The debate about AI in investigations has been about accuracy. It should have been about arithmetic. Validation is the cost that makes the efficiency case collapse, and almost no matter budget contains a line for it.
A partner tells the client the AI review will cut six weeks to two. The client approves the budget on that basis. The tool performs roughly as advertised. And then somebody has to check it.
That last step is where the argument for AI in investigations quietly falls apart, and it falls apart on economics rather than on capability. If an AI-generated chronology is going to inform a privilege call, a disclosure decision, or a termination, someone competent has to trace every material assertion back to source. That tracing is slow, senior, and unglamorous. It is also the only thing standing between the output and a methodology challenge. Priced honestly, it eats most of the saving.
So it does not get priced honestly. It gets absorbed, deferred, sampled, or skipped, and the matter proceeds on a foundation nobody has actually inspected.
Why Validation Is the Whole Problem
The market conversation has fixated on hallucination, largely because of a handful of sanctioned filings where a lawyer cited a case that did not exist. Those are memorable and they are not the corporate risk. Fabricated citations are easy to catch precisely because they are absurd. The case is either in the reporter or it is not.
The corporate risk is subtler and worse. It is a chronology that is 90% right, competently written, internally coherent, and wrong about one date in a way that changes who knew what and when. Nothing about the document signals which 10% to distrust. Fluency is uniform across the accurate and the inaccurate parts, and the reader has no way to tell them apart without going to source.
The output does not degrade gracefully. It fails with the same confident register it succeeds in. That is the property that makes checking mandatory and expensive.
This is the inversion nobody has priced. A junior associate who is unsure signals uncertainty. They hedge, they flag, they ask. Their work product tells you where to look. An AI-generated summary offers no such signal, which means you cannot allocate your checking effort intelligently. You have to check all of it, or you have to accept that you do not know which part is load-bearing.
What This Costs, Concretely
Assume a document review where AI clustering and summarization genuinely compress the timeline. The tool saves you the reading. It does not save you the verifying, and verifying an assertion is not much faster than forming it. The saving is real at the triage layer, where being wrong is cheap and recoverable, and it approaches zero at the conclusion layer, where being wrong is neither.
The practical consequence is that AI is enormously valuable for deciding what to look at and close to worthless for deciding what is true. Every organization deploying it in investigations has to draw that line explicitly, in writing, at matter inception, because the line will not hold if it is drawn later under time pressure by someone with a deadline.
The Uncomfortable Question for the Profession
If validation costs what it costs, then AI-enabled investigation is not cheaper. It is faster at the front and more expensive at the back, and whether it nets out depends entirely on the matter. That is a defensible position. It is not the position being sold.
Organizations should be suspicious of any proposal that promises AI-driven cost reduction in an investigation without a corresponding line item for verification. The absence of that line does not mean the cost is not there. It means somebody has decided not to tell you about it, or has not thought about it, and it is not obvious which is worse.
What a Serious Protocol Looks Like
The distinction that matters is between using AI to narrow the field and using AI to reach a finding. Everything else follows from getting that boundary right and enforcing it when the schedule slips.
- Set the boundary at matter inception, in writing, and define what happens when it is under pressure. A protocol that has never been tested against a deadline is not a protocol.
- Price the validation as its own line. If it is not visible in the budget, it will not happen in the matter.
- Require source linkage for any assertion that touches a privilege call, a disclosure decision, or a person losing their job. Assertions below that threshold can carry lighter treatment. Assertions above it cannot.
- Label evidentiary status on the face of the document. Lead, hypothesis, verified fact, disputed fact, conclusion. The absence of these labels is what allows a hypothesis to harden into a finding through nothing more than repetition in successive drafts.
- Keep the trail. Tool, data scope, prompts or query logic, reviewer, and what changed after review. If you cannot reconstruct how a conclusion was reached, you cannot defend it, and the fact that a machine was involved makes the reconstruction question sharper rather than softer.
The Honest Position
AI belongs in investigations. It finds patterns across volumes of data that no team can read, in languages the team may not speak, at speeds that matter when a regulator is waiting. Refusing to use it is not a defensible posture and will not be for much longer.
But it belongs there as an instrument that produces leads, operated by people who understand that a lead is not a fact and that the distance between them is measured in verification hours somebody has to pay for. The organizations that get this right will be the ones that budgeted for the checking. The ones that get it wrong will discover the cost anyway, in a deposition, when someone asks how a particular date was established and the answer is that a tool said so.
Ankura designs AI-enabled investigation protocols and validates AI-generated chronologies, summaries, translations, and issue lists against source, combining forensic collection, analytics, forensic accounting, and regulatory response.
© Copyright 2026. The views expressed herein are those of the author(s) and not necessarily the views of Ankura Consulting Group, LLC, its management, its subsidiaries, its affiliates, or its other professionals. Ankura is not a law firm and cannot provide legal advice.
