How Do CFOs Justify AI Due Diligence Spending to Boards?

Updated August 2026 · 7 min read · Deal Room Intelligence Series

Why the efficiency framing collapses in the room

Efficiency claims fail board scrutiny for three specific reasons, and they fail in the same order every time.

The saving is not bankable. Hours freed from document review do not become cash unless someone leaves or unless the firm does more deals with the same team. A board that has seen a few technology business cases knows this, and the question arrives within minutes.

The baseline is undocumented. “60% faster” than what? Almost no firm has instrumented its current review process by activity, so the percentage rests on an estimate the board cannot test, from a vendor with an interest in the answer.

The number is borrowed. If the figure came from a vendor case study or a sponsored survey, a single sceptical director asking for provenance ends the discussion — and damages your credibility on the next paper too.

The case that holds: coverage, not speed

Reframe the purchase from doing the same work faster to changing what the firm can demonstrate it examined. This survives scrutiny because it is verifiable, and because it speaks to a risk the board already owns.

The argument in one paragraph, suitable for a board paper:

Note what this does not claim. It does not promise fewer hours, fewer people, or that nothing will be missed. It claims a change in the evidentiary position — and that is a claim a board can test, because either the report shows coverage or it does not.

The four numbers a board will actually engage with

1. Documents examined versus documents read

Most firms have never calculated this gap. Do it for your last three transactions. It is usually uncomfortable, it is entirely defensible because it is your own data, and it makes the problem concrete in a way no percentage does.

2. Cost of the workstream you would have stopped

Take a deal your firm abandoned. Identify when the disqualifying issue was found and what the diligence spend was between kick-off and that moment. That figure — real, from your own history — is the value of finding it earlier. It is usually the largest single number available to you and it needs no vendor to substantiate it.

3. Blended rate of triage hours

Separate the hours spent deciding what to read from the hours spent reading. Triage is disproportionately senior time, billed or costed at senior rates, producing only an ordering. Boards understand seniority mix immediately.

4. Standing verification cost

Include it, prominently. Someone must check escalated findings against source text, permanently — grounded commercial legal AI has been measured hallucinating between 17% and 33% of the time in an adjacent task^[1], and no vendor has eliminated that. A case that shows this line is credible. A case that omits it invites the question that unravels everything else.

The risk argument, stated for a board

Boards fund risk reduction more readily than efficiency, provided the risk is specific. Three formulations that work:

Asymmetry. The cost of examining a document is small and known. The cost of a material provision missed in an unexamined document is large, unknown, and lands post-close when leverage is gone. Any control with that payoff shape is worth buying well past the point where its efficiency case makes sense — this is the insurance argument, and boards are fluent in it.

Defensibility. When a miss is investigated, there are two possible findings: the document was scored and ranked below a documented threshold by a named reviewer, or it was never examined and nobody knew. The first is an ordinary professional judgment; the second is an unmanaged gap. Screening moves misses from the second category to the first. That is worth real money in an indemnity dispute and it is worth more in reputation.

Obligation. For regulated professionals this is no longer purely discretionary. ABA Formal Opinion 512 (29 July 2024) holds that lawyers using generative AI must “fully consider their applicable ethical obligations,” including competence and supervisory responsibility.^[2] More broadly, the governance surface has hardened fast and on the record: the NIST AI Risk Management Framework in January 2023, ISO/IEC 42001 the same year, ABA 512 in 2024, and the EU AI Act applying from 2 August 2025 with penalties reaching €35,000,000 or 7% of worldwide annual turnover for prohibited practices.^[3] A board asking “why now” can be answered with dates rather than with market forecasts.

Structuring the paper

1. The exposure, in your own numbers. Documents examined versus read, from your last three deals.
2. What changes. Every document scored against every category; reading order set by risk; unread documents ranked below a pre-set threshold.
3. What does not change. Materiality judgment stays human. Headcount stays. Verification is a new standing cost. Say all three explicitly — the credibility of the paper rests on this section more than any other.
4. Full cost. Licence, overage, verification labour, integration, security review. Not the quote.
5. The value range. Triage hours at blended rate, plus a range — clearly labelled as a range — for earlier abandonment, anchored to the real deal from point 2.
6. How you will know it worked. A committed measurement plan: coverage gap, second-pass triggers, override rate. Boards fund things that come back with evidence.

Three questions to have answers ready for

“If it's so efficient, why aren't we cutting cost?” Because we are converting the saving into coverage rather than into headcount. We will read a similar number of documents; we will have examined all of them. If you want the saving taken as cost, that is a separate decision and I would advise against it in the first year.

“What if it misses something?” It will. So will we. The change is that a miss becomes a scored item ranked below a documented threshold rather than a document nobody opened. We are not buying certainty; we are buying a defensible record and a better allocation of attention.

“Where did these numbers come from?” Ours — instrumented across the last three deals. This is why the paper carries no vendor statistics, and it is the answer that ends the question rather than extending it.

Bottom line

Lead with the coverage gap in your own data. Frame the purchase as changing what the firm can demonstrate it examined, not as doing the same work faster. Include the verification cost. Say plainly what does not change.

A board will fund a smaller, harder claim it can test far more readily than a large one it cannot — and in this category, the smaller claim happens to be the true one.

Produce a coverage record from a real document →