Akshan KumarFinance leadership & transformation
← Back to insights
Pillar guide · AI in Finance

AI in finance: controls, exceptions and human review

AI works best in finance when it helps a named owner investigate an exception, interpret evidence or draft a conclusion after reliable rules have established the facts.

By Akshan Kumar · Chartered Accountant and finance transformation leader · Reviewed

The short answer

Do not begin with an AI assistant that writes commentary over unreliable numbers. Begin with a defined finance decision, governed source data and deterministic rules that separate normal transactions from exceptions. AI can then make the exception queue easier to understand, prioritise and resolve while finance retains approval and accountability. This is part of a wider finance transformation operating-model change, not a standalone technology deployment.

What this guide covers

  • Where deterministic finance controls should replace or precede AI.
  • How to design an exception queue with ownership, ageing and evidence.
  • Which finance tasks AI can assist without taking consequential approval.
  • How to measure value through decision time, control quality and reviewer effort.

How to implement AI in finance safely

  1. Define the population. Decide which transactions, balances, contracts or close activities are in scope.
  2. Apply rules. Test completeness, matching, thresholds, period, approval and accounting conditions.
  3. Auto-process the normal. Let repeatable, low-risk items move through the workflow with recorded evidence.
  4. Create an exception queue. Show the failed rule, value, risk, age, owner and supporting evidence.
  5. Use AI for interpretation. Classify patterns, summarise evidence, suggest questions and draft commentary.
  6. Escalate to a human. The controller or process owner decides, records the action and owns the audit trail.

What AI should do

AI is useful when the task involves language, investigation or prioritisation. It can retrieve an approved policy clause, group recurring reconciliation breaks, identify the evidence most relevant to a variance, or prepare a first draft for a controller. These tasks improve the speed and consistency of review without pretending that an interpretation is a control result.

What deterministic controls should do

Rules should calculate, match, validate and gate. Examples include actual-versus-budget calculations, reciprocal intercompany matching, missing-field checks, ageing thresholds, sequential revenue gates and approval limits. The same inputs and rules should produce the same conclusion. If the result depends on judgment, route it visibly rather than hiding it behind a generated answer.

How to measure value

Measure time to decision, exception ageing, repeat causes, value at risk, resolution at source, reviewer effort and evidence completeness. Automation percentage by itself can reward the wrong behaviour. A finance team creates value when it resolves the right exceptions faster and leaves a stronger control trail.

Questions finance leaders ask

Can AI approve a payment or journal?

It can prepare evidence and a recommendation, but a defined finance owner should retain approval for consequential actions. The rule, approval limit and decision record should remain visible.

When is an AI workflow actually ready?

When the population, rules, exceptions, owners and escalation path are clear enough for someone new to operate the process without relying on undocumented judgment.

Where should a finance team start?

Start with one recurring exception queue where the underlying data and policy are understood. The finance control framework gives you a practical starting card.

Supporting guides

Further reading

I keep the technology discussion tied to established risk and control thinking. These are useful reference points, not substitutes for designing the finance workflow itself.