AI Over Structured Financial Data

1The difference a number makes

Much of the discussion about professional AI concerns prose. In an accounting firm, the output may also include figures.

A wrong rate, total or balance can look exactly like a correct one. Fluency offers no warning. The figure must be checked against the underlying records and the method that produced it.

That is why structured financial data needs stricter controls than drafting.

2Never let the model do the arithmetic

The central rule is simple: a language model may help decide what to calculate, but it should not perform the calculation.

Arithmetic belongs in deterministic code, such as a query, spreadsheet function or reconciliation routine. That calculation can be reproduced and tested. The model can help formulate the query, read the result and draft an explanation.

Re-asking a model is not verification. Re-running a controlled calculation is.

The model may help write the query. It must not be the calculator.

3What this makes possible

Within that boundary, structured data can support several useful accounting workflows.

Reconciliation and matching. Examples include bank-to-ledger records, intercompany balances, supplier statements and payables, or e-invoice records and posted entries. A model may propose matches and explain residual items. The matching rules and calculations should run deterministically.

Anomaly and exception surfacing. A system may flag entries outside the usual pattern of an account, round-sum postings, period-end clusters, possible duplicate payments or unusual approval patterns. AI can help rank and explain the exceptions. The accountant decides what they mean.

Variance and commentary drafting. AI can prepare a first draft explaining movements against the prior period or budget. The draft should be grounded in figures returned by a controlled query, not in the model's recollection. The accountant checks both the figures and the explanation.

Plain-language querying. Staff may be able to ask a question without writing SQL. Malaysia's MyInvois programme increases the amount of transaction data available in a structured form, subject to the quality and consistency of each client's records.

4Every figure carries its lineage

Every figure surfaced by AI should be traceable to the calculation and source records that produced it.

The file should show the figure, the query or calculation, the rows used and the source system. Without that chain, the figure is not ready to support a professional conclusion.

In practice, the workflow should preserve the query alongside the answer so that a reviewer can re-run it. If the number cannot be reproduced and traced to source, it should not enter the file.

5Anomaly detection is not a conclusion

Anomaly detection can examine a full ledger population and rank entries that warrant attention.

An anomaly is a question, not an answer. A flagged entry is unusual according to the method used. It is not, by itself, evidence of error or irregularity. In an audit, treating a ranked exception list as a conclusion would substitute pattern matching for the auditor's scepticism.

Anomaly detection can change what the team examines. It does not determine the conclusion. Accountants must understand the method, investigate the selected items and apply professional judgement.

6Why this work should remain on premises

Structured financial data is the most concentrated form of client confidential information a firm holds.

A ledger extract may reveal trading patterns, customers, margins, bank relationships and payroll in a machine-readable form. Sending it to a public or third-party AI model can disclose far more than a single document.

The distinction is not between cloud storage and on-premises storage. It is between keeping AI processing under the firm's control and sending confidential client numbers to an external model.

Secure Accounting runs on premises and supports work with authorised client files without sending them to a public or third-party AI model. It sits beside the firm's existing systems. The ledger remains the official record.

7The honest bounds

Three limits matter.

Secure Accounting is not a system of record and does not post entries. Poor source data will limit the quality of any analysis. AI output can also be wrong, even when it is grounded in the firm's records.

Accountants must review the work, reproduce important calculations and verify the supporting records. They remain accountable for every figure, filing and professional conclusion. These professional obligations do not move to the tool.