The Traceable Thesis: artificial intelligence across the private equity investment lifecycle

Artificial intelligence in private equity compresses the work around a decision, and the return comes from what the firm can still reconstruct afterwards.

Nothing in this article constitutes legal or investment advice. It is offered to support discussion and further review with qualified advisers.

What does artificial intelligence in private equity actually deliver?

Artificial intelligence in private equity compresses the information work around an investment decision, and the clearest gains sit in AI due diligence, deal sourcing algorithms and portfolio monitoring. The return depends on traceability, because a firm that lets a fluent synthesis stand in place of its evidence owns a conclusion nobody can reconstruct.

• Adoption across financial services is broad and thin at once, and the enterprise-scale figure is the more informative one.
• No independent study yet shows that artificial intelligence improves private equity returns.
• UK supervision is technology-neutral, which lightens the rulebook and raises what a firm must be able to show.
• Five of the six ways this fails inside a fund produce no signal at the time.
• A source-linked record converts verification from re-performance into spot-checking, which is what makes a use case pay.

An investment committee meets, the papers are excellent, and a decision is taken. Six months later a portfolio company misses its plan, and nobody can say which diligence findings carried the vote or who held the contrary view. The memo survived. The reasoning evaporated.

That pattern predates the technology, and a system that produces a fluent synthesis of a data room in minutes makes it worse before it makes it better. In November 2024 the third Bank of England and Financial Conduct Authority survey put usage at 75 per cent of 118 responding firms. Against that, a survey of 200 fund and operating leaders published in May 2026 by an advisory firm put the share working at enterprise scale at 7 per cent. Adoption is broad and shallow at once.

A general partner deploying it is making three separate bets. The first is that the technology compresses the information work around a decision, and the available evidence supports that. The second is that compressed work improves investment outcomes, which the evidence has yet to show. The third is that the firm can describe what it does without misdescribing it, and enforcement has already landed there.

My paper argues a position that runs against the sales cycle. Buy compression, pay for traceability, and treat any workflow that cannot produce a source-linked record as unfinished, whatever its speed.

Why has the advantage moved from holding the data to reconstructing the decision?

For two decades the analytical edge in private markets came from structured data, and competition removed it, since every serious manager now buys comparable databases and deal sourcing algorithms from comparable suppliers. What changed is the ability to read unstructured material with tolerable reliability. A data room is unstructured, and so are board packs, contract sets and a committee’s own history of theses and post-mortems. Compression is the honest word for the gain, because it makes it possible to ask, across 3,000 documents in an afternoon, a question nobody would have bothered to ask.

Each paper in the series uses the same five-stage cycle, because a professional decision follows that sequence whether or not anybody records it.

• Gather, where material is collected from data rooms, management and markets.
• Ground, where each assertion is bound to the document it rests on.
• Weigh, where a named professional judges what the evidence supports.
• Decide, where the committee acts and the reasoning is written down.
• Compare, where outcomes are read back against the recorded reasoning.

Funds perform the first and fourth stages well. Grounding, weighing and comparing are where institutional knowledge accumulates or evaporates.

Does artificial intelligence improve private equity returns?

The most substantial quantitative work I located segments 471 private-equity-backed companies by artificial intelligence maturity, and reports a median revenue multiple of roughly 31 times for the highest maturity group against roughly 13 times for the lowest. Read carefully, that is an association between two things which both correlate with company quality, because better-run companies adopt operating technology earlier and command higher multiples. The authors make no claim to have separated them. The research comes from a firm that sells transformation advice.

I looked for a counterweight from a source without a commercial position and found none. That absence should shape how a partnership talks to its investors, since a claim of improved returns currently has no published basis. General partners are themselves more sceptical than the marketing, and 39 per cent of the 103 professionals in one outlook survey expected no material impact on portfolio companies during 2026.

What do UK supervisors expect from a fund using artificial intelligence?

The FCA states that it plans no extra regulations for artificial intelligence and will rely on existing frameworks, among them the Consumer Duty and the Senior Managers and Certification Regime, which locate responsibility in named individuals. That is a lighter rulebook and a heavier evidential burden, since the firm must show an existing outcome was met. There is no box to tick. My own view, going beyond what the FCA states, is that supervisory interest would open by asking which senior manager owns the output.

