For the last decade, “AI in tax” meant pattern-matching on bank narrations and the occasional dropdown that guessed your category. That has now shifted. In 2026 a new category of tool has emerged — tax intake automation — and it sits in a different part of the firm’s workflow than the bookkeeping and filing software it gets compared to. This guide is for partners and senior staff evaluating where, and whether, to invest.
Intake is the unloved middle of the firm
A UK accountancy firm’s production line has three rough stages. Intake is everything between “client owes us their return” and “we have a complete, categorised dataset ready to prepare from.” That covers the chase emails, the paper bank statements, the missing dividend voucher, the “what was this £842 transfer?” conversation, the receipts mailed in a margarine tub. Preparation is the tax-rule work — the bit your people went to college for. Review and filing is the sign-off and submission to HMRC.
The middle stage is where the year disappears. Most firms spend 60–75% of their compliance hours on it. It is also where margin disappears, because intake work is hard to scope, hard to charge for, and hard to delegate to juniors when the client is unresponsive. For low-value compliance clients — the typical £350–£600 personal return — intake costs alone can wipe out the fee.
A worked example
Take a representative book of 200 personal-tax clients at an average fee of £450. Gross revenue is £90,000. If average partner / senior time per file is six hours all-in at a £75 blended rate, delivery cost is £90,000 — break-even. Where does the time go? On a sample we ran with three UK firms, the split was roughly:
- Chasing missing items: 1.8 hours
- Reading and categorising what arrived: 1.6 hours
- Resolving exceptions with the client: 0.9 hours
- Actual tax preparation (rules, judgement, computation): 1.1 hours
- Review, sign-off, filing: 0.6 hours
Four hours of intake per file, an hour of tax work, and half an hour of review. That ratio is the problem the new category of tool is going after.
What “tax intake automation” actually does in 2026
The tools that work in 2026 cluster around four capabilities. None of them are speculative; the underlying model quality crossed the line in late 2024, and the UK-specific tax tooling has caught up over the eighteen months since.
1. Document parsing that holds up on real client mess
P60s, P11Ds, P45s, dividend vouchers, interest certificates, completion statements, mortgage interest letters, SA302s, bank statements as PDF or as CSV, receipt photos with a finger over the corner. The 2026 generation of vision-capable models reads these reliably, including handwritten amounts and rotated scans. Importantly, they extract structured output — a dividend voucher becomes { payer, amount, date, tax_year } — not free text for a human to retype.
2. Context-aware chasing
The harder, less obvious capability. The tool has to know what is missing for this specific client’s return — not a generic checklist. A landlord with a new buy-to-let mid-year needs the completion statement; a director who started taking dividends needs a fresh dividend voucher; an employee who changed jobs needs both P60s. Then it has to ask for those items on the channel the client actually uses, in a tone that isn’t robotic, and follow up without bothering the partner.
In practice this is what separates intake automation from a smarter inbox. A smarter inbox files the document; intake automation knows which document is still owed and goes and gets it.
3. Categorisation against UK tax rules
Categorisation that maps to the actual lines on the SA100, SA105, SA106, or future MTD ITSA quarterly schema — not to a generic chart of accounts. That means the tool understands Section 24, the property allowance, the replacement of domestic items relief, Form 17 splits on jointly-held property, and the difference between capital and revenue expenditure. If the categorisation layer can’t hold these distinctions, the firm ends up redoing it manually and the time saving evaporates.
4. Audit trail
Every decision the system makes — every category assigned, every chase sent, every piece of evidence relied on — has to be inspectable by the responsible accountant. Not just “the AI categorised this as repairs” but which bank line, which receipt photo, which client message confirmed the purpose. This is the bit firms underweight when evaluating and regret six months in.
