AI for accounting firms: what to automate first, what to leave alone, and how to get seniors to trust it
Most accounting firms start with AI by letting a few staff try a chat assistant on their own, with no shared rules and no measurement. Typically that saves the enthusiasts some time and nobody else any, and the seniors who review the work stay sceptical, often for good reasons.
There's a better order to do this in. This guide is for accounting and bookkeeping firms of 10-60 people - a people business, meaning a business that sells its people's expertise and time, like accountants, lawyers, agencies and consultants. It covers where AI genuinely helps an accounting practice today, where it doesn't, and how to roll it out in a way your managers and partners will sign off. Our view is simple: AI drafts, people decide.
Why it happens
Firms struggle to get value from AI for three reasons, none of which is the technology.
The data is spread across apps
The useful context for any client sits in Xero or QuickBooks, the practice management tool (Karbon, TaxDome), the proposal tool (Ignition), receipt tools like Dext and Hubdoc, and a lot of email. A general chat assistant sees none of it unless someone pastes it in, and pasting client data into an unapproved tool is exactly what a careful firm should avoid.
Nobody owns it
AI arrives as a personal habit rather than a firm decision, so there are no agreed uses, no review step and no way to tell what's working.
The wrong work goes first
Firms often try AI on judgement work - tax positions, technical answers - where errors are costly and seniors are rightly cautious, instead of the repetitive work around the edges where AI is reliable and mistakes are easy to catch.
The fix with what you already have
You can make real progress without building anything.
- 01
Write a one-page AI policy
Which tools are approved, what client data may go into them (often: none, unless the tool is covered by your firm's contract and data terms), who reviews the output, and how AI use is recorded on the file. Keep it short enough that people read it.
- 02
Start with the AI already inside your apps
Many accounting tools now include AI features: categorising transactions, extracting data from receipts, drafting emails, summarising client history. Because they run inside a tool you already have terms with, they're usually the safest place to start. Check each vendor's documentation for what the feature does and where the data goes, and turn on one at a time.
- 03
Automate first where errors are cheap and visible
Good first candidates:
- Drafting routine client emails: document requests, deadline reminders, "your return is ready to sign".
- Summarising long email threads before a manager picks up a client.
- First-pass categorisation and receipt extraction that a bookkeeper reviews.
- Turning meeting notes into action lists in the practice tool.
- Drafting the commentary for management accounts from figures a person has already checked.
- 04
Leave these with your people, for now:
tax positions and technical judgement, anything signed off as a professional opinion, final review of returns and accounts, and difficult client conversations. AI can help someone prepare; it shouldn't make the call.
- 05
Name an owner and a review rhythm
Pick one manager to own AI use for a quarter. Their job is small: keep the approved list current, collect examples of what worked and what went wrong, and bring them to a monthly team meeting. Hearing a colleague explain how they used a drafted reminder, and what they had to fix, does more for adoption than any training video. It also surfaces mistakes early, while they are cheap, and gives sceptical seniors a regular, low-stakes place to raise concerns.
- 06
Measure before and after
Ask each person to log a week of repetitive work before you start, and again a month later. Our free hours lost worksheet gives you the structure.
Setting up the built-in AI features, app by app
Every vendor changes its menus often, so treat these as the order of work rather than exact clicks, and check each vendor's help pages for where the setting lives on your plan.
Xero and QuickBooks
Start with bank reconciliation suggestions and transaction categorisation, because a bookkeeper already reviews every line there. Before you turn anything on, ask your bookkeeping lead to tidy the chart of accounts and the bank rules for your five largest bookkeeping clients - suggestions learn from past coding, so messy history produces messy suggestions. Then turn the feature on for those five clients only, and have the bookkeeper note each suggestion they change in the first two weeks. If the change rate is falling, widen it to the next batch of clients.
Dext and Hubdoc
Data extraction from receipts and bills is the most mature AI in most practices. The setup work is in the rules, not the AI: supplier rules that set the default account and tax rate, and a clear owner who clears the review queue daily. A bookkeeper should still check anything over an agreed value, anything from a new supplier, and anything where the tax treatment isn't routine.
