How AI Is Changing Tax Preparation

For decades, preparing a UK Self Assessment tax return has followed the same pattern: gather your documents, manually enter figures into a form (either on paper or on screen), double-check everything, and submit. The process is tedious, error-prone, and stressful for the 12 million people in the UK who file each year. Artificial intelligence is now changing that — not by replacing human judgement, but by automating the most time-consuming and error-prone parts of the process.

The traditional process and its problems

Traditional tax preparation — whether done by an individual or an accountant — involves several manual steps:

1. Document gathering: Collecting P60s, P45s, dividend vouchers, bank statements, mortgage statements, and other paperwork. For landlords or the self-employed, this can mean dozens of documents.
2. Data entry: Manually reading each document and typing the relevant figures into HMRC's online forms or accounting software.
3. Classification: Deciding which figures go where — is this payment a deductible expense or a capital cost? Is this income from employment, self-employment, or property?
4. Calculation: Working out tax owed, applying the correct rates, thresholds, allowances, and reliefs.
5. Review and submission: Checking everything before filing.

Steps 2 and 3 are where the vast majority of errors occur. Transposition errors (typing £12,570 as £15,270), missed income sources, and incorrect classification of expenses are all common. HMRC estimated that errors cost UK taxpayers an estimated £8.5 billion in the 2022/23 tax gap, with a significant proportion attributable to simple mistakes in returns.

How AI reads tax documents

The first major application of AI in tax is automated document reading. This involves several technologies working together:

Optical Character Recognition (OCR)

OCR converts an image of a document (a photo, scan, or PDF) into machine-readable text. Modern OCR systems — particularly those powered by deep learning — can handle:

• Poor-quality photos taken on a phone
• Documents at an angle or with wrinkles
• Handwritten notes alongside printed text
• Multi-page PDFs with varying layouts

Modern OCR is highly reliable on clean printed text, which makes it practical for financial data extraction, provided a person still reviews the results.

Document classification

Once the text is extracted, the AI needs to understand what kind of document it is looking at. A P60, a P45, a dividend voucher, and a bank statement all look very different and contain different types of information. Modern AI models are trained to recognise document types automatically based on layout, keywords, and structure.

Data extraction

After classifying the document, the AI extracts specific key-value pairs: total pay, total tax deducted, employer name, PAYE reference, National Insurance contributions, and so on. This is where AI excels — it can process a document in seconds that would take a human several minutes to read and transcribe.

Intelligent classification and categorisation

Beyond reading individual documents, AI can also classify and categorise transactions. This is particularly powerful for:

• Self-employed taxpayers who need to separate business and personal transactions in bank statements
• Landlords who need to identify rental income, mortgage interest, repairs, and other expenses from bank and credit card statements
• Investors who need to match dividend payments to specific companies and share classes

AI classification works by analysing transaction descriptions, amounts, frequencies, and counterparty names. A payment to “British Gas” from a landlord's bank account, for example, is very likely a utility bill for a rental property. A payment labelled “HMRC PAYE” is clearly a tax payment.

However, classification is not infallible. A payment to “Amazon” could be a business expense (office supplies) or a personal purchase (a book for leisure). This is why AI-assisted classification always requires human review — the AI makes a best guess, and the user confirms or corrects it.

Accuracy and confidence scoring

A critical feature of well-designed AI tax systems is confidence scoring. Rather than simply presenting extracted data as fact, the system assigns a confidence level to each piece of information:

This approach is vastly superior to a system that either requires full manual entry or presents AI-extracted data without any indication of reliability. Users can focus their attention on the items that need it, rather than re-checking everything.

What AI cannot do (yet)

It is important to be honest about the limitations of AI in tax preparation. Current AI systems cannot reliably:

• Make tax judgements: Deciding whether a cost is a repair or an improvement, whether income is trading income or investment income, or whether a non-standard relief applies — these require human expertise and knowledge of the taxpayer's circumstances.
• Handle complex tax planning: Multi-year loss relief, pension carry-forward calculations, capital allowances elections, and other advanced planning decisions require specialist knowledge.
• Guarantee accuracy: Even the best OCR and extraction systems produce errors. The legal responsibility for a tax return always lies with the taxpayer (or their agent), not with the software.
• Replace professional advice: For taxpayers with complex affairs — multiple properties, overseas income, trusts, capital gains — professional advice remains essential.

TaxGo's approach: AI where it helps, deterministic logic where it matters

At TaxGo, we use a hybrid approach that combines the speed of AI with the reliability of deterministic computation:

• AI handles document reading and data extraction. Upload a photo of your P60 or bank statement, and our models extract the relevant figures in seconds.
• AI assists with expense classification. For self-employed and landlord returns, our models categorise transactions and flag potential deductions.
• A deterministic tax engine handles all calculations. Every tax figure — Income Tax, National Insurance, student loan repayments, payments on account — is computed using published HMRC rates and rules. No language model is involved in any calculation.
• Human review is always part of the process. We show confidence scores for every extracted value and highlight anything that needs attention.

We believe this is the right balance. AI is transformative for speed and convenience, but tax calculations must be precise, auditable, and deterministic. A language model that “guesses” your tax liability is not acceptable, no matter how sophisticated it is.

The future of tax technology in the UK

The trajectory is clear: AI will continue to make tax filing faster, cheaper, and more accessible. Several developments are likely in the next few years:

• Making Tax Digital (MTD): HMRC's MTD programme is expanding to Income Tax Self Assessment from April 2026 for taxpayers earning over £50,000, and from April 2027 for those earning over £30,000. This will require quarterly digital reporting, making automated document processing more valuable than ever.
• Real-time bank integration: Open Banking APIs allow tax software to pull transaction data directly from bank accounts (with the user's consent), eliminating the need to upload statements at all.
• Pre-populated returns: HMRC already receives employment income data via RTI. As more data sources become digital, it is likely that much of a Self Assessment return will be pre-filled — with the taxpayer reviewing and confirming rather than entering data from scratch.
• Natural language interaction: Rather than navigating forms, taxpayers may be able to describe their situation in plain English and have AI determine which parts of the return are relevant.