How to test and control AI use in an accounting practice before relying on it

AI can support narrow accounting tasks, but it does not remove professional judgement, authority or responsibility. Start with a bounded use case and failure consequence, then decide whether automation, rules or ordinary software would solve the problem more reliably.

What to take away

Candidate applications

Potential uses include extracting fields from invoices, suggesting transaction categories, identifying duplicates or anomalies, prioritising exceptions, summarising controlled records, drafting routine client messages and supporting internal knowledge search.

Each application needs a precise input, output and decision owner. "AI bookkeeping" is too broad. A tool that proposes a code for reviewer acceptance has a different risk from one that posts and submits without a human checkpoint.

Map information before testing

Record people, data fields, sources, purpose, model or provider, location, recipients, retention and deletion. Decide whether real personal or confidential records are necessary for evaluation. Prefer sanitised test data until the arrangement is approved.

The ICO's artificial intelligence guidance covers data-protection considerations, AI risk tools and explanation resources. Apply current guidance to the actual system and role rather than treating an AI policy as approval for every feature.

Review whether provider settings permit reuse for model training or product improvement. Optional assistants embedded in existing software can introduce a new purpose or recipient even when the main contract is already approved.

Test realistic accounting cases

Build a test set containing ordinary and difficult records: blurred images, duplicate invoices, credit notes, mixed VAT treatment, handwritten detail, unusual suppliers, changed bank information and missing context. Protect the test data and preserve expected answers.

Measure false acceptance and missed exceptions, not only overall accuracy. One invented figure in a tax calculation can matter more than many correct low-risk fields. Test consistency after product updates and across relevant client groups.

Keep human review meaningful

Assign a reviewer with competence for the underlying task. Show source and suggestion together, preserve edits and prohibit automatic acceptance through workload pressure. Train users to recognise confident but unsupported text.

HMRC's standard for agents expects agents in scope to prevent unsupported or speculative figures, maintain relevant knowledge and keep contemporary records. An AI output does not change those expectations.

Control consumer-facing AI

If a chatbot or agent gives service information, quotes or recommendations, test the complete customer journey. The CMA's AI agents consumer-law guidance explains that businesses remain responsible for compliance when deploying AI agents in covered consumer activity.

Make role and limitations clear. Escalate technical, vulnerable-customer, complaint and high-consequence questions to a suitable person. Do not let generated language promise HMRC endorsement, guaranteed savings or work outside the engagement.

Monitor and preserve an exit

Track corrections, unsupported outputs, user overrides, incidents, response time and displaced work. Review changes to the model, supplier, subprocessor, data use and integration.

Compare the new process with its previous baseline. Include reviewer time, client clarification and downstream correction, not only faster first output. A use case that saves preparation time but increases technical review or hides uncertainty may not improve the full service.

Assign a release owner for product changes. Retest critical cases before an updated model or feature enters live accounting work, and document any accepted difference in behaviour.

Maintain a manual or alternative route for critical deadlines. Test export and deletion before dependence grows. Stop the use case when evidence falls below its acceptance threshold or the practice cannot explain and correct the result. Responsible adoption is a continuing control, not a one-off procurement decision.

Before you act

Common questions

What should a practice record before testing an AI tool?

Record the people involved, the data fields and sources, and the purpose. Also record the model or provider, the location, the recipients, and retention and deletion. Decide whether real personal or confidential records are necessary for evaluation, and prefer sanitised test data until the arrangement is approved.

How should a practice measure an AI tool's performance?

Test on ordinary and difficult records, and measure false acceptance and missed exceptions rather than overall accuracy alone. Check consistency after product updates and across client groups.

When should a practice stop using an AI use case?

Stop the use case when evidence falls below its acceptance threshold, or when the practice cannot explain and correct the result. Maintain a manual or alternative route for critical deadlines, and test export and deletion before dependence grows.