Fewer deals, bigger checks: why speed stopped being the diligence bottleneck

Something strange happened to private equity in the first half of this year, and most of the software being sold into the industry has not noticed.

According to PwC's midyear outlook, the first half of 2026 produced roughly two-thirds fewer private equity transactions than the same period in 2025. That sounds like a collapse. It isn't. Over the same period, aggregate deal value went up by about a tenth. Firms did not stop buying. They stopped buying broadly. The capital that used to be spread across a dozen mid-sized bets is now concentrated into a handful of larger ones.

Ask any deal partner what that feels like from the inside and you get a version of the same answer: every deal now carries more weight. When you close twenty deals a year, a bad one is a line item. When you close four, a bad one is the vintage.

This is the context in which an entire category of software is selling itself on speed.

The answer to a problem we no longer have

Read the marketing for AI diligence tooling and you will find the same promise in eight different fonts: summarize the data room in minutes, not weeks. Ingest ten thousand documents overnight. Compress the first pass.

That promise was well-aimed in 2023. Deal volume was high, teams were thin, and the binding constraint on a diligence process really was throughput — how many documents an analyst could get through before the exclusivity window closed. Speed was the right thing to sell because speed was the thing that was scarce.

It is not what is scarce now.

A firm running four high-conviction processes a year is not lying awake about whether the first pass takes nine days or three. It is lying awake about whether the first pass asked the right questions. The scarce resource has moved from reading capacity to conviction — and those two things are not improved by the same tool.

There is a second reason speed has stopped being the useful metric, and it comes from an unlikely source: our competitors' own research. Hebbia commissioned a survey this spring of several hundred US finance professionals and found that while they trust AI for research more than they trust many of their colleagues, they verify its output anyway, at the operational level, before anything reaches a decision. Every serious diligence guide published in the last year says the same thing in prescriptive form: no AI finding reaches an investment committee memo without an analyst confirming it against the source document.

Sit with the arithmetic of that for a moment. If a model produces a first pass in three minutes, and a human then has to trace every material claim back to a page in the data room before it can be used, the three minutes were never the expensive part. The verification was. Selling a firm a faster generator without changing anything about verification is selling them a faster way to create work.

What actually goes wrong

Deals do not blow up because someone read the data room slowly. They blow up because of something that was in the data room and nobody thought to look for, or something that was conspicuously not in the data room and nobody thought to ask about.

A customer concentration problem that only appears when you cross-reference the revenue schedule against the contract expiry dates rather than reading either one on its own. A management team whose tenure numbers look fine until you notice that everyone senior arrived in the same eighteen-month window after the last sponsor took control. A working capital pattern that is unremarkable annually and alarming monthly. A supplier agreement whose change-of-control clause is standard, and a second one, four folders away, whose is not.

None of these are retrieval failures. Every one of them is a question failure. The information was sitting there. The model was never asked the thing that would have surfaced it — and a language model, however capable, answers the question you put to it. It does not know which question you should have asked about this business, in this sector, at this hold stage, given what went wrong on the last three deals your firm looked at in this space.

That knowledge exists. It just does not live in the model. It lives in the heads of the four or five people at your firm who have been doing this the longest, and it leaves the building when they do.

Which points at the capability that actually matters, and it is not summarisation. It is whether a system can look at a data room and tell you what is missing — which questions this business ought to have been asked and wasn't, given its sector, its size, and what your firm has learned from the last nineteen deals that looked like it. A summary tells you what is there. A gap is the thing that costs you money.

The uncomfortable implication for everyone selling AI to dealmakers

Here is a thing we say to prospects that our sales team would prefer we did not put in writing: most of what AI diligence platforms do is now a commodity.

We know how this lands, because a prospect said it back to us on a call recently, and more concisely than we would have. He summarised our product as a few templates that draft an investment memo from data room materials, plus a chatbot to ask follow-up questions. He was not being unkind. He was describing what he had seen, accurately, and checking whether there was more to it.

He was also describing roughly every product in this category, including several that have raised a great deal of money.

Extraction is a commodity. Summarization is a commodity. Cross-document retrieval, citation-linking, generating a clean first-pass memo from a messy data room — a general-purpose frontier model with a competent wrapper does all of it, and does it better every quarter, at a price that trends toward zero. Any vendor whose differentiation is "we can read the data room" is selling something the model providers will give away.

What does not commoditize is the question set. The accumulated, firm-specific, sector-specific, painfully-acquired knowledge of what to interrogate — which is precisely the thing that no general model has, because it was never written down anywhere for the model to learn from.

That is the asset. And it is an asset with a strange property: it belongs to your firm, it should never leave your firm, and it gets more valuable the more deals it sees. The twentieth deal a firm runs through a system that remembers should be diligenced better than the first, not because the model improved, but because the firm's own questions accumulated.

We call this firm memory. It is the only part of this stack we think is genuinely defensible, for us or for anyone.

Memory you cannot load is not memory

There is an unglamorous precondition here that took us embarrassingly long to treat as strategic.

A firm cannot have memory of deals it was never able to load. If a system's knowledge of how your firm thinks has to be assembled one deal at a time, starting from zero, on the day you sign, then the memory is a promise about year three rather than a capability you have in month one. Meanwhile the four or five people whose heads it is currently stored in are still the system of record, and still a flight risk.

So the questions that sound like IT plumbing — can we import six years of deal folders, can we point it at a Dropbox directory, can a zipped data room go in whole, can it read a SIM and populate the deal record itself — turn out to be the questions that decide whether any of this is real. We have heard every one of them from buyers in the last month, usually phrased apologetically, as though asking about import was a distraction from the interesting part.

It isn't the distraction. It's the entry condition. The interesting part is downstream of it.

What to measure instead

If you are evaluating diligence tooling this year, we would gently suggest that time-to-first-pass is the wrong number to put at the top of the scorecard. It is easy to measure, every vendor will happily compete on it, and it correlates weakly with whether you make good decisions.

Five questions we would ask instead, including of ourselves:

1. What did it ask that we didn't? Run a platform against a deal you have already closed — ideally one that disappointed. The useful output is not the summary. It is the list of questions the system raised that were not on your original request list, and the gaps it flagged in what the seller provided. If that list is empty, the tool is a faster typist. And trace one of those findings back to its source document while you are there; the length of that path is where your analysts' time actually goes.
2. Can it ingest what we already have? Six years of deal folders, in the mess they are actually in. If onboarding starts from an empty account, so does the firm memory, and the compounding argument is a story about a future you may not reach.
3. Where does the question set live? If the intelligence sits in the vendor's prompt library, it sits in every other customer's too, including your competitors'. Ask specifically whether your firm's accumulated questions are yours, and get the answer in writing.
4. What can we take with us if we leave? A buyer put this better than we do: how much switching power are you giving the customer? Ask for the export format before you sign, not after. A vendor confident in the product is comfortable with the answer.
5. Does deal twenty go better than deal one? If the system has no mechanism for retaining what your firm learned on the last nineteen, it is not a diligence platform. It is a document reader with good marketing.

Fewer deals means each one matters more. That is not a market condition that rewards reading faster. It rewards being right — and being right is a function of the questions you thought to ask, which is a function of what your firm has learned and managed to keep.

We are not building a faster way to get through the data room. We are building the thing that remembers what to ask when you get there.

HuxleyIQ builds AI diligence infrastructure for private equity, with firm memory as the core primitive.