The pitch for artificial intelligence in public finance usually arrives with a forecast attached: predict revenue, predict overruns, predict default. Forecasting is the least interesting thing these systems do well. A three-person finance department in a town of forty thousand people does not have a prediction problem. It has an attention problem. Roughly the same number of documents arrive every month as arrived twenty years ago, plus a decade of new reporting obligations, and the same three people process them.
That reframing matters, because it changes what a useful system looks like. A tool that produces a revenue estimate replaces judgement the department already exercises reasonably well. A tool that ranks 9,400 purchase-card transactions so that the forty worth reading appear first does not replace judgement — it delivers judgement to where it is scarce.
Triage, reconciliation, and extraction
Three categories of work in a finance office reward automation disproportionately, and they share a shape: high volume, low individual stakes, catastrophic aggregate stakes if nobody looks.
- Triage. Ranking transactions, invoices, journal entries, or grant charges by how unusual they are relative to the entity's own history. The output is an ordered worklist, not a verdict.
- Reconciliation. Matching records across systems that were never designed to talk — a utility billing system, a general ledger, a bank statement, a grants portal. Fuzzy matching on vendor names, amounts and dates handles the ninety per cent of pairs that are obvious and surfaces the residue.
- Extraction. Pulling structured fields out of unstructured documents: lease terms out of PDFs for GASB 87 inventories, subscription terms for GASB 96, insurance certificates, bid tabulations.
None of these are glamorous. All of them are places where a finance director currently chooses between doing the work badly and not doing it at all.
A useful test before adopting anything
Ask what happens when the model is wrong. If the answer is "a person reads a transaction that turned out to be fine", the risk is a few wasted minutes. If the answer is "a vendor payment is blocked" or "a resident is flagged", the tool needs governance, an appeal path, and a human decision-maker on the record. The first class of tool can be piloted this quarter. The second should not be piloted at all until the governance exists.
What does not work as advertised
Three claims deserve scepticism.
"It learns your entity's controls automatically." Anomaly detection learns what is statistically unusual, which is not the same as what is improper. In a jurisdiction where an improper practice has been routine for six years, the improper practice is the baseline, and the model will flag the first correct transaction as the outlier. Detection models find deviation from history. They do not find deviation from policy unless someone encodes the policy.
"It replaces the need for reconciliation staff." In practice, headcount rarely falls. What changes is the mix: less matching, more investigating. That is a better use of a public accountant's time and usually a better outcome, but it should not be sold to a council as a personnel saving that will not materialise.
"The model explains itself." Feature attributions explain what the model weighted. They do not explain why a payment was improper. The gap between those two things is where audit findings live.
Sequencing an adoption
The sequence that tends to survive contact with a real finance department starts with the least consequential work and moves outward only when the previous stage has produced documented value.
| Stage | Work | Human role | Typical horizon |
|---|---|---|---|
| 1 | Document extraction into a register (leases, SBITAs, contracts) | Reviews and corrects every record | One quarter |
| 2 | Transaction triage for post-payment review | Investigates the ranked worklist | Two quarters |
| 3 | Cross-system reconciliation with exception queues | Resolves exceptions; owns match rules | Two to three quarters |
| 4 | Forecast support for revenue and expenditure lines | Sets assumptions; publishes variance analysis | A full budget cycle |
Nothing in that sequence touches a payment before it leaves the treasury. Pre-payment blocking is a fundamentally different risk posture: it can delay legitimate obligations, it interacts with prompt-payment statutes, and it needs an override procedure that is auditable and fast. Entities that jump to stage five before establishing stages one through four generally end up with an override that is used on everything, which is the same as having no control at all.
The governance that has to exist first
Before any model influences a public spending decision, four things should be written down and approved by someone accountable: what data the system may use, who may override it and how that override is recorded, how often the model is re-evaluated against a labelled sample, and what the entity will publish about it. The fourth is the one most often skipped and the one most likely to matter when a local reporter asks how the flagging works.
Public finance has an advantage here that the private sector does not: the norm of documented, reviewable process is already the culture. An entity that treats a model like any other internal control — documented, tested, periodically reassessed, owned by a named person — will get most of the way to responsible deployment without inventing a new discipline.
This publication is general information and is not legal, accounting, audit or financial advice. See our Disclaimer. Found an error? Write to [email protected] — we correct in place and note what changed.