AI helps most in operations when it’s applied to real workflow — automating admin, drafting, data entry and triage — rather than bolted on as a gimmick. The value comes from an operating lead who integrates it into how the firm genuinely works.
Where AI actually helps
- Automating repetitive admin, data entry and reconciliation prep
- Drafting communications, summaries and first-pass documents
- Triaging queries and routing work faster
- Surfacing information from across systems so people spend less time hunting
Why most AI efforts underdeliver
Tools get bought and half-adopted, disconnected from the workflow they’re meant to improve. AI only pays off when it’s embedded into how the team actually works — the same reason any software fails or succeeds.
Treat it as an operations project
The firms getting real value put someone in charge of integrating AI into their operating model, testing it against genuine bottlenecks, and embedding the ones that work. That operational lead is exactly what a fractional COO provides.
The profession has split, and the gap is not the software
Recent industry research points the same way: practices with AI genuinely embedded in daily workflows report meaningful time savings per person per week, and the most profitable firms record substantially better margins than their peers. Those findings come from vendor-sponsored surveys, so treat the magnitudes with appropriate caution — but the direction is consistent enough across sources to be worth acting on.
The important part is what separates the two groups, because it is not access. Everyone can buy the same tools at the same price this afternoon. The difference is whether the tool is embedded in a process or sitting alongside one.
That is the same distinction that decides whether practice management software pays back, and it fails for the same reason: a tool records or accelerates a process, it does not create one. Dropped into an undefined workflow, AI produces faster inconsistency.
Where it earns its keep in a practice today
| Job | Why it suits AI | What still has to be human |
|---|---|---|
| Records and document extraction | High volume, structured output, immediate feedback | Exception handling and the judgement calls |
| First-pass review | Catches the obvious before a human looks | Every conclusion, and the sign-off |
| Drafting client correspondence | Volume, repetition, a house style to follow | Anything containing advice, and the send button |
| Query generation | Spots gaps consistently, does not get bored | Deciding which queries are worth asking |
| Summarising a client's position | Pulls scattered facts into one place fast | What it means and what to do about it |
The pattern across all five: AI is good at the first draft of something repetitive, and the value is realised only where a person was previously doing that first draft. If the first draft was already automated, or nobody was doing it, there is nothing to release.
Why most efforts stall, specifically
Three failure modes, in order of how often they occur.
It is nobody's job. A tool is bought, enthusiasm is high for a fortnight, and then January arrives. Nobody owns adoption, so nobody notices the slide back. This is the same mechanism that leaves practice management systems half-used, and it has nothing to do with the technology.
It is applied to an undefined process. If two preparers do the same job differently, AI trained on or dropped into that process produces two different kinds of output, and review gets harder rather than easier.
It is measured on enthusiasm rather than hours. “It is brilliant” is not a result. If the saved hours have not been identified and reallocated to something, they disperse into the general pace of the week and the practice ends up paying a subscription for a feeling.
How to run it as an operations project
- Pick one job, not one tool. Start from the task that costs most and is most repetitive, then choose something for it. Starting from the tool produces a solution looking for a problem.
- Baseline it first. Time the task as it currently runs, for a fortnight. Without this you can never tell whether anything changed, and every subsequent conversation becomes anecdote.
- Write the standard. What the output must contain, what is checked, what is never accepted without a human. This is also your audit trail when somebody asks how the work was done.
- Name the owner. One person accountable for adoption, with time for it. Not the most enthusiastic person — the one who will still be doing it in March.
- Decide in advance what the hours are for. Advisory capacity, review capacity, going home earlier. An unallocated saving is reabsorbed within a month.
- Measure again at ninety days against the baseline, and be willing to conclude it did not work.
The professional obligations, briefly
None of this changes who is responsible. The work is yours, the sign-off is yours, and “the tool produced it” is not a defence to a client, a professional body or an insurer.
Three practical points worth settling before anything touches client data: where the data is processed and whether that is covered by your engagement terms and privacy notice; whether client data is used to train a third-party model, and how to switch that off; and what your professional indemnity insurer expects to see in the way of human review. Ask the insurer directly rather than assuming — the answer is usually reassuring and occasionally not.
The defensible position is straightforward: AI drafts, a qualified person reviews and signs, and the standard for review is written down. A practice that can show that has a policy. One that cannot has an exposure it has not priced.
The honest summary
AI is currently the best available answer to a specific problem — the volume of repetitive first-draft work in a practice that cannot recruit its way out. It is not an answer to an undefined workflow, an absent management layer or a pricing problem, and the practices getting the least from it are usually the ones that needed one of those three instead.
Which is worth checking before spending anything. If work is stuck in queues rather than in preparation time, a faster first draft will not touch the constraint, and the subscription will quietly become another thing that was tried.
Common questions
Will AI replace staff in accountancy firms?
More realistically it removes low-value tasks and frees people for higher-value work — if it’s embedded into workflow properly. Bolted on without that, it mostly adds noise.
Where should a firm start with AI?
With a real bottleneck, not the shiniest tool. Pick a repetitive, high-volume task, test AI against it, and only keep what measurably helps — then embed it.
Where does your firm actually stand?
Score it across the ten things a COO would look at first. Three minutes, an honest read on where you’re strong and where you’re exposed, and a 90-day plan for the areas holding you back.
Or book a confidential call if you’d rather just talk it through.