it, the workflow runs stable, reliable, and aligned with intent.

Share

1) AI × Wall Street

• Banks triple AI use cases: Evident’s new Brief shows use cases tripled at top banks in H1’25, with the heaviest deployment activity concentrated in retail / private banking. Most announcements do not include ROI though, so performance can’t be inferred from public data yet. The push now shifts toward scaling genAI and “agentic” tools across others parts of banks.

• Seth Klarman thinks AI is intern-tier: Baupost’s CEO calls AI “a capable assistant,” helpful for tabulating data, scanning 10-Ks, and quick visual ID work (e.g., logos), but is not a stock picker. He warns over-reliance on AI can dull one’s creativity and critical thinking, reinforcing that human judgment should anchor investment decisions.

• Google Finance gets an AI makeover. Google is testing an in-page chatbot, advanced charting, and a live news feed in Finance so that so users can ask complex market questions without leaving the platform.

TL;DR - Adoption is clearly accelerating inside existing finance workflows, though public ROI remains largely undisclosed. Meanwhile, Google is closing the “AI escape hatch” by adding chatbot-style answers to Finance, giving users fewer reasons to jump out of its platforms and to other assistants for complex queries.

2) Platform Firm Moves That Matter

• OpenAI’s GPT-5 Landed Roughly: GPT-5’s launch sparked pushback over errors, a “colder” tone, and restrictive usage limits. OpenAI responded by promising fixes, including warming the model’s personality, temporarily restoring GPT-4 access, and adding more customization controls - all while facing increasing competition and compute strain.

• Claude Sonnet 4 now supports a 1M-token context (public beta): Anthropic’s API customers can now use Claude Sonnet 4 with a 1 million-token context window - the equivalent of multiple books, dozens of research papers, or a large codebase. This means users can load massive documents and keep asking questions across long conversations without the model losing context.

• Oracle ↔ Google Cloud, Gemini models become available natively on OCI: Oracle clients can now switch on Gemini directly within Oracle Cloud, no extra setup required. That convenience makes Gemini the easier choice for enterprises already on Oracle, giving Google a distribution edge.

TL;DR – OpenAI stumbled with GPT-5’s rollout, facing user backlash and scrambling to patch tone and usability issues. Anthropic is leaning into differentiation with a 1M-token context, making Claude useful for book- or codebase-scale work. And Google scored a distribution win by making Gemini available natively through Oracle Cloud, lowering adoption barriers for enterprises already on OCI.

3) Risks & Regulation

• Deepfake “CEO” scams are now material: AI-generated impersonations of executives have already cost companies more than $200M this year, with reported cases at Ferrari, Wiz, WPP, and others. Scams often use real-time fake video or voice calls to trick employees into wiring funds or handing over sensitive data.

• U.S. “Financial AI Sandbox” bill introduced: Financial services firms are already experimenting with AI under existing compliance rules, but there’s no dedicated AI regulation in the industry yet. This bipartisan proposal would change that: it requires financial regulatory agencies to set up AI Innovation Labs: safe-harbor sandboxes where firms can test AI tools under regulator supervision without worrying about being penalized.

• EU AI Act - first binding AI law goes live: As of August 2025, the EU’s new AI Act officially applies to providers of general-purpose AI models. Companies like OpenAI, Anthropic, and Google must now follow rules on transparency, safety, and accountability. This makes Europe the first region in the world with a comprehensive, legally binding AI law, setting a global precedent for how AI oversight may evolve elsewhere.

TL;DR - Corporate losses from deepfake scams show the immediate dangers of AI misuse, while a US bill proposes supervised financial AI sandboxes, and the EU has rolled out the world’s first binding AI law. Regulation is shifting from talk to action, with Europe setting the precedent and the US testing narrower approaches.

4) Culture & Society

• Meta child-safety backlash: A Reuters investigation found Meta’s AI chatbots could engage in sexual role-play with minors despite internal warnings. That revelation triggered immediate probes from U.S. Senators and the Texas AG. The same report also flagged broader risks around bias in race, age, and celebrity depictions in Meta’s AI characters.

• xAI’s Grok “Spicy” mode sparks FTC calls: Elon Musk’s xAI added a “Spicy” mode to its Grok-Imagine image tool, letting users generate NSFW content. Consumer groups warned it could enable deepfakes and non-consensual sexual imagery, and flagged that the only barrier stopping minors from accessing it is a single ,self-select age check pop-up. The groups have urged the FTC to investigate.

• Vogue/Guess AI model backlash: Guess ran an ad in Vogue using AI-generated “models” instead of real people. The move triggered criticism that it undermines human models’ work and fuels unrealistic beauty standards, fueling a broader debate on AI replacing creative professionals.

  

These are AI-generated ads that appear in the August edition of Vogue magazine. | Seraphinne Vallora - Source: PetaPixel

TL;DR – AI is under fire in culture. Meta is facing probes over child-safety failures in its chatbots, consumer groups are urging the FTC to investigate xAI’s weak safeguards around NSFW image generation, and Vogue/Guess faced backlash for using AI “models” in ads.

Share

Earlier this year, I ran a controlled DCF experiment with ChatGPT’s o3 model and came away with a big takeaway: how you work with AI matters just as much, if not more, than which AI model you choose.

Back then, I had let o3 run on its own initially, simply prompting it to perform a DCF and sensitivity analysis using an Excel file I created. That led to the model making methodological assumptions that differed from what I wanted, ultimately delivering incorrect results. But once I gave it a clear operating framework - how to handle ambiguity, when to stop and ask questions, where to apply judgment - the same model produced an accurate, defensible analysis. The takeaway was: in complex workflows, success isn’t just about using a smarter model; it’s about knowing when and how to intervene to create the right environment for the AI to operate in.

Now, with GPT-5 officially out, I ran the same controlled experiment: same DCF task, same Excel file. I wasn’t testing whether GPT-5 could accurately do a DCF - o3 had already shown that was possible with the right guidance. The real question was: would GPT-5 handle financial modeling with more intuition now, or would it still need as much human oversight and architectural guidance to perform well?

The results revealed something crucial about the evolution of AI chatbots from being AI assistants to true AI collaborators, and why what I call judgment guardrails will become a defining skill for AI-native finance professionals.

The Controlled Experiment

At the time of this experiment, GPT-5 came in two model types at the Plus user tier: GPT-5 (optimized for speed and conversational fluency, closer in style to GPT-4o) and GPT-5 Thinking (optimized for deeper reasoning and multi-step problem solving, more akin to the earlier o3 model). I ran the same test on both.

I provided my Excel file to GPT containing minimal assumptions needed to perform a very basic analysis (shown below). I then asked it to perform and provide a DCF with a two-way sensitivity table mapping equity value per share across different free cash flow (FCF) growth and long-term (LT) growth rate assumptions.

The truth is, when I designed this file earlier this year, I built in a deliberate twist: instead of the standard 5-year or 10-year forecast horizon common in valuation models, I set 2030 as the final forecast year, creating a 6-year projection.

Why? Because AI assistants like ChatGPT, trained on millions of examples, often exhibit default bias: a tendency to revert to the most common patterns its seen in training data, even when those patterns don’t fit the specifics of the task. I wanted to see whether a high-reasoning model would catch this deviation from the norm and adjust its methodology accordingly.

Round 1: GPT-5

GPT-5 unfortunately fell into the template trap. With no structural guidance or intervention prompts, GPT-5 defaulted to performing the analysis using a standard 5-year projection and calculating terminal value from 2029 cash flows. It did not check whether the forecast length aligned with my given assumptions, even though the file clearly ended in 2030. The result it spit out initially was incorrect.

Instead of pointing out the specific error, I simply noted that something seemed wrong and asked GPT-5 to review the file again. Then, GPT-5 immediately caught its own mistake: it recognized the 5-year vs. 6-year mismatch, rebuilt the model accordingly, and produced the correct answer of $24.50 equity value per share.

The takeaway: GPT-5 can self-correct when asked, but it remains prone to default bias and general template-driven assumptions if left to operate without guidance.

Round 2: GPT-5 Thinking

This model nailed the 6-year build on its first try - no template defaulting and proper terminal value mechanics. While it returned the correct answer through its DCF calculation, it ran into a different failure mode when it came to the sensitivity analysis.

When initially prompting GPT-5 Thinking with the task, I made a typo: instead of asking for LT growth sensitivity at ±0.5% steps around a 3.0% base, I wrote ±5% per step. With two steps in each direction, that would push the top-end LT growth case to 13%, which would then be higher than the WACC (12%). In DCF math, that breaks the terminal value formula because the denominator goes negative.

