Adoption Outran Accounting: Why Most Firms Can't Say What AI Returned

Thomson Reuters' 2026 AI in Professional Services report, drawing on more than 1,500 professionals, found that just 18% say their organisation tracks return on its AI investment — with a further 40% unsure whether it is measured at all.

Set that against adoption. Separate 2026 research across 1,300+ legal professionals found 69% now using general-purpose generative AI for work, more than double the prior year. Use is widespread. Accounting for use is not.

The consequence is that most confident statements about AI working are, strictly, statements about how the work feels. That may well be correct. It is not defensible, and it will not survive the first budget review conducted by someone sceptical.

Baselines are the missing piece

You cannot report a return against a state you never recorded. This is obvious when stated plainly and near-universally skipped in practice — because the moment of maximum enthusiasm for a new tool is exactly the moment when stopping to write down current performance feels like bureaucracy.

The fix is one sentence per workflow, written before anything is built:

Today this takes ___ , happens ___ times per week, and fails ___ % of the time.

Three blanks. If they cannot be filled, that workflow is not ready for change — not because the tooling won't help, but because there will be no way to distinguish improvement from relocation.

After the change, the same three blanks get filled again, using the same definitions over a comparable period. The difference is the return, and it is specific enough to defend.

A faster step is not a faster process

The finding that deserves more attention than it has had: automating a task without redesigning how the service is delivered tends to produce weaker returns than changing the service model alongside the tooling.

The mechanism is straightforward. A faster step inside an unchanged process does not produce a faster process. It produces a queue in front of the fast step. The saving is real at the level of the task and invisible at the level of the business.

Capturing the return requires changing what the service is, not only how one part of it is performed.

The hourly-billing trap

For firms billing time, there is a sharper version of this problem, and it is arithmetic rather than strategy.

If revenue is hours multiplied by rate, and the hours fall while the rate holds, efficiency does not produce a gain. It produces a discount — extended to every client simultaneously, without negotiation, and without anyone deciding to grant it.

This is a plausible mechanism behind headline findings like PwC's 29th Global CEO Survey, in which 56% of 4,454 CEOs reported no significant financial benefit from AI to date. In some of those cases the technology likely performed and the commercial model absorbed the benefit before it could reach the income statement.

Governance is the second hole

In that same survey of 1,300+ legal professionals, 43% reported their firm has no formal AI policy and no plans to create one, and only 9% had a written policy that is actually enforced. Separately, 40% reported receiving contradictory instructions from clients — some directing that AI be used on their matters, others directing that it not be.

That combination is operationally unstable. Without an internal standard, individual staff make case-by-case judgments, and the firm discovers its own position only when a client objects to a decision it never knowingly made.

Setting that policy is the firm's own call, and it depends on professional rules that no vendor is in a position to interpret on a firm's behalf. But the absence of a policy is itself a decision, and in a large share of the market it is being made by default.

A defensible sequence

1. Write the before. One sentence, three blanks, per workflow, prior to any purchase or build.
2. Change the smallest piece with a clear baseline. Measurability beats impressiveness at this stage.
3. Measure the same three blanks after. Same definitions, same period length.
4. Then redesign the service, and the price. This is where the return is captured, and it cannot be done credibly without the first three steps.

The first step costs roughly half an hour per workflow. It is the cheapest element of any AI project and the one most consistently omitted — which is a reasonable explanation for why so much of the market reports nothing from investments that are, in fact, working.