Machine Learning for Accountancy and Audit Firms

Sampling was a compromise, and data changes that

Traditional audit testing samples a small number of transactions because nobody could look at all of them. A mid-sized audit firm with 150 audit clients spends a large share of its hours on samples that find nothing, while the unusual entry that matters sits unexamined in the other 99 percent of the ledger.

Machine learning does not remove the need for professional judgement. What it does is let the team examine the whole population, score every entry for how unusual it is, and focus testing on the entries that deserve it. For accountancy practices doing bookkeeping and advisory work, the same techniques speed up coding and highlight clients heading for trouble.

Where machine learning helps accountants and auditors

• Journal entry testing. Scoring every journal for unusualness: odd posting times, rare account combinations, round amounts, unusual users or entries just below approval limits.
• Transaction coding. Suggesting nominal codes for bank transactions from past coding and the counterparty, with confidence scores for review.
• Duplicate and anomalous payments. Finding near-duplicate invoices and suppliers whose payment patterns suddenly change.
• Client risk and profitability. Predicting which clients will send records late, overrun fee estimates or need extra partner time.
• Workload forecasting. Estimating hours needed per job and per season so staffing is planned before the January rush.
• Going concern signals. Highlighting clients whose cash flow and payment behaviour look like early distress.

Journal entry anomaly detection in practice

Rules-based journal testing is well established: entries posted at weekends, by senior staff, to unusual accounts. The problem is volume. On a large client, those rules flag thousands of entries, most of them explicable, and the team either tests a sample of the flagged ones or quietly stops trusting the list.

Unsupervised anomaly detection learns what normal looks like for that client and scores how far each entry departs from it, combining many weak signals into one ranking. The top of the list is much more likely to contain something worth asking about.

1. Load the full general ledger and trial balance for the period
2. Profile normal patterns by account pair, user, timing and amount
3. Score every entry and review the highest-ranked with an explanation of why each scored high
4. Record the review in the audit file with the auditor's conclusion
5. Feed confirmed false positives back so next year's model is sharper

The model finds the entries worth asking about. It does not decide whether they are fraud or error, and the file must show a person reached that conclusion.

Professional responsibility does not move

Auditing standards place responsibility for the audit opinion and the judgements behind it on the auditor, whatever tools are used. Regulators have also made clear they expect firms to understand, test and document any technology they rely on. A firm that cannot explain how its anomaly model works, or show that it was validated, has a quality management problem.

For practical purposes that means keeping models explainable, recording versions and parameters per engagement, and treating the tool as part of the audit methodology rather than a black box a vendor looks after. Client confidentiality also rules out sending ledgers to any service whose data handling you have not reviewed.

Transaction coding for practices doing bookkeeping

For a practice keeping books for 300 small businesses, coding bank transactions is a large chunk of junior time. Cloud accounting platforms such as Xero already offer bank rules and suggestions, and for many practices those are enough.

A custom model becomes worthwhile when the practice has clients with messy, varied transactions where built-in suggestions struggle, or when it wants one model trained across many clients' coding patterns. Our post on machine learning for spend classification covers the technique, and automation for accountancy practices covers the workflow around it.

Predicting late, difficult and unprofitable clients

Every practice has clients whose records arrive three weeks before the deadline in a carrier bag, and clients whose fixed fee quietly became a loss years ago. Partners usually know the worst few by name. What they rarely have is a view of the next tier down, or of which newer clients are heading the same way.

Practice management systems hold the history needed: when records were requested and received, time recorded against each job, write-offs, chasers sent and fee changes. A model trained on that history can score each client before the busy season, so the practice chases the likely late ones early, reprices where time consistently overruns and plans staff around the jobs most likely to land late.

• Days between records request and receipt, by client and year
• Recorded time against fee, to show which fixed fees are underwater
• Volume of queries raised per job as a measure of record quality
• Changes in payment behaviour on the practice's own invoices

When it is the wrong investment

Commercial audit analytics tools exist and are the right answer for many firms. Custom work makes sense for firms with specialist sectors, unusual client systems or a desire to own their methodology.

How we would approach it

At SpiderHunts we would start with last year's completed audits: load the ledgers, run the anomaly scoring retrospectively and ask the engagement teams whether the top-ranked entries were the ones they would have wanted to see. That test costs little and gives an honest read on value before anything touches a live engagement.

Data extraction from varied client systems is usually the largest piece of work, followed by building a review screen that fits the audit file. Our machine learning development service covers both the models and those tools.

Frequently asked questions