What it doesn’t do, and shouldn’t pretend to
Intake automation is intake automation. It is not a tax adviser and it does not replace one. Specifically, the tools worth looking at in 2026 are explicit about staying out of:
- Nuanced professional judgement. Whether a particular extension is a capital improvement or a like-for-like repair, whether the trading-allowance election or actual expenses is better, whether to claim the property allowance — these are partner calls.
- Complex tax-position questions. Non-domicile status, the remittance basis, mixed-use property, partnership transitions, EIS / SEIS chains, employee share schemes.
- Novel structures. Any client whose return isn’t mostly off-the-shelf needs human-led prep, full stop.
- Signing the return. The responsible accountant signs. Nothing should obscure that.
A vendor that blurs these lines — “our AI prepares the return end-to-end” — should be treated with the same scepticism you’d treat a vendor that promised to do your firm’s reviews for you. Press them on the audit trail and the answer will usually clarify what’s really happening underneath.
An honest evaluation framework
If you’re looking at this category, four tests separate serious tools from demoware.
Categorisation accuracy on a real test set
Take 200 transactions from a real client — anonymised — across rent receipts, repairs, capital improvements, mortgage interest, letting agent fees, insurance, and the awkward ones (deposit returns, transfers between joint owners, one-off legal fees on a remortgage). Ask the vendor to categorise them blind, then measure accuracy against your team’s answers. Anything below 90% is not ready. Anything that can’t tell you why it picked a category is not ready either.
Audit-trail completeness
Pick five transactions at random from the system. Can you, in under sixty seconds, see who said what to whom, which document was relied on, what category was assigned, and who confirmed it? If the answer is “the AI categorised it” without an inspectable provenance chain, walk away.
UK tax-engine accuracy
If the tool produces numbers — totals, profit calculations, adjustments — those numbers should match HMRC’s own worked examples to the penny. Ask to see the test suite. Most serious UK tools now publish golden tests against the HMRC manuals (PIM, BIM, SAIM). If the vendor can’t show you theirs, assume they don’t have one.
Client-channel coverage
The intake tool is only as good as the channel your clients actually use. WhatsApp is now the dominant channel for personal clients in the UK; email still matters for older books; a small minority of clients want a portal. The tool should meet them where they are, not force them to a new login.
The integration question: Phase 1 vs Phase 2
The other thing worth getting clear before signing anything is where the tool stops in the workflow. Today there are two credible shapes.
Phase 1 — validated dataset, prepared return handed over. The tool does intake and categorisation, then exports a clean, structured dataset (or a draft return) into the firm’s existing tax software. The firm reviews and submits. This is the right shape today for any firm that already has working tax software, because it leaves your filing relationship with HMRC unchanged and keeps the responsible accountant’s sign-off obvious.
Phase 2 — direct-to-filing. The tool submits to HMRC itself. There’s a place for this — quarterly MTD updates are the obvious wedge — but the demand should be pulled from the firm, not pushed by the vendor. If the tool can’t operate cleanly as a Phase 1 product first, Phase 2 will be worse, not better.
The healthy pattern in 2026 is to start Phase 1, prove the accuracy and audit trail across one tax season, and let the firm decide whether and when to push more of the filing layer into the same tool.
Where this category is heading
The interesting thing about intake automation as a category is that it isn’t a feature of tax software. It’s a separate layer that sits in front of it. Five years ago the firm either did intake by hand or pretended their bookkeeping tool covered it. Neither is true any more, and the firms that figure this out first will quietly take the low-value-compliance market off the firms that don’t — because they can profitably serve clients other firms can’t.
If you’re a partner evaluating this category: the questions to ask are the four above. The answer that matters most is the audit-trail one. Everything else can be fixed; opaque automation can’t.
We built one of these — happy to show you
Otto is a tax intake automation engine built for UK firms. Clients send documents and answers by WhatsApp; we read, categorise, chase what’s missing, and hand you a prepared return ready to review and file in your existing software. If you’d like to see what it looks like on your kind of book, book a 30-minute demo.
Book a demo