TT.TONY THOMASCHARTERED ACCOUNTANT
← All insights

FINANCE AUTOMATION

A faster close needs a stronger review loop

Use AI to organise the work around the close. Keep financial judgment, journal approval and accountability with people.

By CA Tony Thomas · Chartered Accountant · 5 minute read
Published

FINANCE · JUDGMENT · TECHNOLOGY / 01 OF 04

The opportunity is in the handovers

A month-end close rarely stalls because a finance team cannot calculate a number. It stalls because an invoice has no clear owner, a reconciliation is waiting for evidence, or a late adjustment has not reached the reporting pack. Automating those handovers can be more valuable than asking AI to write a fluent explanation of an unfinished ledger.

The design I would start with is an exception-led close: deterministic checks establish what happened, AI helps organise the supporting material, and an accountable reviewer decides what follows. It is a proposed operating model, not a description of a deployment at my employer or a promise of a particular reduction in closing days.

For a UAE business with several branches, the practical objective could be to give the controller one reliable view of unresolved accruals. Success means each material item has an owner, evidence and a disposition before reporting is signed off.

A worked accrual review

Consider an illustrative distributor closing September. An approved service order is AED 120,000. The system contains an AED 80,000 supplier invoice and evidence that the full service was delivered by month-end. A rules-based check flags a potential AED 40,000 unbilled balance. That calculation is straightforward; the accounting conclusion requires more work.

The reviewer must establish whether the invoice is partial, whether another invoice exists in a different branch, whether the service acceptance is reliable, and whether the amount includes recoverable tax. The service period, contract terms, existing accruals and relevant accounting policy matter. An unmatched purchase order alone is insufficient evidence of an expense.

AI can assemble an evidence summary with references to the order, invoice and acceptance record. It should report: “Potential difference: AED 40,000; completeness and tax treatment require review.” It should not report: “Accrue AED 40,000” as an approved accounting instruction.

If the reviewer confirms the treatment, an authorised preparer creates the journal through the normal process. A separate approver signs it off where the control design requires segregation. The AI service has no posting credentials.

Build an exception queue people can use

Start with one reconciliation or accrual population. Freeze a dated extract, reconcile its totals to the source system, and record entity, period, currency and extraction logic. Reject incomplete loads before any analysis starts. A beautiful exception dashboard built on missing branches is a control failure.

  1. Detect: use documented rules for missing approvals, duplicate references, unusual aging and unmatched balances. Set thresholds by risk and account behaviour; one monetary limit will not suit every account.
  2. Enrich: retrieve only the evidence the reviewer is permitted to see. Ask AI to summarise it into fixed fields, with “not available” allowed.
  3. Route: assign an owner, due date and escalation route. Keep high-risk items visible even if their value is small.
  4. Resolve: choose supported adjustment, no adjustment with rationale, further investigation, or escalation. Preserve rejected suggestions as well as accepted ones.
  5. Reconcile again: tie approved journals to the updated ledger and reporting pack. A closed ticket is not proof that the financial statements were updated.

Keep the full population available. Sample items that were not flagged, particularly around period-end and manual journal activity, to detect what the workflow misses.

Make the evidence trail reconstructable

A useful audit trail allows another competent person to reconstruct the decision. For each exception, retain a unique case ID; source record IDs and extract version; the rule and threshold that triggered it; the model and prompt-template version; the generated summary; and links to the evidence actually reviewed.

Add preparer and reviewer identities, timestamps, the final rationale, journal reference where relevant, and the result of the ledger tie-out. Preserve a record of later changes. Use access-controlled, versioned storage with retention aligned to company policy and applicable obligations.

Do not retain every raw prompt indefinitely just because storage is cheap. Logs may contain confidential data. Store the minimum approved evidence needed to explain the decision, protect it, and define who can export it. Record provenance and integrity checks where appropriate, while recognising that an unchanged file can still contain an incorrect assertion.

Documents are evidence inputs, not system instructions. A supplier PDF containing text that tells an AI assistant to ignore its rules must never change permissions, approve a case or initiate a payment.

Run a pilot that can fail safely

Use a shadow run first: compare the workflow against the existing close without changing posting or sign-off responsibilities. Include known duplicates, incomplete documents, conflicting service dates and legitimate unusual transactions in the evaluation set. Agree the expected reviewer response before testing.

Measure reviewer time per resolved case, unsupported suggestions, missed issues found by sampling, cases reopened, and the percentage of final decisions with complete evidence. Track total effort, including extraction, corrections and oversight. Faster drafting is not a useful saving if it creates more reconciliation work.

Set acceptance criteria locally. For example, require every pilot decision to have an accountable reviewer and traceable evidence, with no unauthorised posting. Any leakage, source reconciliation failure or permissions breach should pause the pilot immediately. Repeated unsupported narratives should trigger redesign rather than cosmetic prompt changes.

Give the team a manual fallback and a named owner who can disable the workflow. Revalidate after a model, data source or accounting-policy change. Expand only after the control owner accepts the evidence from the pilot.

The first practical step

Pick the reconciliation with the most repeated follow-up emails. Write down what a reviewer needs to close one exception, then build that evidence packet before building an AI assistant. The strongest opportunity is a shorter path from an unresolved item to a defensible decision.

Suggested pilot deliverable: one reconciled population, one exception register, one reviewer checklist and one documented stop rule. That is enough to test whether automation improves the close without surrendering control over it.

About CA Tony Thomas

Chartered Accountant based in the UAE, with AICA Level 2 (ICAI) and PGD–GST qualifications and previous experience at Ernst and Young (EY). Writing on financial reporting, practical automation and responsible use of AI.

More about Tony ↗
Perspective & scope

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.

Explore the AI Finance Lab roadmap · Discuss these ideas