Self Assessment season: what to automate, and what to leave alone

Ask any practice where January goes and the answer is rarely “tax”. It goes on the client who sent three bank statements but not the fourth, the P60 that arrived as a photo taken at an angle, the dividend certificate that was there last year and has quietly gone missing this year, and the fifteenth email asking whether the return has been filed yet.

None of that needs a qualified professional. All of it lands on one anyway, because it sits between the parts that do.

That is where automation earns its keep in Self Assessment. Not in the calculation, and certainly not in the filing, but in everything that has to happen before a professional can sit down and do the work they are actually trained for.

The workflow, step by step

What follows is the sequence we build for accountancy practices. It runs from September to the end of July, because a Self Assessment season does not really end on 31 January.

Before the documents arrive

1. Segment the client list (September to October). Each client is sorted by type of income — employment, self-employment, property, dividends, capital gains — by whether they already fall within Making Tax Digital for Income Tax, and by the state of the practice’s agent authorisation with HMRC. The AI helps with the sorting, working from last year’s return. The practice checks the result and signs it off.

2. Ask each client for what they, specifically, need to send. Instead of the same generic checklist for everyone, each client receives a list built from their previous return: P60, P11D, letting statements, interest and dividend certificates, pension contributions, Gift Aid. If something does not arrive, reminders go out automatically and at increasing intervals, so nobody on the team has to remember who to chase.

While the documents come in

3. Receive and file every document. Clients upload to a secure portal. The AI recognises what each document is, pulls out the key figures and attaches them to the right client and the right tax year.

4. Check for gaps. What has arrived is compared with what was expected. The useful gaps are the ones nobody would spot by eye: last year this client declared interest from Bank X, and this year there is no certificate. Each gap triggers a reminder to the client and a task for the team.

5. Categorise the figures for sole traders and landlords. Bank transactions are mapped to the boxes on the SA103 and SA105. Anything that moves sharply against last year is flagged, and so is any expense that looks doubtful.

Preparing the return

6. Load the data into the practice’s tax software. The figures go into the HMRC-recognised software the practice already uses — TaxCalc, IRIS, CCH, Xero Tax or another — by import or through its API. The tax is calculated by that software, not by the AI.

7. Prepare a review memo for the professional. The AI drafts a summary of what has changed since last year, reliefs that may have been missed (higher-rate relief on pension contributions, Gift Aid, marriage allowance), cases caught by the High Income Child Benefit Charge, and clients who may be candidates for reducing their payments on account. A qualified professional reviews it and signs off.

Closing it out

8. Client approval. The client receives a plain-English summary of what they owe and when it is due, and approves it with an electronic signature.

9. Filing. Someone at the practice submits the return from the tax software. This step is not automated, and should not be.

10. After filing. The submission receipt is filed, the client’s status is updated in the CRM, and payment reminders are scheduled.

Seeing the whole season

11. A single view of the work. One dashboard shows every return by stage, how many days are left to the deadline, which clients are holding things up, and how much each person on the team is carrying. The partner no longer has to ask round the office to know where things stand.

What the AI does not do

This is worth saying plainly, because it is the first thing a partner will ask:

• It does not file returns or calculate the tax. That stays in the practice’s own software, with a person pressing submit.
• It does not make judgement calls on tax treatment. It flags; the professional decides.
• It does not replace professional review. The standard set out in PCRT, which ICAEW, ACCA, ATT and the other bodies expect their members to follow, still applies in full.

Everything the AI produces is a draft, and every draft leaves a trail: who checked each step, and when. If HMRC ever asks how a figure was arrived at, the answer is on record.

Two things to settle before building anything

UK GDPR. Client tax records are about as sensitive as personal data gets. Before anything is connected, you need a data processing agreement with every AI provider involved, a contractual commitment that your clients’ data will not be used to train their models, and an explicit decision about where the data is stored. Not a default somebody accepted during setup: a decision.

Do not automate a step you do not already run well by hand. If the practice has no consistent way of chasing documents today, automating the chase just sends inconsistent reminders faster. The manual process comes first. Once it works, it can be automated; before that, automation only hides the problem.

---

The practices that get the most from this are not the ones with the most technology. They are the ones who are clear about which part of the season is tax work and which part is simply moving paper, and who decide to stop paying qualified people to move paper.