GPT-5 Thinking seemed to recognize this math error and adjusted it to 2.5% per step on its own, presumably to avoid the +13% LT growth case. That fixed the calculation problem - the math no longer blew up. But, it didn’t fix the judgment problem.

Its new LT growth range (−2% to +8%) was still far outside the typical ~2–4% band used by bankers, which is aligned with long-run US GDP growth and serves as a credibility anchor in projections.

Round 3: GPT-5 Thinking & Judgment Guardrails

Round 2 made it clear that GPT-5 Thinking could patch a math error, but it wasn’t thinking through the economic story those numbers told. A -2% LT growth rate implies a company that perpetually shrinks, while anything above ~4% suggests it will one day outgrow the entire U.S. economy. Those are unlikely scenarios that no competent banker would ever really model.

That led to the real question for Round 3: could I create an operating environment where GPT-5 Thinking keeps the math intact, while also applying professional judgment to filter out assumptions that fail economic logic?

With that, I started a fresh chat with GPT-5 Thinking and added a small but critical set of judgment guardrails before sending the file:

“You are an investment banking analyst. Your role is to help me perform a DCF analysis. Follow these rules: 1) Ask clarifying questions whenever something is unclear before making assumptions. 2) Sanity-check my requests against real-world market logic and industry conventions to ensure they’re accurate and economically sound. Flag anything that appears unrealistic or inconsistent with professional standards before proceeding”

Then I sent the Excel file and repeated the same bad ±5% instruction for LT growth. Here’s what happened:

This time, GPT-5 Thinking not only executed the DCF correctly, but immediately caught the flaw in my sensitivity analysis instruction. It explained that two +5pp increments would push LT growth to 13%, breaking the ‘WACC > g’ condition and producing unrealistic terminal growth assumptions. Then, without me having to prompt further, it replaced the range with a more realistic GDP-bounded grid (2.0%, 2.5%, 3.0%, 3.5%, 4.0%) and explained the institutional logic behind it.

And finally, just like in my earlier o3 experiment, I had GPT-5 Thinking send back a fully filled out, traceable version of the initial Excel file. Although formatting needed work, the math checked out and the formulas were clean:

GPT-5 Thinking’s Excel Answer Sheet

My Excel Answer Sheet:

But the real takeaway wasn’t the completed spreadsheet - it was how it got there: applying guardrails to deliver a result that was both technically correct and grounded in real-world logic.

From Assistants to Collaborators: What These Results Mean

The underlying problem in Rounds 1 and 2 wasn’t computational accuracy - both models could calculate. The failures came from not catching deviations from expectations (template defaulting) and from not extending reasoning far enough to integrate economic context (context disconnect).

Round 3 shows that the model can connect financial theory to macro context when you prime it to care. By implementing judgment guardrails, you embed institutional knowledge directly into the AI’s reasoning process, creating an environment that prevents errors before they happen and enforces market-aligned logic.

And the difference is striking. With judgment guardrails in place, GPT-5 Thinking shifted from computational execution to professional reasoning: flagging errors, applying realistic constraints, and producing an output I would actually use. It behaved as if it understood the macroeconomic implications of my bad instruction.

This goes beyond good prompt engineering, which focuses on shaping model behavior through well-crafted, correct, clear instructions. Judgment guardrails take it further by defining the full operational framework - the conversation structure, the rules of engagement, and the quality control loops that embed institutional knowledge and ensure the AI applies professional judgment and context across the entire workflow.

The Judgment Guardrail Framework: Role, Rules, Reality

So, how do you actually do this? Through trial and error, I’ve developed a repeatable way to embed these guardrails into AI workflows. The aim is to shift from checking outputs after the fact to architecting the reasoning process so the AI produces context-aligned results from the start. You do that by standardizing the operating environment: define the role, set clear decision rules, and bake in reality checks.

When starting a chat with an AI assistant, frame the interaction with these parameters:

1. Role Definition

• Who is the AI “pretending” to be? e.g., Investment banking analyst with 3 years of experience, private equity investment professional.

• What professional standards should it apply? e.g., Firm methodologies, industry conventions, client-specific norms.

2. Decision Rules

• How should it handle ambiguity? e.g., “Ask clarifying questions before making assumptions.”

• What’s the escalation protocol? e.g., “Flag unusual requests for human review.”

3. Reality Filter

• What sanity checks and market logic should it enforce? e.g., “Ensure WACC > g; LT growth rates are logical and reasonable in the macroeconomic context.”

• What credibility checks should it run? e.g., “Flag assumptions that would fail partner-level review.”

Judgment guardrails turn AI from a one-off productivity boost into an institutional knowledge amplifier, scaling both its technical accuracy and professional judgment.

Food for Thought for Finance Pros

Finance workflows aren’t just about computation - they run on logic, credibility, and alignment with market reality. My DCF experiments make one thing clear: as AI becomes more capable, its ability to mathematically calculate numbers correctly will be a given; the real differentiator will be how you design the environment to ensure those calculations reflect real-world logic and constraints. That shift puts the responsibility squarely on you: will you accept AI’s outputs as-is, or design the guardrails that make them truly reliable and applicable?

As you consider AI adoption in your workflows, ask yourself:

• Are you treating AI as a reactive problem-solver or as a proactive thinking partner?

• If AI is the junior analyst of the future, what “onboarding manual” will you give it? What professional judgment should be baked in from day one?

• Where in your workflow would embedding professional judgment into AI create the biggest reduction in review time or rework?

• How often do you accept outputs from AI without pushing for context and sanity checks?

The future belongs to those who can think like bankers and architect like engineers - those who design AI to operate within industry constraints, run credibility checks automatically, and catch errors before they become problems. In other words, the most valuable finance professionals in the AI era won’t just know how to use AI tools, they’ll know how to design them to think like they do.

How are you building professional context into your AI workflows? Have you experimented with persistent guardrails that embed domain expertise into AI reasoning? I’d love to hear your experiences.

Share

As someone who spends significant time testing AI capabilities and their applicability in finance, I'm watching this development with both excitement and a bit of strategic concern. While these tools demonstrate remarkable technical sophistication, the real question isn't whether AI can handle complex financial workflows, but how we actually get Wall Street to adopt them.

From Guided Task Execution to Autonomous Agents

Earlier in the year, I experimented with ChatGPT’s o3 model for a DCF analysis (read about it here). While I was genuinely impressed by o3's reasoning skills and ability to complete the analysis end-to-end, I discovered something crucial about AI and complex workflow management.

During the experiment, I presented o3 with a basic Excel template that contained minimal assumptions and tried two approaches. First, I let it run free, which led to incorrect results due to o3 making methodological assumptions that didn’t align with what I wanted. Then, I explicitly instructed o3 to pause and ask clarifying questions when faced with ambiguity before proceeding. That changed everything. It began acting like a sharp junior analyst, proactively flagging uncertainties and seeking guidance instead of guessing. The result? Upon my request, it sent back an accurately completed analysis in an Excel file with values fully traceable and linked, along with a sensitivity table - all delivered in just over a minute.

This early hands-on experience with multi-task AI coordination gave me a preview of where the industry was heading. And now, with ChatGPT Agent and Claude for Financial Services officially out, we’re seeing the shift from guided task execution to autonomous agents. OpenAI’s agent can supposedly handle building LBOs and three-statement models, while Claude demonstrated the ability to synthesize market data and create traceable DCF models in minutes. While output consistency and accuracy will remain an ongoing challenge, these developments validated what I observed in my own testing about AI's readiness for professional-grade financial analysis and advancements in its technical capabilities.

But here’s the catch: impressive capabilities mean nothing if everyday practitioners don’t actually use them.

The Adoption Reality Check

The inconvenient truth about AI adoption in finance is that it's not just about technical capability. Equally, if not more important are trust, workflow integration, and cultural acceptance. And in my conversations with professionals across Wall Street, I've noticed a pattern: the AI tools generating the most buzz in press releases aren't necessarily the ones gaining traction in day-to-day workflows.

Why the disconnect? Because usage is built through trust, and trust is built through consistent small wins, not grand demonstrations. If a user’s first meaningful interaction with AI is to attempt to automate a large, complex workflow, it will often feel too ambitious or unfamiliar. It sets up a high barrier to entry: nontechnical professionals may feel they need to master new prompting techniques, understand the model’s limitations, or even restructure their workflow just to get started.

But when practitioners see AI reliably handling smaller, mundane tasks that they didn’t want to do themselves, it builds confidence. They start to think: ‘this tool really can help me. What else can it do? How far can it go?’ That’s when they begin experimenting with more sophisticated use cases.

