FINANCIAL ANALYSIS
Let the ledger raise questions. Let evidence answer them.
An unusual journal is a starting point for investigation. A convincing AI explanation is still only a hypothesis until it is supported.
By CA Tony Thomas · Chartered Accountant · 5 minute read
Published
FINANCE · JUDGMENT · TECHNOLOGY / 02 OF 04
From a variance to a reviewable question
A general ledger can contain thousands of perfectly valid transactions and a small number that deserve immediate attention. Intelligent investigation should help a finance professional find those items and ask better questions. It should not turn an unusual pattern into an allegation or infer an accounting error from a statistical score.
The most useful output is an evidence packet: the movement, the transactions contributing to it, the tests performed, the missing information and the next decision owner. This builds on AI-assisted commentary, but starts earlier in the process, when the explanation is still uncertain.
My preferred discipline is to maintain three separate fields: observed fact, possible explanation and verified conclusion. The boundary between those fields is where much of the professional value sits.
A variance bridge that actually ties
Imagine a fictional UAE trading business whose monthly freight expense rises from AED 200,000 to AED 310,000, a 55% increase. A model might immediately blame fuel prices or higher sales. Neither explanation is established by the account balance.
A reproducible bridge could instead isolate AED 45,000 of additional shipment volume at the prior-period rate, AED 25,000 of rate and shipment-mix effects, AED 60,000 of late invoices for an earlier service period, and an AED 20,000 credit. Together these explain the AED 110,000 movement: 45,000 + 25,000 + 60,000 − 20,000.
Those figures are illustrative findings, not facts inferred from the two totals. They require shipment records, matched invoice lines and an explicit methodology. A volume effect might use the change in comparable shipment units multiplied by the baseline rate. The remaining rate/mix effect then needs investigation for routes, weights and service levels; it is not automatically a supplier price increase.
Keep currencies, tax treatment, units and cut-off consistent. Avoid double counting the late invoices in both volume and timing buckets. Tie every bucket to transaction IDs and reconcile the bridge to the ledger. Unexplained residuals must remain visible.
Make detection transparent
Before ranking anomalies, reconcile the extract to the trial balance for each entity and period. Check row counts, debit and credit totals, currency fields, reversal links and the treatment of opening balances. Confirm whether the report includes posted entries only. Missing journals can make a normal account look unusually quiet.
Combine several explainable tests rather than relying on one opaque score:
- Movement tests: compare absolute and relative changes against budget, prior month and a relevant seasonal baseline. A tiny prior balance can create a dramatic percentage.
- Transaction tests: review new account/vendor combinations, potential duplicates and manual entries near cut-off. Document how references are normalised.
- Context tests: compare entries with contract periods, branch activity, reversal patterns and approved allocation logic.
- Coverage tests: review missing expected entries as well as unexpected ones. An absent payroll accrual may matter more than a large, valid invoice.
Round values, weekend postings and new vendors are review signals, not proof of misconduct. Tune thresholds with the account owner and record why a test exists. Preserve a view of the entire population so investigators are not trapped inside the model’s selection.
Constrain what the AI can say
Use ordinary calculations or approved analytical code for totals and bridges. Give the language model a structured, minimised evidence set and require it to cite record references for each factual assertion. Do not ask it to calculate the truth from an unrestricted ledger dump.
Using only the supplied evidence, separate observations from hypotheses. For each observation, cite record IDs. For each hypothesis, identify the evidence needed to test it and an alternative explanation. If evidence is insufficient, state that explicitly. Do not allege fraud, prescribe a journal, invent a document or treat a detection score as a probability.
A controlled output for the example would say: “Freight increased AED 110,000. Records F101–F109 contain AED 60,000 relating to earlier service dates. Confirm whether these items were previously accrued before assessing cut-off.” It would not say: “The finance team failed to accrue expenses.”
Check the cited records yourself. A reference can exist and still fail to support the statement attached to it. Treat retrieved descriptions and document text as untrusted inputs; they must not instruct the system to export records or change its rules.
Close the case with a decision
Give every investigation a status: open, awaiting evidence, supported explanation, adjustment proposed, or escalated. Record who reviewed the evidence and what changed. Where a potential duplicate is identified, verify invoice number, supplier identity, amount, service period and payment status before deciding what action is justified.
A financial adjustment follows the normal preparation and approval process. A suspicion of misconduct follows established confidential escalation procedures, with careful language and restricted access. Neither decision should be made by the model.
For management reporting, distinguish operational performance from timing effects. In the freight example, moving AED 60,000 out of the current-month analytical view may help explain trends, but it does not authorise changing the ledger or presenting an alternative performance measure without appropriate context and approval.
Measure investigation quality
A pilot should track which alerts lead to a supported finding, how many valid transactions require unnecessary review, and what material issues appear in sampled unflagged items. Include time to resolution and evidence completeness. An increasing number of alerts is not, by itself, evidence of a stronger control.
Retain challenging examples: a valid unusual entry, a quiet account with a missing accrual, and a plausible explanation that the documents contradict. Re-run these cases when rules or models change. Review performance by account and branch so a large population does not hide weaknesses in a smaller one.
The practical standard is simple: a colleague should be able to reproduce the numbers, follow the evidence and understand the decision without trusting the AI’s tone. That is what turns ledger analytics into financial insight.
Personal educational perspective, independent of my employer. All business examples are fictional and illustrative. Proposed workflows are not claims of completed deployments or measured results. Apply professional judgment and your organisation’s policies.
Further reading: NIST’s Generative AI Profile provides a voluntary cross-sector risk-management resource. The workflows and examples above are original proposals, not NIST-prescribed controls.