The New Diligence Stack

Most AI-in-diligence content is underselling the change.

The common pitch is: the memo gets faster. The CIM summary is cleaner. The market map takes 20 minutes instead of 6 hours. The customer call notes get organized automatically.

All true. Also not the interesting part.

That is just labor substitution.

The deeper shift is that agents make new forms of diligence possible — workflows that were either too expensive, too slow, too fragmented, or too annoying to run in the pre-agent era.

That is the frame I care about now:

What kinds of business and investment diligence can we do now that we realistically could not do before?

Not because the models are magical. Because the operating model changed.

A good agent system can decompose work, pursue parallel lines of inquiry, preserve state, challenge its own conclusions, and keep running after the first memo is written. Once that exists, the diligence process stops being a one-time document exercise and starts looking more like a living research system.

Here are the seven categories that feel genuinely new.

1. Adversarial memo generation at scale

Pre-agent diligence usually produced one main synthesized view at a time.

You had the deal team memo, maybe an IC pre-read, maybe a consultant workstream, maybe a few pointed comments from a partner who had seen the movie before. But running ten distinct, high-quality analytical frames against the same company was expensive and organizationally awkward.

Agents change that.

You can now set up structured opposition by default:

• one agent builds the best long case
• one builds the best short case
• one acts as the null hypothesis and asks whether the whole debate is overstated
• one focuses only on customer concentration and revenue quality
• one focuses only on management incentives and governance
• one tries to falsify every major assumption in the base case

Then you force evidence-backed rebuttals across rounds instead of accepting the first polished answer.

That matters because many of the worst diligence misses are not caused by lack of information. They come from premature convergence around a coherent narrative.

The banker has one. The management team has one. The deal team often forms one too early. A single LLM memo, no matter how elegant, tends to reinforce that convergence.

A multi-agent setup can be designed to do the opposite.

Not “be more critical.”

Structurally create criticism.

That is different.

2. Dynamic evidence pursuit instead of static document review

Traditional AI workflows are still mostly passive. You upload a dataroom. The model reads what you gave it. It summarizes, organizes, maybe flags inconsistencies.

Useful, but still fundamentally document-bound.

Agentic diligence can be configured to work more like a hungry associate:

• identify the 5 claims that matter most
• map what evidence would actually validate or break each one
• detect what is missing
• generate a next-request list
• route targeted follow-ups by function: finance, sales, ops, product, legal
• revise the thesis after each new evidence pack arrives

This sounds small until you realize what changes.

The process is no longer “read everything and synthesize.”

It becomes:

1. Form a view.
2. Identify the load-bearing assumptions.
3. Pull the next best piece of evidence.
4. Update confidence.
5. Repeat.

That is much closer to how good investors actually work.

And pre-agent, it was hard to maintain that level of organized persistence across dozens of micro-questions without burying a team in project management overhead.

3. Stakeholder simulation before the world forces you to learn

This is where things start to feel genuinely non-obvious.

A lot of important diligence is not just “is the model right?” It is “how do different groups react if this company changes?”

That includes:

• customers after a price increase
• employees after a cost-cutting plan
• channel partners after a strategy shift
• regulators after a category incident
• media after an acquisition announcement
• power users after a product simplification

Pre-agent, you could approximate this with consultant interviews, expert calls, surveys, or pure judgment.

Now you can run structured stakeholder simulations using hundreds of personas with different priors, incentives, and sensitivities — then study where opinions converge, where they do not, and which narratives are most fragile.

This is not a crystal ball. Anyone presenting it that way is selling theater.

But as a stress-testing layer for narrative durability, it is surprisingly powerful.

I think of it as:

Monte Carlo for human reaction.

That is interesting for M&A integration, consumer brands, politically exposed industries, pricing decisions, and any deal where second-order reactions matter as much as the spreadsheet.

4. Diligence graphs that connect scattered facts across workstreams

In most processes, the most valuable facts live in separate silos:

• finance has the margin bridge
• customer calls have the qualitative complaints
• product demos reveal hidden implementation friction
• legal has the contract exceptions
• ops surfaces the single-point-of-failure vendor
• market research explains why one cohort is weakening

Humans are bad at holding all of that in working memory at once.

Agents are useful here not because they “know more,” but because they can keep building a machine-readable map of claims, evidence, contradictions, and unresolved questions.

Imagine a diligence graph where:

• every major thesis node has attached support and contradiction
• each claim is scored by evidence quality, not rhetorical confidence
• unresolved items stay live until explicitly closed
• one new customer interview can automatically downgrade three other assumptions
• recurring themes across management calls, support tickets, and cohort behavior are linked instead of living in separate notes

Pre-agent, this kind of cross-workstream integration was theoretically possible but operationally painful. In practice, it depended on one unusually organized person.