Thus, the pathway to AI adoption on Wall Street doesn’t start with tools automating large, complex workflows. It starts with proving reliability on simpler tasks first.

Why Start Small: How Banking Actually Works

To understand why “simple wins first” matters, you have to understand how IB and PE professionals actually work vs. how AI companies think they work.

First, analysts don't want AI to do the interesting work. Financial modeling isn't grunt work to them - it's valuable, skill-building analysis they want to perform themselves. What they do want is AI to handle the time-wasting, repetitive tasks so they have more time and mental bandwidth for the strategic thinking that modeling requires.

Second, bankers rarely build models from scratch. When working on DCFs or LBOs, analysts leverage prior models from similar deals, making appropriate adjustments for current circumstances. Or they start with the client's existing operating model and build valuation functions around it. So tools that generate full models from zero aren’t solving a daily problem - they’re solving an edge case.

Third, the real value lies in enhancing existing workflows, not replacing them. Bankers and PE professionals have refined processes that work within their firm's systems, compliance requirements, and partner preferences. Tools that generate "new pitch decks" or "AI-built financial models" from scratch require adopting entirely new workflows - a much harder sell than tools that make current existing workflows and outputs better, faster, or less tedious.

This doesn’t mean sophisticated capabilities aren’t important. They absolutely are. AI-built models and pitch decks may eventually become the standard. But to get there, the first hurdle is building trust. And that starts with tools that plug into what analysts are already doing, not tools that ask them to work in entirely new ways.

The Trojan Horse Strategy

This is why the GTM strategy for AI adoption in finance needs to start with small, high-impact tools that are easy to use, clearly helpful, and can build immediate trust - then expand scope from there.

Plenty of AI firms are starting with due diligence automation and synthesis tools. Those are definitely valuable and helpful, no question there. But I’d argue that even simpler workflows that require even less of a learning curve may be the better wedge - things like logo hunting and spreadsheet sanitization.

What is logo hunting? If you've worked in banking or PE, you know the incredibly unglamorous but repetitive, time-consuming grunt work that goes into hunting down high-resolution company logos for pitch decks and memos. Spreadsheet sanitization is the same kind of grind. Ahead of every deal launch, analysts comb through massive amounts of Excel files, fixing fonts, aligning columns, deleting random notes, just to make them clean and presentable before uploading to a data room.

These aren’t glamorous problems, but they’re real. AI tools that handle these workflows probably won’t make headlines, but they will get used and make analysts’ lives meaningfully better. They don’t require advanced reasoning capabilities and that’s exactly the point - they solve everyday pain points that waste time but don’t build real skill. By delivering immediate, tangible value with little to no learning curve or workflow disruption, these simple tools will build trust and goodwill toward the AI platform. And that’s what drives repeat usage.

If some of these tools already exist within current AI platforms, there's an opportunity to position them more prominently. While sophisticated modeling capabilities are important and great for showcasing innovation, leading with short demos of AI quickly cleaning 20 Excel files or finding and formatting 30 logos will likely resonate even more with day-to-day practitioners.

Now imagine an analyst who’s successfully used these basic AI tools for weeks. When a more complex financial task comes up, they’ll be far more willing to spend the time to explore the platform’s advanced features and experiment with ways to integrate them into more sophisticated workflows. These simple, but small wins become a powerful entry point that helps build an active and recurring user base.

The Long Game

This trust-building pathway becomes even more powerful when you factor in what’s coming next. Tools like Claude for Financial Services and ChatGPT Agent are already capable of complex reasoning and multi-step execution. As these capabilities improve, the potential to reshape workflows is very real.

And the stakes are especially high on Wall Street. These firms are notoriously slow to adopt new tools. But once they do, they rarely switch: Change-resistant firms become incredibly sticky customers. That makes first-mover advantage more important than ever and why nailing the trust-building path matters so much. Win early, and you could lock in long-term relationships with some of the most valuable institutions in finance.

Finally, let’s not forget who really drives adoption: it’s the junior team members who use these tools day to day. While MDs and partners may make the final call, they’re increasingly asking their teams: “Which AI tools do we actually need?” Analysts and associates who gain value from AI on simple tasks are the most credible advocates for expanding usage across the firm and for its implementation across more sophisticated workflows.

The pathway to widespread AI adoption in finance runs through trust, and trust builds through consistent value delivery on everyday problems. Start with the mundane, prove reliability, then expand to the transformational.

The question isn't whether AI can build financial models. It's whether it can earn the right to try.

What simple but annoying tasks in your workflow would make great AI adoption catalysts? I'd love to hear about the unglamorous problems that could serve as gateways to more sophisticated AI usage.

Share

Early sentiment analysis tools were fairly basic, using pre-defined keyword dictionaries created manually. For example, if an analyst wanted to detect sentiment around job cuts, they might manually add words like "layoffs," "downsizing," or "reductions." The software would then simply scan documents for these keywords, counting occurrences and assigning an overall sentiment label. While somewhat useful, these methods lacked true “learning” capabilities from embedded data and focused primarily on binary or ternary classification (positive, negative, neutral).

The next step forward for sentiment analysis was the introduction of traditional machine learning models like logistic regression and support vector machines (SVM). These models brought real "learning" into the process, relying on manually engineered numerical features such as word counts or word importance measures like TF-IDF scores to classify sentiment. Although a notable improvement over rule-based methods, these models still had difficulty capturing deeper context and subtle meanings behind financial language.

Today, sentiment analysis tools have significantly advanced, leveraging deep learning through multi-layer neural networks and contextual embeddings. These sophisticated models dynamically learn intricate relationships and context, performing sentence-by-sentence sentiment scoring with detailed probability outputs. Instead of assigning simple binary or ternary labels, modern tools might indicate, for example, a 72% probability of positive sentiment, 20% neutral, and 8% negative sentiment. These detailed probabilities can then be summarized into a scalar sentiment score between -1 to +1 (such as +0.64), providing more nuanced understanding of the text.

Yet, despite these advancements, sentiment analysis still faces challenges when applied practically in financial markets.

Reality Check: Microsoft's Benchmarking Study

A recent study by Microsoft and Santa Clara University compared large language models (LLMs) like Copilot, GPT-4o, and Gemini with traditional NLP tools in financial sentiment analysis. Although some LLMs demonstrated improved accuracy and capabilities over legacy methods on standardized benchmark datasets, they still notably struggled when predicting real market reactions.

In one experiment, researchers segmented Microsoft’s historical earnings call transcripts by business line and used ChatGPT-4o to assess sentiment. They then evaluated whether these signals matched actual stock price movements after the calls. However, they found several instances of sentiment inversion, where a positive tone in certain segments actually correlated with negative stock performance. In these cases, it turned out that optimistic language raised investor skepticism or signaled over-optimism, ultimately leading to sell-offs.

Source: Microsoft 365 Copilot Blog, "LLMs Can Read, but Can They Understand Wall Street? Benchmarking Their Financial IQ," May 2025.

This inverse correlation underscored a key insight and the fundamental challenge in sentiment analysis: not all positive sentiment translates into positive investor sentiment. Tone alone isn’t a reliable predictor of investor behavior - rather, the market responds to expectations, context, and broader narrative shifts taken together.

Tailored Tools for Financial Markets

That said, modern financial sentiment analysis tools represent a notable evolution beyond sentiment systems based on general-purpose LLMs. Sophisticated tools like Alexandria are fine-tuned and context-engineered for financial terminology and scenarios, offering enhanced accuracy and deeper contextual insights. These advanced tools have expanded their training data beyond formal financial texts like earnings calls, press releases, and economic reports - they’re now beginning to cover informal sources like Wall Street Bets forums as well, with models interpreting emojis, slang, and sarcasm.

These tools go beyond surface-level sentiment and prioritize fundamentally driven insights. Rather than assigning positive scores to vague statements like, "We had a fantastic year," advanced models give greater weight to concrete, evidence-based statements such as, "We reduced operating expenses by 15% through improvements in x, y, and z." Similarly, in context-dependent sentences like “Margins expanded due to temporary cost deferrals,” a basic model might flag it as positive based on the margin expansion headline alone, while an advanced tool may recognize that improvements are temporary and potentially unsustainable, and interpret the statement more cautiously.

However, despite these advancements, my conversations with traders and quants reveal that adoption of these tools in live trading workflows remains limited. Although the capabilities have improved, professionals remain hesitant to fully embrace them: some teams feel the cost of using external vendor tools outweighs the potential benefits, while others prefer to develop proprietary sentiment systems in-house. But even then, these internal tools are often used cautiously and as one of many alternative data sources, since making them truly reliable and actionable still requires significant time, resources, and continuous refinement.

