Debt recovery automation needs a credible evidence trail

In a hypothetical case, a collections platform is about to escalate a £12,000 balance. The invoice is 47 days overdue and the reminders have gone unanswered. But an account manager has agreed a service credit by email, while the ledger still carries the original amount. The system has correctly identified lateness. It hasn't established what’s owed.

That gap deserves board attention before another automated collections workflow is approved. Faster correspondence can reduce administrative effort, but it can also move an unresolved commercial disagreement towards legal escalation before anyone has checked the underlying claim. The cost appears elsewhere: settlement concessions, management time or a damaged customer relationship.

Beauhurst’s current UK fintech startup list records total funding above £20m for several businesses classified as seed stage. Those figures establish that capital has been raised; they say nothing about collection performance. For businesses assessing financial technology, the distinction matters. A supplier’s funding story cannot substitute for evidence that its proposed workflow improves the buyer’s economics.

As receivables software moves from recording activity towards recommending or initiating it, purchasing decisions become decisions about delegated authority. AI can classify correspondence and draft a plausible response. Connecting those capabilities to account restrictions or legal referrals changes the exposure substantially, even when each individual software component appears to work.

The strongest business case for debt recovery automation therefore starts with an evidence gate: what must be established before the system can take a consequential action? That question is more demanding than asking whether the platform integrates with the accounting package. It’s also much closer to the board’s responsibility.

What the ledger cannot establish

An aged debtor report describes an accounting position, not a fully evidenced claim. An invoice might identify the wrong contracting entity. A payment milestone might depend on acceptance that hasn’t been recorded. A purchase order discrepancy might be an administrative obstacle rather than a defence to payment. Treating all three as equivalent exceptions wastes effort; ignoring the distinctions produces decisions that are difficult to defend.

Evidence should attach to the obligation itself. That means connecting the invoice to the agreed terms and relevant performance records, with correspondence that varies the amount or payment date preserved alongside them. Where terms have changed, the system needs to distinguish the original agreement from the amendment. A newly uploaded document shouldn’t silently replace the record on which an earlier decision was based.

Payment reconciliation matters just as much. Unallocated receipts can make a paying customer look delinquent, while a credit note posted after a referral can leave an external collector pursuing the wrong amount. Interfaces need controls for duplicate messages and delayed updates. A successful data transfer only proves that information moved; it doesn’t prove that the receiving system now holds the correct balance.

These are different problems from predicting whether an account will pay. A propensity model may identify which customers respond to reminders, but its score doesn’t establish contractual entitlement. Boards should keep those functions separate. Prediction can prioritise work within an authorised population. It shouldn’t decide which balances qualify for legal escalation simply because a model assigns them a low probability of voluntary payment.

For AI systems reading customer emails, apparent confidence is especially misleading. A message referring to a defect could describe a genuine dispute or an already resolved issue. Classification should route the message for appropriate treatment, with its source visible. Customer text is evidence to assess, not an instruction that can change collection policy or override an approval limit.

The necessary control isn’t human approval of every reminder. That would consume much of the benefit. Routine correspondence can proceed where the balance is reconciled and no relevant exception is open. Consequential actions need stronger conditions: a documented claim and an authorised decision. The workflow should record who approved an exception, on what evidence, and which version of the policy applied at the time.

A genuinely disputed claim belongs in commercial resolution or legal review, rather than an automatic insolvency escalation sequence. Insolvency procedures aren’t a substitute for determining a disputed debt. Software that treats increasing pressure as the default answer to silence needs a different decision model.

Make authority part of the investment case

Consider an illustrative monthly business case. Assume automation genuinely brings forward £500,000 of receipts by ten days, without changing eventual recoveries. At an assumed marginal annual funding cost of 10%, the financing benefit is approximately £1,370, calculated using a 365-day year. The £500,000 isn’t additional revenue, and its full amount cannot sensibly be counted as the economic benefit.

Suppose the same workflow releases 60 staff hours a month, valued at an assumed loaded cost of £35 an hour. That adds £2,100 of capacity value. It becomes a cash saving only if the business actually avoids expenditure, such as overtime or external support. If employees use those hours to resolve difficult disputes, measure the resulting benefit rather than counting both their released time and the full value of the same work.

Together, those assumptions produce about £3,470 of monthly benefit before technology costs and additional oversight. A single avoidable £4,000 settlement concession would exceed that amount. That’s a sensitivity test, not a forecast of failure frequency. It demonstrates why exception handling belongs inside the investment model, rather than in a compliance appendix written after procurement has effectively finished.

The evidence for incremental recovery needs equal care. Customers who would have paid anyway often dominate collections results. A supplier can report substantial cash collected while adding little economic value. Where volumes permit, compare similar account cohorts assigned to different treatments, with assignment made before outcomes are known. Randomisation at customer level helps avoid contradictory treatments across invoices owed by the same business.

Measure results over a period long enough to distinguish acceleration from additional recovery. Segment by dispute status and customer concentration where sample sizes allow. Report against the original assigned population, including accounts subsequently removed from automated treatment. Otherwise, difficult cases can disappear from the denominator and make the system look better precisely when it’s shifting more work back to employees.

Smaller portfolios won’t support clean experiments, and concentrated customer books make statistical confidence harder still. A staged rollout with documented case reviews may be more useful than an elaborate model built on thin data. Management should state that limitation openly. A narrow pilot can demonstrate operational competence without proving that the same economics will hold across the entire ledger.

There are second-order effects to watch. Credit controllers may stop recording context if they believe the platform owns the next action. Sales teams may grant informal extensions outside the workflow to protect relationships. Customers may learn that mentioning a dispute suspends contact. Each response changes the information reaching the system, so its apparent performance can deteriorate even without a software change.

Decision rights must therefore follow the case across departments and suppliers. Who can pause activity, and who can restart it? Who may agree a repayment arrangement or settlement concession? An external recovery provider needs explicit authority limits and a reliable route back to the creditor when new evidence arrives. Outsourcing contact doesn’t remove the creditor’s responsibility for the instructions it gives.

Legal routing also needs jurisdictional discipline. For company claims in England and Wales, the applicable pre-action requirements should inform referral decisions; Scotland and Northern Ireland require their own routes. Ordinary trade receivables shouldn’t inherit a generic workflow designed for another product or jurisdiction. The gate should establish that the proposed action fits the claim, rather than treating a template letter as evidence of compliance.

Tax and accounting decisions need separate authority. Writing off a receivable in the accounts doesn’t, by itself, extinguish the legal claim. A settlement or release can have a different effect. HMRC’s conditions for VAT bad debt relief require separate assessment; a ledger write-off alone doesn’t establish entitlement. Systems should preserve those distinctions rather than collapsing every adjustment into a single closed-account status.

The supplier’s own controls belong in the acceptance test. Contact names and individual email addresses can remain personal data in a B2B process. Access permissions and retention rules need to fit the information held. Test whether staff can export the underlying evidence and decision history, including after termination. A downloadable balance file is inadequate if the reasoning behind previous actions stays inside the supplier’s platform.

Before wider deployment, ask the team to reconstruct one difficult case without relying on the vendor’s dashboard. They should be able to establish the balance at referral and explain why escalation was authorised. They should also show what happened when contradictory evidence arrived. If that requires searching personal inboxes, the workflow hasn’t solved the control problem.

Return to the £12,000 balance. The decisive capability is the one that finds the agreed service credit before the next threat is issued. A platform that sends the letter faster, while leaving that evidence behind, has accelerated the wrong decision.