AI adoption in accounting firms usually stalls for five reasons: hourly billing makes speed look like lost fees, busy season wipes out practice time, partners won't approve client data going in, nobody owns the change, and the first project is hard technical work. Restart with a named owner, fixed fees, and one admin workflow run for 30 days after a deadline peak.
First check it has stalled rather than slowed. The signs are consistent: licences that nobody opens (the Copilot usage report in the Microsoft 365 admin centre shows active users over the last 7, 28, 90 or 180 days), staff quietly using personal chat apps instead of the approved one, and nobody able to name a single workflow that is done differently from a year ago. If two of those are true, a reminder email won't fix it. The cause needs finding.
Five stall points peculiar to accounting practices
Every small business meets general barriers to AI, such as time, skills and cost. Practices meet five specific ones, and they tend to arrive in this order.
Hourly billing turns every saved hour into a smaller invoice
If a set of accounts is billed on time, faster work is lower income. The arithmetic is visible to everyone. Say a manager uses AI to cut a sole trader's year-end from nine hours to six, at $75 an hour: the fee falls from $675 to $450. The manager has done good work and the partner sees $225 disappear. Nobody says "stop using AI", but nobody pushes for more of it either.
Run the same job on a fixed fee and the sum turns round. Quote that sole trader's year-end at $650 at renewal. At nine hours, the practice earns about $72 an hour on it; at six hours, about $108. The client pays slightly less than last year's time-based bill, and the three saved hours become margin or capacity for another client instead of a smaller invoice. The awkward edge case is the client whose records are always a mess: fix the fee on stated assumptions ("bank statements and receipts supplied by 30 June, reconciled monthly"), and bill extra work when they aren't met, or the fixed fee quietly absorbs their chaos.
Restart move: move recurring compliance to fixed fees at renewal before asking anyone to go faster. The fee conversation is covered in whether to charge clients less when AI speeds up the work.
The deadline calendar eats the learning time
Practices launch AI in a quiet month, then a filing peak arrives three weeks later. Under pressure, people go back to the method they know, because a new tool always feels slower for the first dozen uses. When the peak passes, the new habit has gone and nobody restarts it.
Restart move: start in the four to eight weeks after a peak, never in the six weeks before one, and plan for the next peak explicitly: which AI routines will continue under pressure and which will pause.
Write that plan down before the peak, while people are calm enough to agree it. For an illustrative four-person practice, the peak plan was five lines long:
PEAK PLAN (six weeks before the filing deadline to one week after)
Continue: record chasers, long-thread email summaries
Pause: file-note trial, any new prompts or tools
Owner time: cut from two hours a week to one 15-minute check on Fridays
If a routine breaks: go back to the old method, log it, fix it after
Restart date: the Monday after the deadline, in the diary nowThe restart date is the line that matters. Practices rarely abandon AI during a peak on purpose; they just never pick it back up, because nobody decided when.
One partner's data worry becomes a veto
A single partner says "no client data in AI, full stop", and because the concern is reasonable, the whole practice waits. Months pass. Meanwhile juniors use free tools on their phones, which is the very risk the partner feared, only now without any controls.
Restart move: answer the worry in writing, with specifics, using the one-page memo further down. A vague "it's fine, it's a business plan" doesn't move a careful partner; a list of the controls does.
Nobody owns it, so everybody assumes someone else does
The practice bought licences, maybe ran a lunchtime demo, and then expected adoption to happen. Built-in features in practice-management and ledger software often sit switched off because nobody was asked to go through the settings. The Copilot licences renew every month whether used or not.
Restart move: name one person, give them two protected hours a week for three months, and make "AI" a standing five-minute item at the team meeting. How to choose and support an AI champion covers who suits the role; in a practice, it's often the manager rather than a partner.
The five-minute item works because it has the same three questions every week. Notes from week three of an illustrative restart:
AI, week 3
Worked: chaser drafts. 11 sent, 9 needed only name and figures.
Broke: two drafts asked for receipts the client had already uploaded
to the document inbox. Coordinator now checks the inbox first.
One change this week: add the fixed-fee clients' year-ends to the
chaser prompt so drafts mention the right deadline.Keep it that short. A standing item that turns into a 30-minute debate gets dropped from the agenda by month two.