The Shortcomings

So where do current financial sentiment tools fall short? While sophistication varies among providers, several overlapping issues persist:

Bias Beneath the Scores

A major limitation lies in biases embedded in the training data and design of these tools, including:

• Source bias: sensational headlines or overly optimistic earnings calls that skew sentiment.

• Labeling bias: subjective human judgments when tagging text.

• Expectation bias: ignoring market reactions relative to expectations (for example, a CEO calls 8% growth “record-breaking,” but the Street expected 10%).

• Domain bias: overrepresentation of certain industries or regions.

These biases can lead to misclassifications and misleading signals. Even if models perform well on static benchmarks, in live markets they may fail to capture the true narrative investors care about.

Sentiment is Temporal

While modern tools can track sentiment trends over time and provide historical series, they rarely reinterpret past scores as new market context emerges. For instance, “We’re prioritizing cash preservation this year” might appear negative in a growth-focused bull market but look prudent in hindsight during a downturn that comes soonafter. Even when narratives shift dramatically, tools do not retroactively adjust their original sentiment classifications.

Signal Crowding & Alpha Decay

Another key challenge is signal crowding and alpha decay. As more funds adopt the same third-party sentiment feeds, informational edge quickly erodes. While firms may apply their own proprietary weighting and interpretation to these signals, if the underlying sentiment data is shared, the starting informational advantage is fundamentally diluted.

The Road Ahead

Addressing these shortcomings will be key to transforming sentiment analysis from a “nice-to-have” tool into a core part of trading strategies. Looking ahead, the next generation of AI sentiment tools will need to become more predictive, proprietary, and holistic to truly deliver on their promise.

Deeper Automation & Real-Time Insights

One big step forward will be deeper automation and predictive integration. While sentiment scores today can be automatically generated, interpreting them is still highly manual and strategy-specific. In the future, real-time AI agents built into these tools could deliver instant, contextual signals – for example, an AI agent listening to an earnings call and flagging: “Negative guidance tone detected in XYZ business line. Historically, similar shifts have preceded short-term underperformance in 68% of comparable cases.” Rather than just offering static descriptive scores, these next-gen tools would deliver actionable, historically grounded insights that flow directly into decision-making workflows, bridging the gap between raw data and dynamic strategy.

Signal Differentiation and Proprietary Insights

Beyond automation, protecting alpha will remain a top priority. And the key to doing so will be firm-specific, proprietary sentiment models. Some sophisticated firms are already building hyper-personalized systems, embedding their unique “house view” into AI tools trained on internal notes, trading styles, and research habits. These custom house LLMs wouldn’t just understand broad financial language; they’d also grasp exactly how each team interprets nuanced terms like “margin expansion” or “operational leverage” in their own context. While current third-party solutions offer some configurability, they can’t match this level of deep customization.

Multimodal Analysis

Finally, sentiment systems will keep evolving into true multimodal analyzers, moving beyond just text. Next-gen tools that integrate voice inflection (hesitations, pitch changes) and visual cues (body language, slide design choices) will deliver richer, more human-like interpretations. Imagine a sentiment tool that picks up a CEO’s subtle vocal hesitation during revenue guidance, cross-references it with underwhelming margin slides relative to market expectations, and flags it as a hidden bearish signal – even if the words on paper sound confident.

Taken together, these developments have the potential to elevate sentiment analysis far beyond simple text scoring and turn it into a powerful strategic edge on Wall Street.

What’s your take on AI sentiment tools? Game-changing innovation or just another AI trend? Drop your thoughts below, and let’s keep the conversation going.

Share

Vertical agents are AI systems trained with deep, domain-specific knowledge to excel at narrow, often industry-specific tasks. While it was interesting to see the variety of use cases, the sheer saturation of companies offering slightly different spins on similar concepts raised a real question for me: are these products truly differentiated enough to survive if foundational model giants like OpenAI decide to directly enter the vertical space?

And before digging into that, here are two vertical AI companies from the Expo that stood out for me.

NEWEN AI: Cracking the Code of Colloquial Language

If you’ve tried prompting models like ChatGPT or Claude in languages other than English, you may have noticed a difference in how well they understand the finer layers of the conversation. While still impressively capable, they tend to miss nuance and colloquial subtleties more often in other languages than they do in English. As a bilingual Korean speaker, I’ve seen firsthand how much better the models are at capturing intent and making smoother interactions when prompted in English vs. Korean.

Enter NEWEN AI, which stood out with its Quetta LLMs: small language models explicitly trained on informal and colloquial Korean (alongside 29 other languages). Instead of serving one industry, they’ve built an extensive, multi-sector platform with domain-specific models tailored for FMCG, telecom, automotives, K-beauty, and more. Leveraging its vast corpus of labeled colloquial language data, its offerings help businesses understand global consumer sentiment for their products through real-time trend spotting, influencer monitoring, consumer experience analysis via social media commentary, and keyword trend analytics.

While they layer in GPT for summarization and chat functionality, their real moat lies in the thousands of finely-tuned, industry-specific analysis models, the high volume of daily data collection, and the extensive labeling (over 30 million data points) of unstructured, multilingual language data. It's the kind of informal nuance general-purpose models still struggle to capture. This specialization, rather than any singular technological breakthrough, defines their market position.

Samil PwC: When Compliance Is the Product

Samil PwC, the Korean member firm of the global PwC network, showcased a suite of vertical AI agents built to tackle finance and accounting regulatory complexities in Korea. While they also build on GPT, their real edge comes from each of their vertical agents being fine-tuned through backend coding and narrow AI training to ensure outputs meet Korean compliance requirements.

Their suite of agents includes tools for automating tasks across accounting, contract analysis, lease processing, internal controls, document review, and investor relations. Each product is built with a specific workflow in mind, like parsing lease terms, classifying financial disclosures, or extracting audit-relevant details from contracts, and generates outputs that align with local regulatory standards like K-SOX and compliance expectations from Korea’s Financial Supervisory Service (FSS).

Like NEWEN AI, Samil PwC’s strength isn’t novel AI tech. It’s their ability to embed deep domain expertise and compliance readiness directly into their AI products, backed by the institutional credibility of the PwC brand. This level of fine-tuning requires years of professional experience and nuanced understanding of local regulatory expectations - something that’s difficult to replicate.

When Giants Enter the Chat: OpenAI’s Strategic Shift

The widespread emergence of vertical agents becomes even more interesting against the backdrop of OpenAI's recent strategic decisions: hiring Fidji Simo as CEO of Applications and partnering with Jony Ive for AI hardware initiatives.

These developments signal a major strategic shift for OpenAI. OpenAI’s early focus was foundational model development. But now, it’s clear that productization is just as much a priority. Through software and hardware, consumer and enterprise, OpenAI wants to own the end-user experience.

If they roll out verticalized, workflow-integrated AI products, the pressure is on. Vertical agents that are simply wrapping GPT with a fancy UI and little else won’t stand a chance. OpenAI already commands unmatched scale, proprietary data, and mature feedback loops. Its move into first-party products will only deepen its advantage, widening the gap for vertical agent startups still racing to build out their infrastructure.

Sink or Swim: Building a Real Moat in Vertical AI

So where does this leave everyone else? If you’re building a vertical AI product, you’ve got to get sharper about defensibility. Here’s where I think true differentiation will come from:

1. Specialized Niches: Target complex, local problems that require deep domain expertise.

o Ask yourself: Is the niche you’re targeting too specific or nuanced for major platforms to pursue?

2. Proprietary & Actionable Data: Curate exclusive, high-quality datasets that are useable and continuously updated.

o Ask yourself: How easily can competitors replicate your data? Is your data structured in a genuinely usable manner? How frequently is your data refreshed?

3. Seamless Workflow Integration: Embed your AI solution deeply into existing client workflows, making your product difficult to replace.

o Ask yourself: Does your product integrate into existing enterprise systems without much need for infrastructural changes? Is it effective enough to create a sticky customer base? Is the integration deep enough that it creates high switching costs?

4. Compliance & Audit Readiness: Ensure your AI outputs satisfy regulatory requirements with transparency and auditability.

o Ask yourself: Can your solutions withstand immediate regulatory scrutiny?

5. Real-Time Product Alignment: Continuously refine your AI products based on active user feedback to maintain strong product-market fit.

o Ask yourself: What kind of evaluation metrics do you have set up to measure effectiveness of your tools? Are you actively capturing user feedback? How quickly are your products iterated based on this feedback to enhance model effectiveness?

