Leanrs

Almost every research team we speak with believes they’ve already “adopted AI.”

When you look closely, that usually means summarizing PDFs, transcripts, searching news, asking ad-hoc questions or rewriting notes. Useful, sure. Transformative, no.

We see this gap repeatedly because we sit inside real workflows. We talk to buy-side teams, boutique funds, family offices, and institutional research groups every week. What’s being said externally about AI adoption and what’s actually happening internally are very different things.

The uncomfortable truth: most analysts are using AI as a convenience layer, not as a structural change to how research gets done.

The Wrong Questions are Being Asked

“Will AI replace analysts?”, “Can it do end-to-end research?”
The real question is: what parts of analyst work should humans no longer be doing at all?

Very few teams are willing to ask that question honestly.

As a result:

Data extraction from filings is still copy-paste work.
Analysts still build financial models from scratch, cell by cell.
Quarterly updates are still manual and time-consuming.
News tracking is either manual or tied to expensive subscriptions.
AI tools exist, but no firm-wide frameworks or workflows exist to use them properly.

AI has dramatically lowered the cost of automation and integration, but most research teams are still operating as if nothing fundamental has changed.
‍

Where AI Clearly Works and Everyone Agrees

There are areas where AI is already incredibly effective, and almost no one disputes this:

Summarizing earnings calls, transcripts, filings, and news
Doing deep background research on industries or themes
Pulling together relevant metrics and data points
Understanding business models and drivers
Creating first-pass company descriptions

These use cases work because they’re bounded, contextual, and don’t require final judgment. Nearly every analyst today uses AI for some version of this, and the productivity gains are real.

But this is only the surface layer of research.
‍

The Problem With “End-to-End AI Research”

At the other extreme, AI tools, finance influencers, and course sellers frequently promise end-to-end AI research: full reports, full models, full investment theses.

They usually add a small disclaimer somewhere: outputs may be inaccurate or hallucinated.

That disclaimer matters more than the demo.

Investment research requires large context, historical continuity, nuanced judgment, and accountability. As context increases, AI costs rise quickly. Even if you’re comfortable paying those costs, you still can’t take raw AI output and send it to a portfolio manager without review.

If you do, you’re importing risk, not saving time.

Ironically, the more important the decision, the more essential the analyst becomes.

AI can attempt almost anything you ask it to do, even vaguely. But once tasks become even moderately complex, accuracy collapses unless the problem is precisely framed. That reality reinforces, not replaces, the role of analysts.
‍

Where AI Delivers the Highest ROI

AI performs best where tasks are well-defined, repetitive, verifiable, and structured.

For example: extracting financial data from multi-year PDFs, normalizing disclosures across formats, arranging data into predefined templates, and creating clean first drafts for models or notes.

Here’s what this looks like in practice inside real research workflows: instead of an analyst spending hours pulling numbers, cleaning tables, and re-keying disclosures into a template, AI produces a structured first pass and the analyst spends their time verifying, reconciling, and applying judgment. The work shifts from manual production to review and decision support.

When an analyst knows how to give precise instructions, AI can save hours and compress days of work into minutes.

But this changes the analyst skillset.

Modern analysts need more than financial intuition. They need the ability to break problems into logical steps, frame instructions clearly, review long outputs patiently, and validate results rigorously.

In many cases, the ideal workflow looks like this:

AI produces a strong first draft. Analysts review, correct, and enhance. Humans make the final judgment call.

That’s not failure. That’s leverage.
‍

Why Full Model Automation Is Still Hard

Financial model updates, especially complex or custom ones, remain one of the hardest problems to automate cleanly.

AI can assist. With investment and effort, parts can be automated. But today, for many edge cases, it’s still faster and safer for experienced analysts to handle complicated updates manually.

This isn’t a contradiction. It’s reality.

AI reduces effort where structure exists. Humans step in where judgment, exception handling, and accountability matter.

End-to-end AI investment research workflows are not feasible today. Hybrid workflows are.

Why Small Teams Are Pulling Ahead

The most advanced AI usage we see isn’t inside large banks or massive KPOs.

It’s inside small, focused research teams.

Not because they have better models. Because they have less friction.

Small teams can do four things quickly: pick one workflow, standardize it, automate the repetitive steps, and ship the output into the team’s existing template. If it works, they scale it. If it fails, they delete it and move on.

Large organizations struggle to do the same thing. Every workflow is slightly different. Every output format has exceptions. Every automation touches multiple systems. And every change requires approvals, controls, and alignment across stakeholders.

KPOs have an additional structural problem: their commercial model rewards hours. If automation reduces analyst hours, it reduces revenue. So even when “AI adoption” is discussed, it rarely translates into redesigning workflows in a way that meaningfully cuts labor.

That’s why many large clients we speak with see no real reduction in headcount or turnaround time despite years of “AI adoption.” The constraint isn’t technology. It’s incentives and execution speed.
‍

The Internal Build Trap

Despite real progress in AI, internal adoption remains difficult.

There are too many tools, too many models, too much noise. Teams subscribe to products like ChatGPT, Claude, Perplexity, Hebbia, or Shortcut and assume that buying access equals adoption.

It doesn’t.

Most teams get stuck in the same loop: they pilot AI on unstructured work, see inconsistent results, and then conclude “AI isn’t reliable.” The real issue is that the workflow was never made reliable. Reliability comes from structure: clear inputs, defined outputs, templates, and checks.

Then comes the second trap: “We’ll build it ourselves.”

Building a demo is easy. Keeping it working across real-world variation is the hard part. Filings change formats. Tables break. New edge cases appear. One-off exceptions become permanent rules. And suddenly the tool needs constant maintenance to avoid silent errors.

That’s why internal AI succeeds only when it’s treated like a workflow system, not a one-time project: narrow scope, strict templates, verification steps, and someone responsible for upkeep.

You don’t need elite AI researchers. You need people who understand the research workflow, can connect systems, and can enforce structure so outputs stay consistent over time.
‍

How Lean Research Approaches AI Differently

At Lean Research, AI adoption starts with a simple question:

How much of this work should not require a human?

Every new engagement begins by breaking workflows into parts:

What can be automated reliably?
What needs structured AI assistance?
What requires analyst judgment?
How many analysts are actually required after automation?

Our analysts are AI-enabled by design. They understand both the analytical problem and how to solve parts of it with modern tools, whether that’s prompting, workflow automation, or integration using tools like n8n.

The result isn’t flashy. It’s measurable:

Lower costs. Faster turnaround. Higher consistency. Cleaner outputs. Better decisions.

That’s real adoption.
‍

What Real AI Adoption Actually Looks Like

Real AI adoption in investment research is quiet.

It doesn’t show up in press releases or demos.

It shows up in fewer hours wasted, cleaner models, faster updates, and sharper questions.

The teams that win won’t talk much about AI.

They’ll just look unusually efficient.

And that’s the point.

‍

Most Research Teams Think They’re Using AI. They’re Not.

Need a Financial Model Tailored Just for You?

Looking for deeper insights, extended data, or custom research? Let us create the perfect solution for your unique needs.