The first project was the hardest one
Partners naturally test AI on the work they know best, which is technical: tax research, complex adjustments, a tricky treatment. That's where language models are weakest and where errors are most embarrassing.
A realistic version (an illustration): a partner asks a chat assistant how a particular relief applies to a client's property sale. The answer is fluent, cites a rule, and gets the qualifying period wrong. The partner spots it, tells the other partners at lunch, and the practice's verdict becomes "it makes things up". That was true for that task. It says nothing about whether AI can draft a chaser or summarise an email thread, but the idea has lost its sponsor.
Restart move: start with admin that repeats weekly and where errors are cheap to catch: email triage, record chasers, file notes, first drafts of routine letters.
Diagnose your practice in twenty minutes
Score each statement 0 (not true), 1 (partly) or 2 (true). The highest scores tell you which stall point to tackle first. Here it is filled in for a three-partner practice with idle Copilot licences (illustrative):
| Statement | Stall point | Score |
|---|---|---|
| Most recurring compliance work is billed on time spent | Billing | 2 |
| AI work started less than two months before a filing peak | Calendar | 2 |
| At least one partner has said client data can't go into AI | Risk veto | 1 |
| Nobody has protected time to run AI adoption | Ownership | 2 |
| The first thing tried was technical research or advice | First project | 1 |
| Built-in AI features in our software are still switched off | Ownership | 2 |
| Staff use personal AI accounts for work | Risk veto | 1 |
This practice scores 4 on ownership, 3 on the risk veto and 2 each on billing and the calendar. So the restart begins with an owner and protected time, and the risk memo comes in week one. The fee change can follow at the spring renewals, because it takes longer.
The one-page memo that gets a careful partner to yes
A partner who has vetoed client data usually wants to see that someone has thought about it as carefully as they have. A filled-in memo like this one, kept to a page, does more than a demo:
Proposal: use ChatGPT Business for drafting client emails and chasers, starting 3 November, for 30 days.
Data allowed: client first names, transaction dates, amounts and payee names. Not allowed: tax reference numbers, bank account numbers, identity documents, anything about health or family circumstances.
Terms: business plan, no model training on our data by default; data processing addendum signed on 20 October, copy in the compliance folder.
Review: every draft read by the sender before it goes out; the manager samples five a week.
Stopping rule: any client complaint or data slip pauses the trial until the partners have reviewed it.
Measured by: hours spent on chasers and routine replies, compared with the two weeks before.
Everything in it is checkable, and it names the conditions under which the trial stops. Careful partners tend to approve trials that can be stopped.
A 30-day restart timed around the deadline calendar
This assumes the restart begins in the weeks after a peak, with an owner named.
- Days 1 to 5: inventory and settings. List every AI tool in use, including personal ones. Switch on and configure the AI features already inside your practice-management and ledger software. Remove licences from people who have left.
- Days 6 to 10: pick one workflow. Use a short time log to find the admin task with the most hours and the lowest risk. For most practices it's chasing, routine client email or file notes.
- Days 11 to 25: run it alongside the old way. Everyone involved uses the new routine for that one task. The owner collects problems daily and fixes prompts and settings the same day, because unresolved friction is what kills the habit.
- Days 26 to 30: measure and decide. Compare hours with the baseline, list what went wrong, and decide: keep, adjust, or drop. Then pick the second workflow.
The inventory in days 1 to 5 is usually where the surprises are. One row per tool is enough, and filled in for an illustrative practice it read:
| Tool | Who uses it | For what | Account | Action |
|---|---|---|---|---|
| Free chat app on phones | Two bookkeepers | Rewording client emails, summarising tax authority letters | Personal, free | Move both tasks to the approved tool; delete the chat history |
| Ledger's built-in transaction suggestions | Nobody | Switched off at setup | Practice | Turn on for one test client |
| Copilot Business | Nine licences, two active | Occasional email drafts | Practice | Reassign two licences in week one |
The first row is the one to handle gently. The bookkeepers weren't being careless; the approved tool was harder to reach than the app on their phones. Treat it as evidence of which tasks people want help with, which is also a shortlist for step two.
The old and new versions of the chosen workflow should be written down side by side, so everyone knows what changed. For record chasing, for example:
Old process:
Coordinator opens each client's ledger, lists missing receipts by
hand, writes an individual email, logs it in the diary. About 20
minutes per client, done for 12-15 clients a week.
