Can AI automate tax return preparation?

AI can automate 50 to 70 percent of straightforward personal and small business tax return preparation. It handles data extraction from bank statements and source documents, categorisation of income and expenses, initial calculations, and draft return generation. Complex returns involving unusual income sources, cross-border elements, or disputed items still require significant professional judgment. The practical goal is not full automation but a workflow where AI does the assembly and the accountant does the review, cutting preparation time by half or more.

Short answer: AI automates 50 to 70 percent of straightforward tax returns: data extraction, categorisation, and drafting. Complex items need human judgment. Preparation time drops by half or more.

What AI actually does in tax return preparation

The tax return preparation workflow has distinct stages, and AI’s capability varies at each one. Understanding this prevents both over-optimism and unnecessary caution.

Stage 1: Data collection and extraction

This is where AI adds the most value. Traditional preparation starts with a box of receipts, a stack of bank statements, or a login to accounting software. Someone manually extracts figures, reads P60s, and types numbers into preparation software.

AI handles this by reading documents (optical character recognition plus language understanding), extracting relevant figures, and mapping them to the correct fields. Modern systems handle:

• Bank statements: transaction dates, amounts, descriptions, and running balances
• P60s and P45s: gross pay, tax deducted, NI contributions
• P11Ds: benefits in kind with descriptions and values
• Dividend vouchers: amounts, dates, and company details
• Rental income statements: gross rent, expenses, net income
• US equivalents: W-2s, 1099s, K-1s, and 1098s

Accuracy on clean, well-formatted documents: 95 to 99 percent for standard items. Accuracy on handwritten receipts, unusual document formats, or poor quality scans: 70 to 85 percent with flags for human review.

Stage 2: Categorisation and mapping

Once data is extracted, it needs to be categorised into the correct tax return boxes. Employment income goes to one place. Rental income goes to another. Business expenses need to be categorised by type.

AI categorises based on transaction descriptions, historical patterns, and learned rules. A payment to “HMRC PAYE” is tax. A payment from “ABC Ltd Salary” is employment income. A payment to “Tesco” from a sole trader might be business supplies or might be personal, and AI learns the client’s patterns over time.

For established clients with consistent transaction patterns, categorisation accuracy reaches 90 to 95 percent. For new clients, accuracy starts lower (75 to 85 percent) and improves as the system learns their specific patterns.

The critical design principle: AI should categorise with high confidence and flag with low confidence. A system that forces every uncertain item through human review is more useful than one that guesses wrongly and creates hidden errors.

Stage 3: Calculation and draft preparation

With data extracted and categorised, AI performs the arithmetic: totalling income categories, applying allowances and reliefs, calculating tax liability, and generating a draft return.

For straightforward returns, the calculations are deterministic once the inputs are correct. AI does not add creative insight here. It adds speed and consistency. A system that processes 100 returns will make the same calculation 100 times without the fatigue errors that affect humans on return number 87.

Where AI adds genuine value in calculations is identifying applicable reliefs and allowances. A well-trained system flags: “This client’s property income suggests they may benefit from the property income allowance” or “Marriage allowance transfer could reduce liability by £252.” These prompts ensure the reviewing accountant considers all options.

Stage 4: Anomaly detection and quality control

AI compares the draft return against prior years, statistical norms, and known risk areas. It flags:

• Significant year-on-year changes in any income or expense category
• Expense ratios that fall outside normal ranges for the client type
• Missing supplementary pages that may be required based on the data
• Inconsistencies between different data sources
• Items that HMRC or IRS algorithms are known to query

This layer catches errors that manual review might miss, especially in high-volume practices where reviewer fatigue is a real factor.

What AI cannot do

Complex tax planning

AI can flag that a client might benefit from a particular relief. It cannot design a tax planning strategy that considers the client’s broader financial situation, future plans, risk appetite, and the interaction between multiple tax regimes. This is where accountants earn their fees, and AI is nowhere near replacing this judgment.

Novel or disputed items

A capital gains calculation involving a property purchased in 1995, improved in 2003, partially let in 2010, and sold in 2026 requires professional judgment that AI cannot reliably provide. The interaction of principal private residence relief, letting relief, and various improvement costs involves interpretation, not just calculation.

Similarly, items where the tax treatment is genuinely uncertain, such as whether an activity constitutes trading, whether expenses are wholly and exclusively for business, or whether a payment is income or capital, require professional judgment that AI should not be trusted with.

Client-facing advisory

The conversation where an accountant explains to a client why their tax bill is higher than expected, what they can do differently next year, and how to plan for upcoming changes is fundamentally human. AI can prepare briefing notes for that conversation. It cannot have it.

The practical implementation

For UK firms: self-assessment and MTD

The UK tax calendar creates natural implementation windows. The optimal approach:

April to June: Implement and configure AI systems. Train staff. Process test returns from the prior year to validate accuracy.

July to September: Begin processing early-filed returns with AI assistance. Refine configurations based on real results. Build confidence.

October to January: Full deployment for self-assessment season. AI handles data extraction and draft preparation. Accountants focus on review, advisory, and complex returns.

Making Tax Digital has been a significant enabler. Firms whose clients maintain digital records through MTD-compliant software have cleaner data, which means higher AI accuracy. The firms that resisted MTD digitisation are now finding that their data is not ready for AI automation.

For US firms: individual and small business returns

The US tax season (January to April) is even more compressed. Implementation should begin no later than September for deployment during the following filing season.

US-specific considerations: state tax returns add complexity, and AI systems need to handle federal-state interactions. Multi-state filers require systems that understand nexus rules and apportionment. These add cost and complexity to the AI build.

Cost and ROI analysis

For a mid-market firm processing 500 straightforward personal tax returns per year:

Without AI:

• Average preparation time: 3.5 hours per return
• Total preparation hours: 1,750
• Staff cost at £35 per hour blended: £61,250

With AI:

• Average preparation time: 1.5 hours per return (review and refinement)
• Total preparation hours: 750
• Staff cost: £26,250
• AI system cost (year one, custom build): £35,000
• AI system cost (ongoing annual): £6,000

Year-one savings: £61,250 minus £26,250 minus £35,000 = break-even Year-two savings: £61,250 minus £26,250 minus £6,000 = £29,000 Year-three and beyond: £29,000 per year ongoing

The economics improve with volume. A firm processing 1,000+ returns sees positive ROI in year one. A firm processing 200 returns may need SaaS tools rather than a custom build to make the numbers work.

The indirect benefits often exceed the direct savings: faster turnaround improves client satisfaction, freed staff time allows advisory upselling, and consistent quality reduces compliance risk.

What we recommend

Start with your most straightforward return type, typically PAYE employees with simple financial affairs, and automate that first. Measure accuracy, time savings, and staff feedback. Once that workflow is proven, extend to more complex return types incrementally.

Do not try to automate your most complex returns first. The gap between AI capability and the professional judgment required is largest for complex work. Start where AI is strongest and build confidence and capability before tackling harder problems.

The firms getting the best results treat AI as an investment in their business model, not just an efficiency tool. When AI handles 60 percent of preparation work, the firm can process more returns with the same headcount, or, better, redirect staff from preparation to advisory work that commands higher fees and builds deeper client relationships.