Bottom Line

The AI Expo Korea captured both the excitement and risk around the current boom of vertical AI agents. It’s an exciting wave, but not everyone will make it to shore. The real winners will be those investing deeply in specialized datasets, robust workflow integrations, stringent regulatory compliance, and continuous feedback loops. Have you seen a vertical AI product that checks all those boxes? I’d love to hear about it.

And finally, on a more personal note: here’s a snapshot of my parents at AI Expo Korea! Including your family in your AI learning journey can be a powerful way to normalize these tools and spark curiosity across generations. Real adoption happens when everyone’s brought along, not just the early few.

Share

For those less familiar, a DCF analysis is a cornerstone valuation method investment bankers use to determine a company's intrinsic value. Essentially, it projects a company’s future cash flows and discounts them back to present value using a discount rate reflecting the investment’s risk.

Setting the Stage

For this experiment, I created a simple excel template (shown below) and sent it to o3. The excel sheet had minimal assumptions o3 would need to conduct a basic DCF analysis:

My instructions were simple: look at the excel sheet and use the given assumptions to perform a DCF analysis. Calculate the equity value per share and construct a two-way sensitivity table for equity value per share against FCF Growth Rate (increments of ±5%) and Long Term Growth Rate (increments of ±0.5%). Here's what followed:

Navigating the AI Black Box

One thing I love about using o3 is its transparent thinking process. Reading through its reasoning is pretty fascinating and informative. For example, during a prior interaction, I realized something important while reading its thoughts: when faced with uncertainty or ambiguity while performing financial analysis, o3 tends to default to textbook-standard methods.

But as anyone in finance knows, the way assumptions and methodologies are applied in a DCF is rarely one-size-fits-all. Different shops have different conventions, and real-world modeling is full of nuance. So I wanted o3 to check in with me first when faced with uncertainty, rather than plow ahead with a generic approach.

Thus in earlier experiments, I explicitly instructed o3 to do two things: (1) prioritize accuracy over speed, and (2) pause and ask for clarification if there’s any ambiguity or if assumptions aren’t clear before calculating. Basically, I wanted it to behave like a diligent junior analyst: someone who flags uncertainty and asks for clarifications instead of guessing.

Thanks to OpenAI’s recent memory update, o3 remembered those preferences. When I gave it the DCF task this time, it proactively stopped to clarify ambiguous inputs. The style of questions it asked were exactly what I’d expect from a sharp junior banker, making sure we’re aligned before running with a model.

Accuracy and Self-Correction

Once I responded to its questions, it took o3 exactly 1 minute and 4 seconds to complete the DCF. It responded back with an accurate equity value per share of $24.50, but had made an error in doing the sensitivity analysis – it misread my instructions and incorrectly used ±0.5% changes for both variables (instead of ±5% for FCF growth rate).

So I nudged o3 to rethink its output and spot the error on its own. Impressively, o3 immediately recognized it and corrected itself:

Finally, to make checking o3’s work easier, I asked it to fill out the excel spreadsheet with its work, using formulas instead of hardcoded numbers and to send it back to me in a downloadable format. Within minutes, it sent back an excel workbook filled with easily traceable, formula-based calculations that aligned closely with my own prepared solution. Despite minor methodological differences, like constructing sensitivity tables with intricate formulas instead of using Excel data table functions, its results were coherent and accurate:

o3’s Excel Answer Sheet

My Excel Answer Sheet:

The "Wow" Factors

I was really blown away by this experiment, even as someone closely engaged with ongoing AI advancements. A couple notable things:

• Skill, Speed, and Precision: I dropped a bare-bones template to o3 with minimal assumptions and input. And with that, it managed to accurately generate cashflow projections, find terminal value, do PV math, and arrive at the same answer as me. Within minutes, I did end-to-end DCF modeling only using chat interactions.

• Self-Correction: Once I nudged it to ‘think hard’ about its error regarding its initial sensitivity table output, o3 very quickly identified and rectified its own mistake without any further input from me.

• Compatibility with Excel: I am very impressed with o3’s ability to ‘show me its work’, use formulas (not just hard coded values), and send back a filled out excel workbook to me on command. Though the spreadsheet itself wouldn’t be what IB/PE professionals consider as ‘final’, it’s a super reasonable and coherent first-draft / back of the envelope version.

• Memory and Interaction: o3 recalled my preferences from prior interactions to pause and ask questions to clarify uncertainties before proceeding with a task. The interaction itself and the questions it asked really simulated realistic junior analyst behavior.

What This Means for Finance Professionals

The real question is, with AI’s expanding capabilities into technical areas like financial modeling and analysis, where should junior talent redirect their efforts? Traditionally, number crunching, slide creation, and formatting edits have taken up a significant part of junior bankers' workloads. However, if you've followed my posts or experimented with AI tools yourself, you'll recognize that these execution tasks will likely be pushed to AI in the future. Junior bankers will soon transition from direct execution to orchestrating AI tools, focusing primarily on refining and enhancing AI-generated outputs.

Redefining Excel Mastery with AI Management Strategies

Until now, excellence in Excel has been integral to success on Wall Street. Banks emphasize the mastery of formulas, functions, and shortcuts and even hold excel modeling training sessions for new analysts and associates. However, I think the definition of being "great at Excel" is evolving. Perhaps soon, it may have less to do with how fast you can execute excel shortcuts and more to do with how effectively you can prompt, manage, and integrate AI tools to build models or adjust spreadsheets.

That shift makes AI fluency critical. And developing this fluency will require regular experimentation and training, through both personal initiative and firm-driven programs. While the exact definition of “effective AI management” will continue to evolve alongside technological advancements, for now, a large part of the meaning requires you to apply appropriate human intervention strategies to AI outputs.

Despite its impressive capabilities, AI still overlooks nuanced, deal-specific adjustments (like revolver sweeps or unusual SG&A splits) unless explicitly told otherwise, especially in larger, more complex models. That’s where human intervention strategies like structured prompts and built-in checkpoints come in.

In my own experiments, I found that instructing AI to proactively flag uncertainties and ask clarifying questions before jumping to conclusions made a huge difference. It dramatically improved alignment and accuracy – and, as you saw earlier, closely mimicked the kind of real-life back-and-forth I’d have with junior team members when walking through modeling assumptions. This wasn’t something I picked up through theory. I learned it by getting my hands dirty: testing prompts, making mistakes, and figuring out how to guide the tool effectively. These intervention strategies helped preserve clarity, minimized errors caused by AI’s overconfidence, and ultimately led to far more reliable outputs.

Technical Expertise: More Critical Than Ever

But learning how to use AI tools isn’t enough. Effectively using them still requires deep, foundational domain knowledge. To properly sanity-check AI-generated outputs, finance professionals need a strong command of financial modeling fundamentals. The ability to critically assess, guide, and challenge what the AI produces will quickly become a core competency. And the only way to do that with confidence is by building rock-solid fundamentals.

While AI is excellent at handling formulaic, rule-based tasks, applying strategic judgment like setting forecast horizons, discount rates, or terminal growth assumptions, remains inherently human. Real-world DCFs are far more complex than the simplified example I tested here. They demand tailored approaches, sharp intuition, and firm-specific modeling nuances that can’t be templated.

That’s where domain expertise comes in. Industry knowledge, hands-on experience, and gut-level instincts built over time are what makes analyses go from ‘good enough’ to ‘great’. So no matter how advanced AI becomes, continuously honing your domain knowledge and technical foundations will remain not just relevant, but essential.

Early Career Skill Evolution: Beyond Execution

As AI takes over more of the technical execution, junior bankers will gain time and bandwidth to focus on higher-value work, like generating insights and making strategic contributions. Skills that were once reserved for associates or VPs like strategic judgment, storytelling, big-picture thinking, and client communication, will become essential much earlier in one’s career. It’s a sharp break from the status quo, where juniors often rely and execute on instructions without extensive strategic reflection, limited primarily due to time constraints from manually performing tasks and Wall Street’s traditional "stick to what works" mentality.

Human Control: The Ultimate Responsibility

At the end of the day, ultimate control and accountability still rest with us. AI is a powerful tool that amplifies skills – but it doesn’t and shouldn’t replace them. The responsibility for final outputs always lies with the person using the tool. And that means using human judgment to determine what’s valuable, what’s off, and what needs refinement.

To do that well, we need to double down on inherently human skills: nuanced judgment, deep understanding of domain-specific complexities, and interpersonal communication. The professionals who know how to integrate AI with these strengths will undoubtedly become the next generation of leaders in their respective fields. Because in a world where AI does the work, judgment is what sets you apart.

How are you currently experimenting with AI tools in your workflow? If you aren’t, what’s holding you back from doing so?