New process:
Coordinator exports unmatched transactions older than 30 days
(client codes only), asks the approved assistant to draft one email
per client, checks each against the do-not-chase list and the
document inbox, fills in names, sends. About 6 minutes per client.The drafting prompt behind the new process can be short, as long as it says what the draft must not do:
Draft a friendly email to client [client code] asking for records for
the transactions listed below. Group them by month. Say why we need
them in one sentence. Ask them to upload to the usual document inbox.
Do not mention penalties or deadlines unless I add one. Under 150 words.
[unmatched transactions: date, payee, amount]A first draft might open (illustrative): "Dear Client, We are writing to inform you that the following items remain outstanding and require your urgent attention." Two fixes before it goes anywhere. The tone is a final demand for what is usually a routine request, so add "write as a colleague, not a credit controller" to the prompt. And "urgent" is a word the prompt didn't ask for; if the assistant keeps adding pressure, give it one example of a chaser your practice likes and ask it to match that. After those changes, drafts in the illustrative trial needed only the client's name and a check against the document inbox.
If a pilot has stalled more than once, why your AI pilot stalled, and how to get it live covers the general causes that sit underneath these accounting-specific ones.
What a restart looks like in a three-partner practice
Take the practice scored above: three partners, two managers, four bookkeepers and admin staff, clients ranging from sole traders to a small chain of guest houses and a campsite. It bought nine Microsoft 365 Copilot Business licences in the spring at $21 a user a month on annual billing, so $189 a month. By September, the usage report shows two active users in the last 28 days. The partners are close to cancelling.
The restart, in outline:
- Owner: the more junior manager, with Tuesday and Thursday afternoons protected for an hour each.
- Timing: the four weeks after the practice's busiest filing deadline, with the plan for the next peak agreed up front: chasers continue, experiments pause.
- Risk: the memo above, adjusted for Copilot, which works inside the practice's own Microsoft 365 files and respects existing folder permissions. The cautious partner approves it with one change: no use on the two clients under investigation by a tax authority.
- Workflow one: Copilot in Outlook for summarising long client threads and drafting routine replies. The manager writes five example replies in the practice's house style and shares them as a reference file.
- Licences: reduced from nine to seven, moving two to people who actually handle client email.
After 30 days, the usage report shows six of seven licences active over 28 days. The team's own log shows routine email time down from about 18 hours a week to 13 across the practice, with the review time counted in. Those are illustrative figures, but they show the shape: small, measured, and enough to justify workflow two, which is file notes from client meetings. How to measure whether Copilot is paying for itself sets out the fuller method.
Signals the restart is holding at 90 days
- Usage is steady through a busy month, not just a quiet one. Check the usage report after the next deadline, and read it alongside the team's time log: someone who tried Copilot once in the period still shows as active, so an active count can look healthy while the saved hours have gone.
- At least two workflows are written down in old-versus-new form and followed by people other than the owner.
- Personal-tool use has dropped, because the approved tool is now the easy option.
- A partner has proposed the next workflow, which is the clearest sign that ownership has spread.
- The fee model is moving: at least some recurring work is quoted fixed at renewal.
If after 90 days usage has fallen back, don't relaunch with more enthusiasm. Go back to the diagnosis table; the score that has risen is where the next fix lies. The change-management plan for small teams and the AI readiness checklist for professional firms are useful at that point, because a second stall usually means a gap that was there before the first.
Further reads
- Barriers to AI Adoption in Small Businesses and How to Clear Them — The general barriers, for comparison with the accounting-specific ones.
- Rolling Out AI in a Bookkeeping Practice Without Losing Control — How a bookkeeping practice rolls out AI without losing control.
- How to Train Staff to Use AI in a Small Business — Training that sticks once the restart has an owner.
- What an AI Implementation Looks Like in a Small Accounting Firm — What a full implementation looks like once momentum is back.
- Turning Hours Saved by AI Into Advisory Work Clients Pay For — Give partners a reason to want the saved hours.
- How to Rescue a Stalled AI Project or Exit It Cleanly — When a restart isn't worth it and you should exit cleanly.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft 365 admin centre documentation on the Copilot usage report; Microsoft 365 Copilot Business pricing. Firm examples and figures are illustrative.