AI for accountants is discussed almost entirely at the wrong altitude. The conversation is about tools — which bookkeeping engine reads receipts best, whether a language model can draft a set of accounts — when the constraint in most UK practices is not analysis at all. It is the volume of low-value contact required to get information out of clients, and the fact that this volume is about to rise.
This guide is written for practice owners rather than procurement. It covers what genuinely changes, what does not, and the order to do things in.
The constraint is chasing, not computing
Ask a practice where January goes and nobody says "preparing returns". They say chasing. Chasing bank statements, chasing a missing dividend voucher, chasing the client who has not opened an email since November.
That work has three properties that make it unusually well suited to automation. It is repetitive, it follows a known sequence, and it produces no professional judgement. Nobody bills for the fourth reminder. The work also has a hard edge: HMRC must receive an online Self Assessment return by 11:59pm on 31 January or a late filing penalty follows, and the same date is the payment deadline. The deadline does not move because a client was slow.
Automating the chase does not mean sending more email. It means the practice stops being the thing that remembers. A system that knows which documents are outstanding per client, escalates on a schedule, and stops the moment a document arrives removes an entire category of work without touching a single judgement call.
Making Tax Digital is the forcing function
The reason this stops being optional is arithmetic.
Making Tax Digital for Income Tax applies from 6 April 2026 to those with qualifying income above £50,000 for the 2024 to 2025 tax year. It extends to £30,000 for the 2025 to 2026 tax year from 6 April 2027, and to £20,000 for the 2026 to 2027 tax year from 6 April 2028. You can check the current position on the GOV.UK guidance, and you should — these thresholds have been revised before.
Read that as a client-communication problem rather than a filing one. Each affected client moves from an annual rhythm to a quarterly one. The practice does not acquire four times the work, but it does acquire something close to four times the touchpoints: four windows in which records must be complete, four rounds of reminders, four opportunities for a client to go quiet.
A practice that handles the annual cycle through personal diligence — a senior who knows who is slow and rings them — does not scale to quarterly by trying harder. That is the moment automation stops being an efficiency project and becomes a capacity one.
What AI cannot do, and why saying so matters
The honest limits are as commercially important as the capabilities, because a practice that automates past them creates regulatory exposure.
Anti-money-laundering risk decisions stay with a qualified person. Automation can collect identity documents, run the checks, chase what is missing and evidence the sequence. It cannot decide whether a client relationship is acceptable. Supervision for accountancy sits either with a professional body — ICAEW, ACCA, AAT or CIOT among them — or, for accountancy service providers not supervised by a professional body, with HMRC. Either way the responsible person is a person.
Advice is not a generation problem. A model can draft a paragraph explaining a tax treatment. It cannot decide whether that treatment is right for this client, and it will produce a confident paragraph either way. The failure mode is not obvious errors; it is plausible ones.
Client relationships do not automate. The reason a client stays is usually that somebody knew their business. Automating the parts that never needed a human protects that. Automating the parts that did erodes it.
A useful test: if the task requires professional judgement, or the client would be annoyed to learn it was automated, it stays with a person.
A sensible order
Most practices try to do this in the wrong sequence, starting with the most visible tool rather than the biggest constraint.
- Fix the intake. Before automating anything internal, look at what happens when a prospective client finds the practice. If enquiries arrive as an unqualified email with no indication of turnover, structure or incumbent accountant, every downstream step inherits that mess.
- Automate record collection. The highest-volume, lowest-judgement work in the practice. It is also the safest place to start, because a reminder that fires wrongly is an inconvenience, not a compliance event.
- Sequence onboarding up to the AML gate. Collect, verify and evidence — then stop and hand the risk decision to a qualified person.
- Then look at MTD comms. By this point the quarterly rhythm is a scheduling change rather than a redesign.
Notice that three of those four are communication problems, not accounting ones. That is the actual finding.
Where the website sits in this
The part practices most often skip is the front door. Record chasing and onboarding are both downstream of enquiry quality, and enquiry quality is set by the website.
A site that states pricing plainly, qualifies on the things that determine whether a client is a fit — sole trader or limited company, turnover band, VAT registration, whether there is an incumbent accountant — and books the call itself, does more for practice capacity than any internal tool. It changes what arrives, rather than processing the same mess faster.
We have written about that specifically for accountancy practices on our AI for accountants page, and about the build itself under services.
The short version
AI for accountants is worth doing where the work is repetitive, sequenced and free of professional judgement. That is a genuinely large share of a practice's January, and MTD is about to make it a share of every quarter.
It is not worth doing where judgement lives, and pretending otherwise is a supervision problem rather than a productivity one. Start with the front door, automate the chase, stop at the AML gate.