AI in auditing and tax: the use cases that work

Every accounting firm has sat through the same AI demo by now. A chatbot answers a tax question, everyone nods politely, and six months later nothing in the practice has changed.

That isn’t because AI doesn’t work in auditing and tax. It’s because the demos focus on the wrong thing. The useful work isn’t a chatbot. It’s the unglamorous middle of the job: extracting data from documents, matching transactions, and flagging anomalies.

Here are the use cases that hold up in a real practice.

Full-population audit testing instead of sampling

Traditional audit testing relies on sampling for a simple reason: no human can read a million ledger lines. Software can.

Running anomaly detection across an entire general ledger surfaces the entries a sample would statistically miss. Postings made at weekends, round-number reversals sitting just under approval thresholds, or a supplier who invoices exactly once a quarter for exactly the same amount.

The auditor still makes the judgement call. The difference is that they now make it with the full population in view, not a 5% sample and a hope that nothing important fell outside it.

Document extraction that finally works

The same shift from sampling to full coverage applies to paperwork. Turning invoices, bank statements, and receipts into structured data used to mean OCR software that broke on every new template. Modern AI models read documents the way a person does: they find the total, the VAT line, and the date regardless of layout.

For a tax practice, that changes the economics of the worst part of the job. When a client sends a shoebox of receipts as photos over WhatsApp, that’s now a categorised expense schedule rather than a lost weekend.

One caveat: extraction is probabilistic, not perfect. Anything that feeds into a filing needs a human review step. Firms that skip the review end up redoing the work, with less trust in the tools than they started with.

Making Tax Digital and eTIMS are forcing the issue anyway

Even firms with no appetite for this are being pushed towards it, because digitising is no longer optional in either of the markets we work in.

In the UK, Making Tax Digital for Income Tax begins in April 2026 for sole traders and landlords earning over £50,000. It requires quarterly digital submissions and mandatory digital records. Practices still running on annual spreadsheet handovers will feel that immediately.

In Kenya, eTIMS already requires electronically generated tax invoices, and the KRA increasingly cross-checks what businesses claim against what their suppliers reported. When the tax authority is running matching algorithms on your data, running none on your own side leaves you at a real disadvantage.

The pattern is the same in both countries: the regulator digitised first. Adopting AI on the practice side is largely about keeping pace with a counterparty that already automated.

Reconciliation and transaction matching

Bank-to-ledger matching, intercompany eliminations, supplier statement reconciliations. This is matching under messy conditions: truncated references, bundled payments, currency noise.

Rules-based tools handle the clean 80% of transactions well. AI models are good at the ugly remaining 20%, like the “PAYPAL *JOHNSMI” reference a human would recognise in two seconds. Handing that 20% to software is where the real hours come back, because the difficult cases were always where the time went.

Where AI falls short in accounting

Some honesty, because the sales pitches won’t offer any.

Tax advice from a general-purpose chatbot is a liability. Models give confident answers to questions that depend on jurisdiction, tax year, and facts they don’t have. Fine as a starting point for a professional who will verify everything. Dangerous as an answer handed straight to a client.

Confidentiality is not a footnote. Client financials pasted into a consumer chatbot may be retained and used for training. The data-handling terms matter more than the model’s benchmark scores.

Professional judgement doesn’t automate. Going concern assessments, materiality, whether a director’s explanation actually adds up. The tools sharpen the evidence in front of the professional. They don’t replace the professional.

Where to start with AI in your practice

Not with a strategy deck. Pick one painful, repetitive, easily checkable process: expense categorisation, statement reconciliation, or invoice data entry. Automate that one thing, keep a human review step, and measure the hours before and after.

If the hours don’t move, stop. If they do, you’ve earned the right to automate the next thing.

That’s the whole method. It’s also roughly the AI automation we build for clients: small, verifiable automations that survive contact with real work, rather than platform rollouts that stall at the demo stage. If you run a practice and want a second opinion on where AI would actually pay for itself, we do those conversations for free.