The Resolution Receipt: Auditing What UK AI Service Channels Do After They Claim Success

AI agents now sit in front of web chat, voice menus and messaging threads across UK service operations. In 2026, the board question is no longer whether to deploy them, but whether they are reducing the total effort of service. The most common answer is a containment rate: the share of conversations closed without a human. That number is useful, but it is not a resolution measure. It describes an exit, not an outcome.

A bot can contain a conversation and still fail the customer. It can close a chat after giving a plausible answer that does not apply to the customer's circumstances. It can suppress a handoff because the queue is long. It can mark an intent as resolved while the customer immediately opens a second channel to ask the same question. None of these behaviours will show up in a containment dashboard unless the organisation deliberately audits the period after closure.

The metric that stops too early

Containment is an exit metric. It records that the automated channel closed without a human. It does not record whether the issue was resolved, whether the customer understood the answer, or whether the same person returned two days later. In 2026, service leaders are under pressure to show that AI is reducing cost to serve. Containment is convenient because it is easy to report. It is also easy to optimise in ways that push work downstream.

The containment illusion is not that AI fails. It is that a closed conversation looks like a closed case. The audit needs to follow the query, not the session.

Four tests after the resolution receipt

A resolution receipt is the record created when an AI-fronted channel marks a conversation complete. At minimum it should capture the timestamp, the intent, the bot's confidence, the source used, the next step promised and whether a human handoff was offered. The audit then samples these receipts and tests what happened next.

1. Context retention

When a case escalates after closure, does the human receive the transcript, the bot's reasoning, the customer's stated constraint and any promise made? If the human starts from scratch, the bot has not contained work. It has created a second discovery phase. Context retention is the difference between a handoff and a restart.

2. Escalation debt

Count human contacts linked to an AI-closed case within 14 days. Distinguish repeat contact, new intent and correction. Escalation debt is the labour deferred by the bot, often arriving in another channel with lower tolerance. A voice queue that absorbs chat closures is still part of the same service estate, even if the reporting line does not show it.

3. Loopback rate

Measure customers who return to any channel with the same reference or intent within 72 hours and 14 days. Segment by intent and channel. A high loopback rate after containment is a signal that the bot closed a conversation, not a case. It also exposes the intents where automation is least ready for unsupervised closure.

4. Handoff trail

Audit whether the handoff is documented, timed and owned. Include silent handoffs, where the customer is transferred without notice, and broken handoffs, where the bot promises a callback that no queue owns. The trail should survive the channel switch. If it does not, the organisation has created a new form of shadow work.

Run the audit as a channel exercise, not a model review

The audit should be led by service design and operations, with data and compliance support. It should not be delegated to the AI vendor to mark its own homework. Sample across web chat, voice and messaging, and across the intents that matter most to customers. Link the bot's closure record to the CRM, the contact centre platform and the ticketing system. Then read a sample of transcripts alongside the linked contacts.

A useful discipline is the shadow test. Take a small set of AI-closed cases and ask an experienced adviser what they would have done at the point of closure. If the adviser would have asked two more questions, the bot's confidence score is not the only thing that needs review. The test is not designed to embarrass the model. It is designed to show where the channel design has made resolution too easy to declare.

What good looks like in 2026

Good AI-fronted channels do not maximise containment. They maximise confirmed resolution at the lowest total effort. That means accepting a human handoff when context is thin, offering a callback with a reference, and recording the customer's preferred channel. It also means monitoring the post-closure period as a first-class part of channel performance, not as an exception report.

For UK organisations with regulated service obligations, the distinction matters. A containment figure is not evidence of a good outcome. A resolution receipt that shows what was promised, what was handed over and what happened next is closer to the standard that boards and regulators are likely to expect.

Governance: make the receipt auditable

Add the receipt fields to the channel specification. Assign an owner for post-closure follow-up. Report loopback and escalation debt alongside containment at the same cadence. Set thresholds by intent, not by estate average. And review the handoff trail when a customer makes a complaint, because complaints are often the clearest evidence that a closed conversation was not a closed case.

Conclusion

The containment illusion is a reporting problem with operational consequences. UK teams that audit after the bot says 'resolved' will see the difference between automation that absorbs work and automation that forwards it. The resolution receipt is a small instrument, but it changes the question from 'how much did we contain?' to 'what happened next?' For channel strategists, that is the more useful question to take into the 2026 plan.