It compresses the research and the drafting, which is most of the hours and almost none of the value. The risk is narrow and sharp: a deliverable that reads generated tells the client precisely what they were paying for, and it was not that.
The useful split here is between what AI does to a consulting firm's cost base and what it does to its price. They move in different directions and firms that only notice the first get squeezed.
Where it genuinely helps
Research synthesis. Twenty documents, an industry you have not worked in for three years, a competitor landscape. Reading and condensing is the strongest thing these models do and the downside of a miss is small because you are going to check the important parts anyway.
First drafts. Structure, headings, the dull connective sections of a report. Speed on the parts that were never where your judgement lived.
Transcription and session notes. Genuinely transformative for anyone running back-to-back client sessions, and close to zero risk.
Pre-reading preparation. Summarising what changed for a client since you last met, which is the thing most likely to be skipped when you are busy.
The thing clients actually notice
Not whether you used AI. Whether the output could have been produced without knowing them.
A deliverable that reads generic reprices the engagement in the client's mind, immediately and permanently. They cannot articulate why the last report felt worth it and this one did not, but they adjust what they are willing to pay, and that adjustment does not reverse.
So the practical rule is unglamorous: use it for the parts that would have been the same whoever wrote them, and write the parts that could only have come from you. The second category is smaller than most consultants think and it is the entire product.
The two risks worth taking seriously
| Risk | What it looks like |
|---|---|
| Confidentiality | Client material in a tool whose terms nobody read. Increasingly a procurement question with a written answer expected |
| Commoditisation | If the deliverable could have come from a public prompt, the client eventually works that out |
On the first, a written AI policy is becoming table stakes rather than good practice. What tools are approved, what client material may go into them, where it is hosted, whether anything trains a model. The practice version of that is covered in what goes in an AI policy for an accounting firm, and the content transfers to any professional services firm.
What AI does not touch
The engagement still fails or succeeds on the same thing it always did: whether anything happens in the ninety days after the recommendation lands.
That has nothing to do with document production. It is whether each recommendation has a named owner, a measure, and a review short enough to catch a stall. A faster, better-written report with no owners attached fails exactly like a slow one did, which is the argument in why strategy execution fails.
This is also where the honest opportunity sits for a firm. If AI has taken hours out of your production, the margin can go to the part clients value most and almost nobody does well: staying visibly useful between meetings. GPS for Professional Services is built for that, with a workspace per client holding their plan and their progress, so the client can see movement without waiting for a document, and the firm can see every engagement from one screen.
Related reading: what AI policy belongs in a practice, and business advisory vs business consulting.
See what clients see between meetings, and GPS for Professional Services.
