AI & TECHNOLOGY
Use AI to prioritise collections, not to judge customers
An overdue balance may signal a dispute, a missing document or an allocation error. The next useful action depends on the reason, not just the aging bucket.
By CA Tony Thomas · Chartered Accountant · 4 minute read
Prepared
PRACTICAL PERSPECTIVES / FINANCE · AUTOMATION · AI
Make the objective operational
A receivables team with hundreds of open invoices cannot investigate every item with equal urgency. AI can help organise supporting information and suggest where a collector should look next. The objective should be better prioritisation of work, not an automated judgment about a customer's honesty or willingness to pay.
Define a narrow first use: summarise verified collection history, identify missing evidence and suggest a next action for a reviewer. Keep changes to credit limits, legal escalation and payment arrangements within separately authorised processes.
Begin with reconciled balances. A powerful prioritisation model applied to an inaccurate ledger will generate polished follow-up on the wrong amounts.
Three balances, three different actions
Consider a fictional ledger with three overdue items. Customer A owes AED 150,000 and has confirmed a quantity dispute. Customer B owes AED 90,000 but has sent remittance advice that may relate to an unallocated receipt. Customer C owes AED 60,000 and has not responded after a documented payment promise.
A ranking based only on amount would put Customer A first. A practical work queue may instead assign the commercial dispute to sales, the potential receipt allocation to cash application and the broken promise to the collections owner.
The AED 90,000 item should be checked against bank receipts before another demand is sent. The AED 150,000 balance needs the delivery and contract evidence needed to resolve the dispute. Customer C's silence is a reason for follow-up, not proof of distress or misconduct.
These are illustrative routing decisions. Real priorities must reflect materiality, contract terms, access to evidence and the organisation's policies.
Separate observable signals from assumptions
Useful signals include invoice age, verified due date, unresolved dispute status, missing documentation, broken payment promises and confirmed customer contact history. Maintain the source and last update for each field.
Do not infer sensitive personal characteristics from names, locations or communication style. Do not convert vague descriptions such as “difficult customer” into an automated risk label. Replace subjective notes with documented facts and an accountable owner.
Give collectors visibility into why an item was prioritised. If a rule says a promise is overdue, show the promised date and the record supporting it. If evidence is missing, state that rather than assigning a precise probability of non-payment.
Keep prioritisation distinct from credit-loss measurement. An operational score or AI summary does not determine an impairment allowance or replace the assessment required under the applicable reporting framework.
Draft messages only from confirmed facts
An AI assistant may help turn verified information into a concise draft. Require the invoice number, currency, outstanding balance, due date and dispute context to come from approved records. Flag inconsistencies before creating a message.
A draft for an unresolved dispute should ask for the specific evidence needed to resolve it. A draft for a missed payment promise should refer accurately to the agreed date and invite an update. Avoid threats, legal statements or promises the business has not authorised.
Have a reviewer confirm recipient identity, attachments and the amount before sending. Limit access so one customer's records cannot appear in another customer's email. Keep external communications disabled during the initial pilot.
Received emails and attachments are untrusted inputs. They must not instruct the system to change bank details, waive a balance or disclose other customers' information.
Measure resolution instead of message volume
Sending more reminders is not necessarily progress. Track disputed items resolved, missing documents obtained, receipts correctly allocated and payment promises met. Separate an administrative ledger correction from an actual cash receipt.
When assessing cash collected, consider the baseline. Some invoices would have been paid without the intervention. A small controlled pilot or matched comparison can provide more useful evidence than attributing every subsequent receipt to AI.
Measure incorrect balances in drafts, wrong-recipient incidents, unsupported statements and reviewer correction time. Review whether lower-value accounts are repeatedly ignored by the queue. A balanced process should allow owners to surface material context that the ranking missed.
Maintain a clear manual override with a reason. Overrides help both immediate case handling and later evaluation, provided they are not used to conceal a systematic model weakness.
Start with an internal queue
The first deliverable can be an internal daily queue with verified facts, a proposed action, an owner and an evidence link. Review the queue alongside the existing collections process before allowing the system to draft external messages.
Use approved data, limited access and a named owner who can suspend the workflow. Reassess after changes in source systems, customer segments or the model. Keep the finance team responsible for accounting decisions and the authorised commercial team responsible for customer arrangements.
AI is most useful here when it reduces the time spent piecing together a case and leaves more time for the conversation that resolves it. The customer relationship and the final decision remain human responsibilities.
Sources & further reading
- ISCA–NTU: Responsible Artificial Intelligence Framework in Accountancy
Professional context on responsible deployment and human review. The receivables cases and proposed controls above are original illustrative designs.
Linked materials checked on 1 November 2026. They provide background; the worked examples and workflow designs are original.
Personal educational perspective, independent of my employer. All business examples and figures are fictional and illustrative. Proposed workflows do not claim completed deployments or measured productivity gains. This is not entity-specific accounting, tax, legal or investment advice. Apply professional judgment, the applicable reporting framework and your organisation’s approval and data-protection policies.