How we run due diligence with AI (and why the model is still ours)

A private investment firm called us on a Friday. They were putting a bridge convertible loan into a company they already knew well, but they still had to get comfortable with the latest data room before they committed, and they needed it by Monday. They didn't want a summary, they wanted to understand the whole machine: how money came in, where it leaked out, and what the numbers were quietly assuming.

The traditional answer to that request is a team of analysts and two to three weeks. We did it over the weekend, and by Sunday night the firm had a working financial model, a mapped-out view of every moving part of the business, and a full list of questions ready for a call with the management team on Monday.

People assume the trick is that AI "did the diligence." It didn't. What actually happened is more interesting, and it's worth explaining, because it's how we now run most of our analytical work.

The principle: parallel, not instead

There are two ways to use AI in finance work. One is to hand it the problem and hope the answer comes back right. We don't do that, and you shouldn't let anyone who touches your numbers do that either.

The other way is to work in parallel with it, where the AI does the reading, structuring, and building while we spend that time on the parts that require judgment, like what matters, what's missing, and what the numbers are really saying.

The rule underneath all of it: we control the analysis file from A to Z. Every formula, every assumption, every number that ends up in front of a client has been placed there deliberately and reviewed by a CPA. AI accelerates the work, but it doesn't get a vote on the conclusions.

How it actually works

Here's the workflow on a diligence project, stripped of the details that vary deal to deal.

First, AI builds the model, not the answer. We use Claude to construct the analysis file in Excel: the structure, the schedules, the formulas, the tie-out checks that confirm everything reconciles. This is the work that used to eat the first week of any engagement, and it's exactly the kind of work machines are good at. It's also the easiest to verify, because a formula is either right or it isn't.

Then the data loads into the model. Trial balances, ledger exports, contracts, whatever the data room holds. AI can read a large volume of documents in hours and put every number where it belongs in the structure we already approved. The volume problem, which is what usually makes diligence slow, stops being a problem.

Then the human work starts. With the mechanical layer done in a fraction of the usual time, we spend our time where it counts: pressure-testing assumptions, chasing anomalies, asking why the working capital moved the way it did, deciding what a buyer should worry about. This is the part of the engagement clients are actually paying for, and it's the part that gets more time in our process, not less.

Why Excel, still

It would be fashionable to say we built a proprietary platform. We didn't, on purpose.

The deliverable is an Excel file because our clients can open it, trace any number back to its source, change an assumption, and watch the model respond. Their auditors can review it. Their board can hold it. When the engagement ends, they own the file and everything in it.

A black box that produces answers is worthless in a negotiation, but a transparent model you can defend line by line is leverage.

What AI doesn't do here

This is the part most AI conversations skip, so we'll be specific. In our process, AI does not decide what's material. It doesn't set assumptions. It doesn't write conclusions. It doesn't produce a number that goes to a client without a human who has done this for twenty years looking at it first.

That's not caution for its own sake. It's the honest boundary of what these tools are good at. AI is remarkable at reading, structuring, and building. Judgment about a business, what a normal margin looks like in this industry, which customer concentration is a real risk, when a seller's adjustment deserves a hard look, still comes from experience. The firms that get this backward will learn it expensively.

What this means for you

If you're evaluating a deal, closing a slow month, or trying to understand your own numbers with a team that's stretched thin, the math has changed. Work that used to take weeks of billable analyst time now takes days, and the quality goes up, because the senior person's time shifts from assembling data to interpreting it.

That's the trade we offer: the speed of AI, the control of a hand-built model, and the judgment of someone who has sat in the CFO chair through audits, acquisitions, and everything in between.