Karbon and TaxDome
Use the AI here for reading and writing around the work: summarising a long client email thread before a manager picks it up, and drafting routine replies. Agree as a team that a drafted email is always read in full and edited before it goes, and that nobody sends a draft that mentions a figure they haven't checked against the ledger.
Ignition
Proposals and engagement letters carry legal and pricing weight, so keep AI to first drafts of the covering message and the plain-English service descriptions. Your partner or practice manager still owns the scope, the price and the terms.
A checklist before you switch anything on
- The one-page policy is written, agreed by the partners and shared with everyone.
- Each feature has a named owner who will check it weekly for the first month.
- You know where the vendor says the data goes, and it fits your policy.
- Staff know how to tell a client, if asked, how AI is used on their work.
- You've picked a small group of clients to start with, not the whole book.
- You've logged a baseline week of the task you're trying to speed up.
For example: one week at a 20-person firm
For example, picture a 20-person firm with two partners, four managers, eight accountants and bookkeepers, and a small admin team, a month after rolling out the steps above.
On Monday, the practice manager clears the receipt review queue from Dext before 10am, as the policy says. Two extracted bills have the wrong tax rate because a supplier changed its invoices; she fixes the supplier rule, so the mistake doesn't repeat. On Tuesday, a manager taking over a client from a colleague on leave asks Karbon's assistant to summarise a 40-message thread, reads the summary, then skims the last five emails himself to check nothing was missed. He finds one open question the summary glossed over and adds it to his task list.
On Wednesday, the admin team uses drafted reminders for the 30 clients with returns due next month. Each draft is read and edited; three get rewritten completely because those clients need a phone call, not an email. On Thursday, a senior accountant drafts management accounts commentary from figures she has already reviewed, then cuts it by a third because the draft repeats the numbers instead of explaining them. On Friday, the AI owner spends 20 minutes collecting the week's examples - what helped, what needed fixing - for the monthly meeting.
Nothing in that week was decided by AI. Every draft was read by a person who knew the client.
Common mistakes
Turning on every feature at once.
Nobody can tell which one caused a problem, and trust drops across the board.
Skipping the baseline.
Without a before-week, you can't show the partners anything except feelings.
Letting drafts go out unread.
One wrong figure in a client email undoes months of goodwill. Reading every draft is the rule, not a suggestion.
Starting with the sceptics' work.
Begin with the people who want to try it and the tasks they find tedious; the sceptics come round when they see colleagues' results.
Treating the policy as finished.
Vendors add features every few months. Review the approved list each quarter.
Measuring logins instead of hours.
People opening a feature tells you nothing; time back on a named task does.
What good looks like
After three months, a firm that has done this well can say: which AI features are approved and why; who owns each one; how many hours a week the team gets back on three or four named tasks; and what mistakes were caught and how. Seniors can tell at a glance what was AI-drafted. Nobody pastes client data into a tool that isn't on the list, because the approved tools cover the jobs people actually need doing.
When that stops working
Built-in features and a good policy take most firms a long way. These are the signs you've reached the edge:
The useful AI task needs data from several apps
, such as a client summary that pulls the ledger, open jobs and recent emails together.
People are pasting client data
into general tools because the approved ones can't do the job.
Each app's AI works differently
, so staff learn five features and trust none.
There's no review step
that a manager can see, so seniors can't tell what was AI-drafted.
The time saved is real but uneven
, concentrated in two enthusiasts while everyone else carries on as before.
You want AI to draft from your own templates and tone
, and the built-in features use the vendor's.
What a custom build looks like
A custom AI build for an accounting firm is usually small and specific: it connects to your apps through their official connections, does one job well, and always hands the result to a person to check. We build these for accounting firms in four steps - scope, build, roll out, improve.
Examples:
- Client briefing before a meeting. One page per client drawn from Xero or QuickBooks, Karbon and recent emails: what's open, what's late, what changed. The manager reads it in two minutes.
- Drafted chasers and updates in your firm's tone, from live job status, queued for a person to approve and send.
- Management accounts commentary drafted from reviewed figures, with every number linked back to the ledger so the reviewer can check it.
- Inbox triage that files client emails against the right job and flags anything urgent.
- An approval trail. Everything AI drafted is labelled, and every approval is recorded, so seniors can see exactly what was checked and by whom.
