Nearly four out of five accounting firms experimenting with AI this year will close their books in December no better off for it. Not because the technology failed — because the firm never built the foundation to capture the value it created.

The gap isn’t the tools. It’s everything around them.

Across the 10–200 staff firms we work with, almost every partner has already tried something — a Copilot licence here, a document-summarising tool there. Very few have seen that translate into hours saved, fees protected, or new capacity for higher-value work. The pattern is consistent enough to name three specific failure points.

1. Pricing didn’t move, so the savings evaporated

If a task that used to take four billable hours now takes one, and you’re still billing by the hour, you’ve just cut your own revenue on that engagement by 75%. Firms that see real ROI from AI have almost always already moved — or are actively moving — toward fixed-fee or value-based pricing on the work AI touches. Without that shift, efficiency gains are a pay cut, not a profit centre.

2. No policy means no confidence — so adoption stalls at the individual level

Without a clear, written policy on what client data can touch which tools, most partners quietly restrict AI use to low-stakes, low-value tasks — which is precisely where the ROI is smallest. The firms getting real value have governance in place first, which lets them extend AI into higher-value, higher-sensitivity work with confidence instead of caution.

3. Leadership never made a real decision — they made a purchase

A licence isn’t a strategy. The firms seeing ROI can answer specific questions: which service lines are we changing, which roles are we redeploying, what does success look like in 90 days. Firms without those answers are running a pilot with no endpoint — which is indistinguishable, financially, from doing nothing.

What the 21% did differently

Across the firms we’ve advised, the ones capturing real value followed a consistent sequence rather than a scattershot rollout. They started with an honest audit of where they actually stood — not where leadership assumed they stood. They wrote the governance policy before extending tool access, not after an incident forced their hand. They renegotiated pricing on the specific services AI touched before declaring victory on efficiency. And they gave at least one partner explicit responsibility for making the calls, rather than leaving adoption to individual initiative.

None of this requires an enterprise-scale innovation team. It requires doing the four things in the right order, with fixed scope and a fixed fee, in a matter of weeks rather than a multi-year transformation programme.