The specific instruments either sit outside the activity or have yet to exist. The July 2026 consultation on a new UK regime for alternative fund managers says nothing about artificial intelligence, and the ESMA guidance of May 2024 is scoped to MiFID II retail services. Enforcement has fallen in one place only. In March 2024 the SEC charged two investment advisers over false artificial intelligence claims, with penalties of $225,000 and $175,000, and every matter since has concerned a description of capability.

Where does AI due diligence fail without anyone noticing?

The failures fall into six recognisable patterns, and the grouping is my own. Confidential material enters a system outside the firm’s contractual perimeter. A system answers accurately on the material it was given while nobody establishes what it was never given, which is coverage illusion. A fluent summary circulates until the evidence behind it stops being read. Retrieval surfaces material across boundaries a person would have been walled from. A supplier update changes behaviour mid-transaction, yet later output is compared with earlier as though nothing moved. Capability overstatement is the sixth, and the only one that eventually announces itself.

Coverage illusion is the likeliest to affect a decision and the hardest to catch, because a system asked whether a data room contains a class of risk answers about the data room, while the partner believes the question was whether the risk exists. The remedy is a design requirement. Any system used in AI due diligence should report its own denominator, naming the documents it read, those it could not parse, and the counterparties it found nothing for.

How should investment committee governance handle a generated output?

The organising question is who bears the cost when an output is wrong. A memo summarised badly costs an hour, a diligence finding missed badly costs a fund, and tiering by consequence gives a model people can apply without convening a committee. Four tiers run from public material at tier one up to investment decisions and investor communications at tier four, which require a senior accountable owner, independent verification and a tested rollback.

Six technical controls carry most of the assurance burden. Five of the six are restrictions, but the fourth is a record, requiring every production system to capture the model, prompt, retrieval corpus, permissions and evaluation version behind each output. My own view is that the record does more work than the other five together, because a fund able to reconstruct which model and which documents produced a finding can investigate an error. A fund that cannot is left arguing about recollection.

Verification is where the cost lands, so source links decide whether a use case pays

Most business cases count the hours saved in production and stop there. The decisive cost arrives later. Generation becomes nearly free, and the binding constraint moves to verification, performed by the senior people whose time is the most expensive in the firm. Compare the cost of producing the work by hand against the cost of generating it plus checking it, and fund the cases where the second total comes in below the first.

Source links stop being a nicety on that reading, since they convert verification from re-performance into spot-checking. A reviewer opens three citations and knows. Checking an unsourced synthesis means reading the data room, which is the work it replaced. A recommendation with no stated basis cannot be checked until the position is realised.

The delivery sequence runs 90 days and then nine months, beginning with an inventory of approved and unapproved use, baselines and three pilots, each carrying a written stop condition. A programme shows it is evaluating when it stops something.

One decision is worth taking in the room. Name the person who will own the decision record, because tooling and vendor relationships can sit anywhere, and in most partnerships nobody owns that record today. It is the same record that makes investment committee governance defensible and turns LP reporting and AI from an awkward question into a prepared answer. Whether broader coverage produces better decisions, or simply more findings to discount, is unresolved, and a firm with three years of decision records could answer it within one fund cycle.

Frequently asked questions

Where does artificial intelligence add the most value in a fund?

Diligence and deal sourcing carry the clearest return, and respondents in an outlook survey of 103 investment and investor-relations professionals placed them at the top. The measured gain concentrates in coverage, since teams tend to put the time saved into asking more of the same window.

Is AI due diligence reliable enough to act on?

A retrieval system reports accurately on the documents it was given, and a partner usually hears a wider claim than that. The control that closes the gap is a coverage statement, naming the documents read, those that could not be parsed, and the counterparties for which nothing was found.

What should a general partner tell limited partners about artificial intelligence?

Describe what the firm does, and keep a register recording which claims have been approved for external use and who owns the evidence for each. Enforcement so far has punished false statements about capability, and answers on LP reporting and AI deserve the same approval route as a performance claim.

How does portfolio monitoring AI change a board pack?

Retrieval across a company’s own reporting history can compare this month’s explanation of a margin movement with the six that preceded it, and produce the question a chair should ask before the meeting. My own position is that such a system should point at a financial figure and be prohibited from restating one.

Who should be accountable for artificial intelligence inside a fund?

The FCA relies on existing frameworks, among them the Senior Managers and Certification Regime, so responsibility stays with the named individual who holds it. A partnership should place a person between any output and any decision, and name one owner for the decision record, which is where investment committee governance holds or fails.