With agents, it becomes a system property.

5. Always-on diligence after the deal, not just before it

This may be the most important one.

Most diligence is still event-based. A process starts. Everyone sprints. A memo is written. The transaction closes or dies. Then the machinery winds down.

But businesses do not move on process timelines.

The best modern diligence stack should not stop at signing. It should become a persistent monitoring layer:

• track competitor launches
• watch hiring patterns
• monitor customer sentiment drift
• catch regulatory changes early
• compare management promises against operating reality
• flag when the original thesis is strengthening versus quietly breaking

That means the line between diligence and portfolio monitoring starts to disappear.

Pre-agent, this was too labor-intensive unless you had a large team or paid a lot for bespoke research support.

Now a smaller team can maintain institutional vigilance without treating every update like a fire drill.

That is not just efficiency. It changes how much of the company you can actually see over time.

6. Cheap parallelization of weird questions

This one sounds minor, but it compounds.

In many live deals, there are 20 questions that are important enough to matter and annoying enough that nobody fully chases them.

Things like:

• how unusual is this pricing architecture really?
• what do ex-employees consistently complain about?
• does the language in customer reviews shift after implementation?
• which niche competitor is quietly winning among sophisticated buyers?
• what operational detail do customers praise before churn drops?

Each of these might not justify a dedicated workstream.

Together, they often determine whether you actually understand the business.

Agents make it practical to run lots of these narrow, odd, high-signal sub-analyses in parallel without blowing up team bandwidth.

That matters because real edge often hides in the corners of a business, not in the obvious tabs of the model.

7. Living diligence systems for proprietary domains

The final shift is architectural.

The old model was a project. The new model can be a reusable system.

Once a firm knows the recurring questions it cares about — revenue quality, pricing power, service fragility, channel conflict, retention durability, regulatory exposure — those can be encoded into reusable agent workflows.

Over time, the stack gets better because it accumulates:

• prior deal patterns
• sector-specific heuristics
• known failure modes
• preferred evidence hierarchies
• internal judgment about what actually mattered in past wins and misses

So the point is not just “use AI on this deal.”

It is:

Build a diligence machine that gets sharper every time you use it.

That was hard in the pre-agent era because the knowledge mostly lived in people, email threads, and half-remembered war stories.

Now more of it can live in process.

What the setup actually looks like

This is where a lot of writing on this topic falls apart. It stays abstract.

A credible setup is not “one super-agent does diligence.” That is demo logic.

A more realistic architecture looks like this:

Layer 1: Ingestion

• CIMs, decks, QoE, customer lists, contracts, transcripts, support tickets, product reviews, market data
• documents chunked and indexed by source type, date, and reliability
• claims extracted into a structured evidence map

Layer 2: Specialized agents

• thesis builder
• bear case builder
• null / anti-hype reviewer
• evidence auditor
• customer voice summarizer
• pricing and unit economics checker
• management consistency tracker
• follow-up request generator

Layer 3: Debate and scoring

• agents respond to each other across rounds
• unsupported claims get penalized
• repeated arguments without new evidence lose weight
• confidence is tied to evidence quality, not polish

Layer 4: Human judgment

• senior investor reviews only the load-bearing questions
• focuses attention on what could actually kill or reprice the deal
• uses the agent output as attack surface, not gospel

Layer 5: Persistent monitoring

• after close or after passing on a deal, keep selected watchlists live
• compare new evidence against original thesis map
• surface drift early

That is a system. And systems matter more than prompts.

The risk: AI theater dressed up as rigor

There is also a fake version of this trend.

You will see a lot of companies describe “agentic diligence” when they really mean one of three things:

• better summarization
• prettier workflow automation
• role-play theater with no evidence discipline

That is not enough.

If the process does not improve falsification, evidence routing, contradiction handling, and monitoring persistence, it is probably just a nicer wrapper around the same old memo.

The question I would ask any founder or internal team building this is simple:

What can this workflow now do that a smart associate + ChatGPT + a checklist could not do last year?

If the answer is vague, the product is probably cosmetic.

My actual view

I do not think agents replace investor judgment.

I think they change the shape of where judgment matters.

The old world used humans for everything because the work was too unstructured to modularize well.

The new world should use agents for:

• breadth
• persistence
• parallelization
• contradiction surfacing
• workflow memory

And reserve humans for:

• directing attention
• framing the real question
• deciding what evidence matters most
• knowing when the story is too clean
• making the final call under uncertainty

That is the real opportunity.

Not a faster memo.

A new diligence stack.

And the firms that understand that early will not just save time. They will see things that were previously uneconomic to see.