Rolling it out matters as much as the build. We train the team, start with the people who'll use it most, and address "will this replace me?" directly: it won't, and the tools are designed so that a person decides every time. Seniors trust AI when they can see its work and override it easily, so that's how we build it. Internal tools like this often live in a simple hub the team already uses, like the team hub we built for an agency.
What does it cost? Our guide to AI consultant cost covers market rates. Forge's prices are on the pricing page: a two-week ops audit at $2,500 (£2,000), credited to your first month, then a monthly plan at $2,000, $3,500 or $7,500 a month (£1,600, £2,800 or £6,000), with your first build included. The software runs on your own accounts, so you own it, client data stays in your systems, and you pay those tool costs directly. We keep it monitored and fixed.
Most AI builds depend on clean client records, so our guide to syncing Xero or QuickBooks with your practice management tool is a good next read.
Rolling it out and training the team
The rollout decides whether any of this sticks. A pattern that works for most firms of 10-60 people:
- Weeks one and two: a pilot group. Three or four volunteers - ideally one admin, one bookkeeper and one manager - use one or two features on a small set of clients. They keep a short note of every draft they changed and why.
- Week three: show, don't tell. The pilot group shows real examples at a team meeting, including the mistakes. Seeing a colleague catch an error is what convinces sceptics that review works.
- Weeks four to eight: widen by role. Bring in the rest of each role together, so a whole team shares one way of working. Give each person a 30-minute walkthrough on their own client work, not a generic demo.
- Ongoing: a monthly check-in. The AI owner brings the examples, the hours logged and any policy questions. Partners attend at least the first three.
Address the worry head on. Someone in the room will wonder whether this is the first step towards needing fewer of them. Say plainly that the aim is to take repetitive writing and reading off their plate so they can spend more time on review, advice and clients, and back it up by asking each person what they'd do with two extra hours a week.
Measuring whether it worked
Keep measurement simple enough that people actually do it:
Hours on named tasks.
Compare the baseline week with a week at one month and three months, for the same tasks.
Edit rate.
Roughly how many drafts went out lightly edited, heavily edited or rewritten. A falling edit rate means the setup is improving; a stubbornly high one means that task isn't a good fit yet.
Errors caught.
How many mistakes reviewers found before anything reached a client. Some is healthy - it proves review works.
Turnaround.
Days from receiving a client's records to a draft set of accounts, or from a client email to a reply.
Who uses it.
If the hours saved still sit with two enthusiasts after two months, fix the training, not the people.
If the numbers say a feature isn't helping, switch it off. Dropping something that doesn't work is a sign the process is working.
Who gets time back, and what they do with it
The best early AI wins in an accounting firm go to the people with the most repetitive writing and reading: admins, bookkeepers and managers.
When routine emails draft themselves, your admin spends the time on clients who need a real conversation. When a client briefing builds itself before a meeting, your manager walks in prepared and spends the meeting on advice rather than catching up - for example, raising a cash-flow concern the client hadn't spotted. When commentary drafts itself from checked figures, your senior spends the time on the parts of the review that need judgement.
AI doesn't make your accountants less necessary. It gives them more time for the work only they can do.
Questions, answered plainly
What should an accounting firm automate with AI first?
Start where errors are cheap and easy to spot: routine client emails, thread summaries, first-pass receipt and transaction categorisation reviewed by a bookkeeper, and meeting notes turned into tasks. Leave technical judgement and sign-off with your people.
Is it safe to put client data into AI tools?
Only into tools your firm has approved, under contract and data terms you've checked. Many firms start with AI features inside apps they already use, and write a one-page policy on what data may go where.
Does AI change what accountants do?
It changes where their time goes. In a well-run rollout, AI drafts and people decide. It takes repetitive writing and reading off your team so they spend more time on review, advice and clients.
How do you get senior accountants to trust AI?
Make its work visible: label everything AI drafted, keep a clear approval step, link figures back to the source, and start with low-risk tasks so seniors see it working before it touches anything that matters.
How much does a custom AI build cost for an accounting firm?
Market rates vary widely; our AI consultant cost guide covers them. With Forge, a two-week ops audit is $2,500 (£2,000), credited to your first month, then $2,000, $3,500 or $7,500 a month (£1,600, £2,800 or £6,000) with your first build included.