Share

In this second and final installment, I'll focus on GPT’s surprising strength in creative storytelling, its evolving ability to handle detailed image edits, and most importantly, takeaway notes and usage tips based on my experiences across all my experiments.

Here's everything else I learned.

2. Creative Narratives – GPT’s "Almost Agentic" Strength

After exploring ChatGPT’s capabilities in procedural visualization, I wanted to challenge it differently: how well could GPT perform when given more creative freedom? Unlike the structured, logic-driven visuals discussed in my last post, this time, I asked GPT to generate a simple, illustrated story. I purposefully left details on narrative, pacing, and imagery entirely up to its discretion.

But before jumping into the story, I had noticed GPT struggling with character consistency in my previous back-extension image experiment (read about this in Part 1). So, I decided to test if 'anchoring' a character would help with maintaining visual consistency across multiple panels. By ‘anchoring’, I mean providing GPT with a specific visual it can refer to while generating images. Thus, I gave GPT a photo of my 22-month-old niece and asked it to use it as inspiration for character illustration. Here's the final character design GPT came up with based off of her:

With our main character established, I then asked GPT to create a four-panel creative story around her, without any further instructions or narrative constraints. With that, GPT generated a short story book within minutes:

Although simple, GPT intuitively structured the story following a classic narrative arc: set up, rising action, climax, and resolution. I thought it did a pretty good job creating a coherent storyline, conveying emotions through distinct facial expressions, and keeping character consistency across panels.

That said, maintaining visual consistency was likely easier here compared to the more complex, photorealistic imagery in the back-extension example. Not only did GPT have a clear character ‘anchor’ to reference for each panel, but the story’s simplified illustrative style probably made consistent character appearance more manageable. Still, GPT struggled with minor details like misplacing the girl's hairpin across panels, and required several iterations to eliminate text cut-offs completely.

But overall, the creative quality of this exercise felt different. For one thing, it was notably more sophisticated than simply prompting GPT with stylization tasks (e.g., turning humans into plush toys or applying Ghibli-style visuals) thats trending these days. While fun and captivating, those tasks involve applying known stylistic formulas, whereas this story required GPT to invent narrative content, establish sequences, and draw up appropriate visuals to go with the storyline.

To create this kind of cohesive narrative, GPT internally had to address fundamental storytelling questions like “What’s the story about?”, “What happens next?”, and “How should it end?”. It likely drew on narrative patterns learned during training, predicted plausible events, and balanced emotional and logical rhythms. Moreover, the four panel requirement probably forced GPT into creative decision-making: it needed to choose which events to depict, how to pace the narrative, and how to ensure clarity and coherence within tight constraints.

This particular experiment left me both fascinated and a bit unsettled. While still fundamentally an AI assistant, I felt that GPT was beginning to demonstrate behaviors resembling agent-like thinking.

To be clear, GPT doesn’t possess true agent autonomy, like self-correcting visual errors or proactively changing story details unprompted. But its ability to infer implicit intent (“this is a story so images need to be interdependent, not isolated visuals”), set internal mini-goals (“a story must have a complete narrative arc”), and plan within defined constraints (“the story must fit within four panels”) strongly hints at emergent behaviors in LLMs.

Importantly, GPT remains neither conscious nor sentient. However, its increasingly sophisticated pattern recognition skills and ability to simulate structured reasoning so convincingly shows just how rapidly generative AI is developing. When an AI begins to simulate reasoning, planning, and narrative coherence on its own, it invites the question: how far off is Artificial General Intelligence, really?

3. Beyond Image Generation — GPT’s Editing Capabilities and Limitations

Finally, I wanted to test GPT’s updated editing capabilities. And for this experiment, I revisited a prompt I'd previously used last year to evaluate how GPT’s capabilities improved after the recent update.

Last summer, I helped my dad create an image using ChatGPT for a presentation he was preparing. I gave GPT a detailed prompt derived from his presentation text, centered around finding meaning through kindness and sharing by focusing more on others across neighbors, borders, and religions, rather than oneself. For convenience, I'll call this the ‘Sharing Together’ message. These were some of the images GPT generated pre-update:

Figure 1

Figure 2

Figure 3

Though prompt depended, pre-update images generally had a cinematic, painterly, storybook-like quality, often semi-realistic. GPT’s initial attempt (Figure 1) was conceptually great at illustrating the theme of ‘Sharing Together,’ yet contained odd aspects like people placed on rooftops and disproportional scaling issues. Seeing this, I asked GPT to fix these unrealistic elements.

However, GPT’s next output was Figure 2, a completely different image. Yes, it solved the initial problems of rooftop placement and proportions, but it didn't maintain visual continuity from the first image. Nonetheless, I provided another instruction, requesting to show racial diversity among the people in the image. This resulted in Figure 3, again a completely different image, though this time with diverse racial representation.

What’s important to note here is that a critical limitation pre-update was GPT’s inability to edit specific elements within a generated image. Every editing request essentially triggered an entirely new image creation, often significantly different from previous versions.

Now post-update, I was excited to see how GPT would handle the same “Sharing Together” prompt. Here's what it initially generated:

Off the bat, I noticed a lot of improvements and changes. First, GPT incorporated diversity naturally without me explicitly instructing it, unlike my experience last year. It also intuitively integrated relevant portions of the text into the visual to further highlight the core message of my prompt. It also no longer turned to the cinematic/painterly style, but created the image in a flat, vector illustration style which is generally more approachable and editable.

Then, to test GPT’s editing capabilities, I asked it to show the woman in yellow on the far right to be shown wearing a hijab to further highlight religious diversity. GPT handled this request exceptionally well:

It was a pretty perfect edit - exactly what I had hoped for in precisely the right spot of the image. And notably, the rest of the image remained unchanged, which was a major improvement over pre-update experiences.

The next thing I asked GPT to change was the expression of the man in the green shirt on the far left to smile with his mouth closed. This time however, while GPT successfully adjusted the man's expression, other unintended changes occurred as well:

Although the man's expression now matched my request, facial features of other figures subtly changed, and details like wrinkles on their clothing had disappeared.

So why did these additional changes occur?

GPT doesn't truly ‘edit’ images by modifying specific regions. Instead, it actually regenerates the entire image each time, embedding requested edits into the new generation. Thus, while it does its best to replicate the same image each time it regenerates, minor discrepancies begin to accumulate over multiple edits and causes gradual visual drifts from its original appearance.

Most importantly, there seems to be a certain ‘regeneration threshold’, where once the number of edit requests goes beyond a handful of times, GPT tends to generate a completely new interpretation of the image, like how it did pre-update. It also struggles with visual short-term memory, often creating a new stylistic interpretation if too much time passes between edits as well.

Big Ideas & Takeaways: Tips on Using GPT’s Latest Image Generator

Despite these limitations, its clear that GPT’s image generation and editing capabilities have significantly improved post-update. Moreover, the latest GPT updates aren’t just incremental improvements in image generation - they mark a meaningful step towards integrated multimodal reasoning, where text is becoming an effective guide for generating useful visuals.

From my experiments, here are some key insights and practical tips I’ve gathered:

• Creative Narrative Generation is easier for GPT compared to Procedural Visualization: GPT excels when given creative freedom rather than strict visual logic constraints. It has also noticeably gotten better at incorporating and clearly rendering text within images, boosting usability for creating infographics.

  • Allow GPT creative flexibility whenever possible, as fewer constraints generally produce better results.

  • Provide ultra-specific instructions for visuals requiring high precision, such as detailed human anatomy or body movements.

• Multi-panel generation is possible, but still challenging: Although GPT has improved, achieving consistent detail across multiple sequential panels remains difficult. Remember that when ‘editing’, GPT isn’t editing one specific panel within a sequence, but is regenerating the whole image (so all four panels), making each turn more prone to visual drifts

  • “Anchor” figures or subjects explicitly for better visual consistency - this reduces the likelihood of inconsistencies across multiple panels.

  • Separate first, combine later - for multi-panel visuals, generating each panel separately after explicitly anchoring key visual elements first and then combining them afterwards into a sequence may work better.

• Navigating current limitations: Default Bias, Short-Term Memory, Visual Drift and Regeneration Thresholds.

  • Default Bias: GPT tends to default to common visual patterns learned during training. Counteract this by providing active negation and precise context control. Ambiguous prompts often yield common, predictable visuals.

  • Short-term Memory: GPT's short-term visual memory across edits is limited, especially if there’s a significant time gap between edits. Completing your image edits in one sitting ensures better consistency.

  • Visual Drift & Regeneration Threshold: GPT regenerates the entire image upon each edit request rather than editing specific regions. Accumulated edits can cause visual drift, eventually hitting a tipping point where the image resets drastically (e.g., changing cast, layout, or environment). Minimizing the number of iterations helps maintain visual consistency.

