Change Management When Adopting AI for Deals

Updated August 2026 · 6 min read · Deal Room Intelligence Series

The objection under the objection

Stated objections are usually about accuracy. The real one is almost always about accountability, and it is entirely rational.

A senior reviewer who accepts a ranking is accepting that documents below the line will not be read — and if something is later found in one of them, their name is on the file. 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.^[1] That duty does not move to the vendor, and experienced people know it.

The answer that works is structural rather than reassuring: with screening, an unread document is examined, scored, and ranked below a threshold the firm set in advance. Without it, an unread document was chosen by folder order and nothing in the record shows anyone considered it. The reviewer is more protected with the tool than without it — and that is a claim they can verify.

Sequence the rollout so trust is earned, not requested

Run parallel before you change anything

On the first live deal, run screening alongside the existing process without letting it change what anyone reads. Reviewers see the ranking next to their own judgment on a deal where the human process is fully intact.

This is slower and it is the step most often skipped. It is also the only way a sceptical reviewer gets evidence rather than a promise.

Expect the first run to look bad

Default severity settings over-flag, because they are calibrated to a generic rubric rather than your sector. If the team sees this before anyone has explained it, the conclusion is that the tool does not work — and that impression is hard to reverse.

Say it in advance: the first pass will flag clause types we have waved through for a decade, and fixing that is what the next three deals are for.

Make disagreement the point

Ask reviewers to override, and capture what they concluded and why. This reframes their role from approving a machine's output — which feels like rubber-stamping and invites resistance — to teaching the system what this firm considers serious, which is a use of expertise rather than a substitute for it.

It also produces the calibration data that makes the tool genuinely better over three or four deals, so the framing is honest rather than a management device.

Address the fee question directly

In hourly-billing firms there is a real concern nobody says aloud: if triage compresses from twenty hours to two, what happens to the twenty?

Opinion 512 is explicit that fees must be reasonable and consistent with time actually spent.^[1] So the honest answer is that the hours do not survive, and pretending otherwise makes the whole rollout feel dishonest to the people being asked to adopt it.

What works better is naming what compresses: the low-judgment, low-rate work. What remains is disproportionately the analysis clients actually value. Firms that reprice around judgment rather than defending hours on triage are in a stronger position when a client eventually asks what the tool changed — and that conversation is coming regardless of your rollout.

Roles, and who actually blocks

The associate row deserves attention because it is usually mishandled. Junior reviewers are told the tool will free them for higher-value work, which sounds like the standard line. It is more persuasive stated concretely: the hours removed are the hours spent opening documents to find out whether they are worth opening, which nobody has ever described as the good part of the job.

Four failure modes

The champion is the only user. If one person runs every scan, the tool has not been adopted — it has been personally acquired. Fix by rotating who runs the screening pass from the second deal onward.

Reviewers approve everything. The failure that looks like success. Track override rate; if it approaches zero, either your rubric is perfectly calibrated, which is unlikely, or nobody is really reading. This is worse than non-adoption because it produces a signature without a judgment.

Verification quietly stops. Any control that costs meaningful effort per finding gets abandoned in week three of a live deal. This is a design constraint, not a discipline problem: if your tool does not quote source text, verification costs a re-read and will lapse exactly when volume is highest.

Nobody owns thresholds. Without a named owner for threshold-setting, override review and the sampling protocol, the rollout stalls after the first deal — available to everyone, nobody's job.

What to say at the kickoff

Three sentences, and they should be the honest ones rather than the ambitious ones:

1. “This does not change who is accountable.” Materiality stays with us. The system ranks; it does not decide.
2. “This changes what we can show we examined.” Every document scored against every category, so the ones we do not read are a recorded decision rather than an accident.
3. “The first pass will be badly calibrated and we need you to tell us where.” Overrides are the product, not an inconvenience.

Notice none of these promise speed. Leading with efficiency invites the question about headcount and makes the whole thing feel like a cost exercise — which is both worse politics and a less accurate description of what the tool does.

Bottom line

Adoption fails on trust and exposure, not on capability. Run parallel first so reviewers get evidence. Warn them the first pass will over-flag. Make overrides the explicit ask. Answer the fee question honestly rather than avoiding it. Name an owner.

And lead with the coverage argument rather than the speed one — partly because it is more persuasive to the people who can block you, and partly because it is the claim that survives the first time something is missed.

Run a parallel scan on a closed deal →