There are two completely different questions hiding inside the phrase "AI due diligence," and the large consultancies have mostly answered the easy one.
The easy one is: can we use AI to run this diligence faster? Spread five years of financials in an afternoon, summarize the data room, draft the IC memo, flag the obvious anomalies. That work is real and it's getting commoditized quickly — every firm of any size now has a diligence-acceleration toolkit, and the differentiation between them is shrinking by the quarter.
The hard question is the one that actually decides whether you should write the check: does this target's advantage survive an AI-enabled market? Not "is there an AI use case here," but "when the cost of the thing this company sells — the analysis, the routing, the inspection, the scheduling, the customer service — falls toward zero for everyone, what is left that this business can still charge for?"
A moat you can describe in a CIM and a moat that survives commoditization are not the same moat.
Why the question is getting sharper, not softer
For most of the last decade, technology diligence in the lower-middle market was a checklist exercise: is the stack modern, is the team competent, is there crippling technical debt. AI breaks that frame because it doesn't just change a target's internal efficiency — it changes the external price of the value the target creates.
A business whose margin comes from doing something faster or more cheaply than its customers could do themselves is exposed the moment that "something" becomes a feature in a model anyone can rent. The financials look fine right up until the renewal cycle where three competitors quietly matched the core offering at a fraction of the price. That risk doesn't show up in a quality-of-earnings report. It shows up eighteen months into the hold, in churn.
What a survivability read actually examines
This isn't speculation dressed up as analysis. It's a structured read on a handful of things that are knowable before close:
- Disruption exposure of the thesis. Which parts of the revenue base sit on top of work that is becoming cheap for everyone, and which parts rest on something durable — proprietary data, physical assets, switching costs, regulatory position, a distribution advantage that AI doesn't touch.
- Data and model maturity of the target. Whether the company's own data is an asset it can actually compound on, or a liability it has been quietly ignoring. Most mid-market targets have never had this looked at honestly.
- Governance and regulatory red flags. Shadow AI in the business, unmanaged model risk, exposure under emerging rules. In regulated end-markets, this is the difference between a clean exit and a re-trade.
- The post-close value plan. The flip side of risk: where AI is a genuine lever for this specific business, sized and sequenced, so the operating partners have a plan on day one rather than a workshop in month nine.
Why it has to come from an operator
You cannot do this read from a desk with a generic framework, because the answer is always specific to the floor. Whether an industrial distributor's advantage survives depends on the actual mechanics of how it wins business — the quoting, the inventory position, the relationships, the service. That requires someone who has stood in the building, can talk to the engineers in their language, and can also sit in the IC and say, in plain terms, "this part is durable, this part is rented, and here's the number."
The framing matters more than it sounds. Sponsors are not short on people who can run AI through diligence. They are short on people who can tell them whether the asset they are about to own is on the right side of the disruption — and then go help make it so after close. That second half is what turns a one-time diligence fee into a relationship across the portfolio.
Diligence at the fund. Embedded leadership across the portfolio. Same judgment, applied twice.
The survivability question is uncomfortable precisely because it can kill a deal that looks good on paper. That's the point. The whole job of a reference point is to tell you where you actually are — not where the pitch says you are.