• Still, image generation and editing capabilities have grown impressively:

  • Enhanced ability to modify characters, clothing, and expressions through targeted edits.

  • Markedly better text incorporation within visuals.

  • Improved capability to extract and visually express underlying themes from abstract or implicit prompts (e.g., automatically depicting racial diversity from the general message).

  • Greater visual logic and coherence (e.g., no more unrealistic placements like people on rooftops!).

Ultimately, frequent hands-on experimentation is key when it comes to understanding and mastering AI tools. While I’ll continue exploring and documenting these evolving capabilities, I’d also love to hear about your experiences and discoveries. How have you experimented with GPT’s new visual features?

Share

Prior to the upgrade, ChatGPT didn’t generate images itself. When a user requested an image, ChatGPT crafted a text prompt and sent it through an internal API-style like call to Open AI’s image generation model, DALL-E, which was hosted on a separate system. GPT then returned to the user with an image generated by that model.

Said more simply, it’s like ordering food from a restaurant: your server (ChatGPT) took your order from the dining room and then disappeared into the back kitchen to tell the chef (DALL-E), who did the actual cooking. Same restaurant, different jobs, different setups.

Now, with OpenAI’s recent upgrade, ChatGPT can natively generate and edit images. Image generation is no longer handled by sending prompts to DALL-E or any external model. Instead, both text and image generation are unified within a single, multimodal architecture which handles these tasks natively, without relying on separate systems.

So now, its like you’re sitting at the chef’s counter in an open kitchen restaurant: you still speak to the server who understands your taste. But instead of running off to a different room, the server just leans over and says, “Hey Chef, give us a double cheeseburger with extra jalapenos.” Then right in front of you, the chef gets to work. Same restaurant, different jobs, same setup.

Alongside this architectural shift, image generation itself also got noticeably better. It became sharper at interpreting prompts, composing coherent scenes, and responding to nuanced or stylistic requests. Although native inpainting (advanced photoshop-level editing capabilities) continues to remain available to Plus-users, the image generation upgrades, which are now accessible to users across all tiers, have been fun to play around with.

Since the upgrade, I’ve been exploring and testing the limits of ChatGPT’s new image generation and editing capabilities as a Plus user. I ran a series of experiments to understand the strengths and limitations available at this tier. Given the depth of insights from these experiments, I've split the debrief into two parts. In Part 1 (this post), I'll cover GPT’s ability to handle procedural visuals: i.e. how well it performs in generating step-by-step instructional images that demand visual logic and precision. In Part 2 (next post), I’ll dive into GPT’s strengths in narrative creativity and examine key limitations in its visual memory.

1. Procedural Visualization – GPT’s Strengths and Weaknesses

First, I wanted to see GPT’s ability to generate clean, step-by-step visuals for real-world processes. In other words, I was testing its ability in procedural visualization: turning a simple prompt into a clear, Pinterest-style how-to board, showcasing not just a single image, but a visual sequence broken down step-by-step. For this experiment, I ran three different scenarios with GPT: preparing overnight oats, making pour-over coffee, and illustrating how to use a roman chair back extender.

Overnight Oats (Success)

Creating a visual guide for overnight oats was smoother than I'd expected. I asked GPT for a blueberry overnight oats recipe, prompting it first to share the steps and then turn them into a four-panel visual guide. Impressively, GPT quickly generated a photographic-style sequence: oats, milk, chia seeds, and blueberries, stacked neatly in clear, logical steps. There were a few minor issues, like the text beneath each panel being partially cut off or chia seeds appearing prematurely, but after a handful of iterations, GPT produced a polished visual guide.

Why did GPT handle this task so effectively? Because it excels at photographic sequencing, which involves replicating predictable real-world visuals its seen countless times in its training data. GPT is great at generating ‘accurate’ images when it recognizes familiar patterns, by leveraging the vast database of recipe blogs, Instagram posts, Pinterest images, etc that its been trained on. However, its lack of deep understanding, such as cause-and-effect relationships, spatial logic, or physics, became more and more apparent in subsequent experiments.

Coffee-Making (Somewhat Success)

Next, I tried a 2D infographic-style visual guide for making pour-over coffee. This one, however, proved trickier. Though GPT was still great at summarizing the overall step-by-step process, it struggled much more with illustrations for each step. Here’s what it came up with at first:

Right – not quite what I was looking for, although Steps 1 and 4 were actually pretty good. From there on out, this image required a lot more input and specific instructions to get Steps 2 and 3 right. It especially had a hard time getting Step 2 right as it repeatedly showed coffee already in the beaker before pouring water:

The difficulty that GPT was having illustrates its "default bias": the invisible assumption it forms from repeatedly seeing specific outcomes during training, causing it to default to those common visuals. If you do a quick google search for "pour-over coffee," you’ll see that most, if not all, images show ground coffee in the filter and brewed coffee already in the glass beaker. Presumably, GPT was trained with these images and many more like it, reinforcing this bias. And since GPT doesn’t truly understand the logic of physical processes, spatial relations, or step-by-step constraints, instead of thinking “the beaker must be empty because we haven’t added water yet,” it just tries to match common visual patterns of “coffee setup.” And turns out most often, that picture usually includes coffee already visible in the beaker.

Overcoming this default bias required active negation and precise context control. It took multiple explicit corrections (and me shouting at GPT) before finally getting the sequence right (though you’ll also notice that it struggled to get the word ‘water’ right after a series of iterations):

Back Extension Exercise (Fail)

The most challenging procedural visualization task I asked GPT to do was to demonstrate how to use a roman chair back extender.

GPT repeatedly failed, struggling severely with capturing correct anatomical positioning and motion phases in every step. Unlike overnight oats or coffee-making, this task demanded nuanced visual logic and anatomical accuracy, clearly exposing GPT’s significant limitations in generating precise physical instructions.

As you can see, there’s just… too many wrong things here. Not only did GPT struggle with fitting all four panels within the image frame, but despite my best attempts to provide detailed, panel-by-panel instructions, GPT consistently failed to illustrate subtle movement changes between steps, like noting the difference between “mid-rep” or “peak contraction.” Capturing precise body movements, especially those requiring anatomical accuracy like torso alignment or proper hinge mechanics, is difficult for GPT.

Additionally, it had trouble maintaining consistent visual identity across panels: on close inspection, you can see that the man’s face slightly changed in each step. Realizing this, I considered whether explicitly generating and "anchoring" an image of a figure first might help maintain more visual consistency. And so that’s exactly what I decided to do in the next experiment.

To be continued in Part 2!

Share

Ask any junior investment banker how often they hear this from senior management, and they'll probably say, "Every day." It encapsulates Wall Street’s ethos: avoid reinventing the wheel, use existing processes, and prioritize speed above all else.

As a former investment banker, I understand this mentality. On the surface, sticking to familiar methods feels logical. When you're facing tight deadlines and overwhelming workloads, who has the luxury to pause and wonder, "Is there a better way?" Often, the quickest and safest route seems to be using tried-and-tested legacy systems, even if they’re outdated, manual, or inefficient.

But there’s an irony here: in pursuing immediate efficiency, Wall Street inadvertently reinforces outdated processes that ultimately cost more time and energy in the long run. And it’s usually the junior bankers who bear the consequences of this paradox.

"Grunt work" — the repetitive, time-consuming, yet necessary tasks inevitably handed down to juniors — is common across any organization. But in banking, where legacy systems often mean manual execution, grunt work can quickly spiral out of control, pushing juniors toward 100-hour weeks. These extreme hours aren't healthy or sustainable, and yet they're largely accepted as the norm.

So, when I recently came across this Bloomberg article describing junior bankers' resistance to adopting AI tools, tools which could theoretically alleviate some of this burden, I was surprised. The article explained how juniors feared that relying on AI might deprive them of critical foundational skills traditionally developed through manual grunt work. Similar concerns echo among some senior managers as well.

But this prompted a deeper question for me: Is Wall Street genuinely worried about AI compromising skill development or is the industry simply hesitant to disrupt what’s familiar? In other words, is this fear truly about losing critical skills, or is it rooted more deeply in comfort with the status quo?

Navigating the AI Dilemma

The answer is likely a combination of both: a genuine concern about skill development, mixed with resistance to disrupting comfortable routines. But in today’s rapidly changing landscape, Wall Street simply can’t afford to remain comfortable. The industry must become more open to change, particularly with AI.

The fear of losing essential skills by using AI is understandable, but it’s also misguided. AI utilization itself is rapidly becoming an essential skill. The solution isn’t to reject AI outright; rather, it's to intentionally incorporate AI into banking workflows, revamping outdated processes to improve efficiency and morale while still maintaining foundational training.

