← All guides

Practical guide

Assessing AI in customer service

AI in customer service is useful when the task, knowledge and hand-off are defined more narrowly than the buzzword. The starting question is not how intelligent a system sounds, but which enquiries it may handle reliably.

In brief

A sound first case answers recurring questions from approved sources and visibly hands unclear or consequential enquiries to a person.

How to recognise a suitable case

Start with a limited subject area. The required information should be findable, current and owned by a subject-matter expert.

  • Similar questions occur regularly.
  • The answer can be derived from named sources.
  • It is clear when no answer should be given.
  • There is a visible hand-off for open cases.

Where a person should take over

Uncertain statements, complaints, contract questions or decisions with significant consequences should not silently become an automated normal path. The exact boundary depends on purpose, data and risk.

Escalation is not a system failure. It is a planned outcome when sources are missing, contradictory or outside the approved scope.

How to assess a pilot

Prepare real but sanitised test cases in advance: typical questions, unclear wording, outdated sources and deliberately disallowed cases. Record the expected behaviour for each test.

Assess answers, source grounding, non-answers and hand-offs separately. Only then can you decide whether knowledge, boundaries or operation need adjustment.

Make privacy concrete early

Instead of making a blanket compliance claim, owners should assess purpose, data categories, legal basis, storage paths, participating providers and deletion rules for the specific use. Privacy measures depend on the actual workflow.

Sources and further reading

Assess a chatbot plan anonymously

Privacy status

No optional analytics or marketing services are enabled in this local redesign. External booking content remains closed without a configured appointment type and an explicit click.