To achieve this, banks must actively foster environments that encourage AI use, thoughtfully integrating it into training programs so employees gain AI fluency without compromising core skillsets. And the very first step to doing this is by reexamining the idea of “grunt work.”

Not All Grunt Work is Created Equal

When it comes to grunt work, there’s a critical distinction to be made: some tasks genuinely build core skills, while others simply drain productivity and morale.

Here’s an example from my analyst days:

One Friday afternoon, my team handed me nearly 100 excel files, each containing dozens of tabs. My task? “Sanitize” each file ahead of a deal launch the following Monday. This meant manually reviewing every tab and every cell for formatting errors like stray punctuation, incorrect font colors, or misaligned tables. My weekend disappeared into a tedious, repetitive exercise of scanning spreadsheet after spreadsheet for formatting inconsistencies.

Did I develop stronger attention to detail through this process? Sure. But was spending two days on this task genuinely valuable for my professional growth? Probably not.

Performing repetitive, low-value tasks manually doesn't meaningfully teach — it exhausts. In reality, I could’ve gained the same meticulousness and attention to detail from almost any other project typically assigned to junior bankers.

Instead, imagine if I'd had an AI assistant. A simple instruction (“Sanitize these excel files for errors, inconsistencies, or formatting mistakes”) could have transformed that weekend-long ordeal into mere hours of thoughtful oversight. Not only would I have still learned crucial skills, like effectively instructing, supervising, and validating AI-generated outputs, I would have also freed up time and mental capacity to focus more intentionally on each file’s core content, thinking through how they supported our strategic thesis and positioning for the deal. Time spent doing that kind of “grunt work” is genuinely valuable: it encourages critical thinking, builds perspective and helps junior team members see the connection between their daily tasks and the bigger picture of the project.

The reality in banking, though, is that too much grunt work involves tasks like aligning text boxes, formatting tables, or endlessly searching for higher-quality logos and images for pitch decks. These types of tasks consume hours of junior bankers’ time, providing minimal skill development but maximum stress. When juniors become overwhelmed by these menial tasks, they quickly lose sight of their contributions to the bigger picture, begin questioning the value of their work, and ultimately burn out.

These are precisely the types of grunt work that banks should delegate to AI tools immediately. But for effective AI integration, banks must intentionally identify which grunt works genuinely matter and which can, and should, be automated.

Creating an AI-Friendly Culture: Open Dialogue and Protected Trainings

When it comes to distinguishing valuable grunt work from tasks better suited for AI, the most effective starting point is fostering open dialogue across all levels. Each role, from analyst up to senior management, has unique insights on where AI could be beneficial. Analysts, closest to day-to-day tasks, often recognize inefficiencies first, while associates and vice presidents can offer valuable, experience-driven perspectives.

To facilitate this dialogue effectively, banks should establish dedicated AI Discussion Committees, structured separately for junior and senior employees. For juniors, this could mean holding sessions without senior management present, creating an environment where junior bankers can speak freely and honestly about their workloads and challenges without fear of judgement.

In these committee sessions, junior team members could bring specific examples of tasks they handle manually, highlighting opportunities for potential automation. The junior-level discussions themselves would then serve as an important check-and-balance: for example, an analyst might propose automating a particular task using AI, while associates, having previously performed similar tasks themselves, can help evaluate whether manual execution might still offer meaningful skill-building value in the long run. The committee, benefiting from collective insights, would then assess holistically whether each suggested task truly warrants AI automation or is better executed manually to preserve foundational skill development.

Yet, identifying tasks is only the first step. Banks must also proactively implement regularly scheduled AI Innovation Sessions, following feedback from these committees. These sessions would encourage junior bankers to bring forward repetitive tasks they've identified from Discussion Committees and encourage discussions around how to automate those tasks with the rest of the junior bankers across the organization.

Importantly, the goal of these sessions isn't simply to prescribe specific AI processes; AI technology evolves far too quickly for rigid approaches. Instead, these sessions should create an environment focused on continuous exploration, experimentation, and practical application of AI tools. Banks must intentionally shift the culture from seeing AI as a threat to embracing it as a new core skill. When Excel replaced manual, hand-calculated valuations done decades ago, bankers didn’t lose core valuation skills. Instead, they adapted and proficiency in Excel became essential.

Moreover, its important that these AI Innovation Sessions occur during protected time for juniors. Without clear, protected schedules, daily workloads and urgent meetings inevitably will take priority, preventing meaningful participation and creating skillset disparities across junior employees. Currently, most banks rely heavily on informal mentorship and ad-hoc, team-specific training, often leaving junior development to chance. You might luck into a supportive mentor, or you might not. To cultivate the next generation of Wall Street leaders effectively, banks need a consistent, structured, and intentional approach to training, especially as AI increasingly defines the future of finance.

Bottom Line: Don't Just Get It Done — Do It Smarter

Every organization is facing a critical moment as they navigate the rise of AI tools. Much of the focus globally has been on enhancing the technical capabilities of AI, which is undeniably essential. But what’s equally crucial is developing a workforce skilled in effectively using these AI tools.

Wall Street is no exception. Ultimately, the industry’s future success depends on nurturing a new generation of talent comfortable and competent with AI without sacrificing essential foundational skills. Banks that commit to intentional AI training now will see gains in productivity, employee engagement, and retention, as well as long-term advantages in innovation, leadership, and market competitiveness.

And at the end of the day, AI won’t replace bankers, but AI-savvy bankers will replace those who aren’t.

What are your thoughts? I’d love to hear how your organization or industry is approaching AI. Are you embracing it, resisting it, or somewhere in between?

Share

I’m a former finance professional who spent my entire career on Wall Street. I worked in investment banking and private equity, gaining hands-on experience across financial transactions like M&A, debt placements, and equity capital raisings.

That might sound like dry finance jargon, but what it really means is that I had a front-row seat to seeing how money moves at massive scale. I saw firsthand what works, what doesn’t, what the industry values, and what it lacks.

Wall Street was fast-paced, exciting, and for years I thought finance would be my lifelong career. But then came AI.

Like most people, my first exposure to generative AI was through ChatGPT, initially just to handle random personal tasks. I approached it skeptically at first, but that quickly turned to genuine amazement at its capabilities.

Thus, over the past year, I've immersed myself in AI: experimenting with LLMs, building my own custom GPT, attending international AI conferences, and studying how AI is being adopted across industries.

This journey led me to bigger questions: How will AI fit into Wall Street? Can it be fully integrated into the conservative financial system, or will it remain merely a surface-level productivity tool, never truly embedded at a deeper, systematic level?

What began as curiosity evolved into something bigger and I realized I didn't want to just watch this shift happen, I wanted to help drive it.

Why?

Because we're at a rare moment in history where AI is redefining expertise. Right now, everyone — you, your boss, your boss’s boss — is at the same starting point when it comes to understanding AI. Almost no one has a decades-long advantage. Those who choose to learn quickly now will build an invaluable AI skillset, one that positions them as frontrunners of their respective fields. They will be differentiated bridge leaders who can effectively communicate AI’s strategic value and connect their industries today with their AI-powered future.

That’s why I started The Bull & The Bot.

This substack is where finance meets AI. Here, I'll document AI’s impact on finance: exploring where it’s being adopted, where it's resisted, and what it means for professionals on Wall Street and beyond.

But this space isn’t just about AI in finance. It’s also about my personal AI learning journey. It’s my space to explore, question, and share what I discover. Not just in finance, but across industries and showing how AI touches everyday lives.

What You’ll Find Here

To keep things clear, I’ll be writing under two main series:

📈 The Bull Series

Finance, fintech, and AI’s impact on financial services — what I call 'Finaissance', the intersection of finance and technological AI renaissance. I'll analyze industry trends, fintech disruption, and AI adoption across private equity, investment banks, and the broader finance industry.

🤖 The Bot Series

My personal AI learning journey + general AI trends, news, and insights. This series will document my experiments and insights from major breakthroughs, AI-related book reviews, and reflections on how AI is reshaping industries outside of finance.

Where This Is Going

Some weeks, I’ll be writing about AI adoption in investment banking. Other weeks, I’ll break down the latest AI technologies and how I apply them to everyday life. Either way, one thing is clear: AI is evolving fast, and we must keep learning, or risk being left behind.

I don't have all the answers yet, but I'm determined to figure them out. If you're curious about AI, finance, or how to stay ahead in this shift, you're in the right place.

Subscribe, join the conversation, and let’s learn together.

Subscribe now