AI customer service: design the human handoff first

On 14 March 2023, Intercom introduced Fin as a GPT-4-powered customer-service chatbot. The company described goals including natural conversation, answers grounded in business-controlled information and reduced hallucinations. Its same-day discussion explicitly described a mix of bots and humans, with difficult questions passed to support teams. These are primary statements about a historical release and its intended design, not independently verified results for every customer. The useful lesson is the architecture of bounded answers and human escalation, not an invented claim that support can be completely automated. 1 2

Build an escalation policy before opening the channel. Candidate triggers include a clear request for a human, repeated misunderstanding, identity uncertainty, contradictory policy information and actions outside approved authority. These are proposed operating controls, not claims about current Fin defaults. Separate a bot’s inability to answer from a case requiring human judgment even when it can produce fluent text. A refund dispute or a vulnerable customer may need a different route from a routine opening-hours question. The policy should identify who owns the queue and what happens outside staffed hours.

A handoff should carry the customer’s question, relevant account context, attempted steps, source references and unresolved uncertainty. Do not force the person to repeat the entire conversation. At the same time, a generated summary must not silently replace the original record: human advisers should be able to inspect supporting messages. Tell the customer what has been transferred and what response expectation is realistic. Design for a failed transfer as well as a successful one. A button that promises a human while routing to an unmonitored queue is not a functioning escalation path.

The launch emphasized business-controlled help content. That shifts work toward maintaining accurate sources, not away from knowledge management. Assign owners to policies, remove obsolete versions and distinguish information by product, region and customer entitlement. Test questions that retrieve conflicting documents. Company knowledge explains this foundation; AI security addresses access and disclosure risks. A source-grounded answer can still be wrong if the source is outdated or belongs to another customer segment. Retrieval quality and policy ownership therefore belong in the support operating model, not only in the initial technical setup.

A customer who stops replying has not necessarily received a correct resolution. Define successful resolution before choosing dashboard metrics. Track repeat contacts, corrected answers, abandoned transfers, complaint themes and the quality of human handoff. Compare like-for-like case types and avoid attributing changes to AI when staffing or policy also changed. Include reviewer sampling for both English and Swedish interactions. Do not use a vendor headline as a local forecast. A lower apparent contact volume can conceal customers who gave up, switched channels or returned later with a more serious problem.

A bot that explains a policy need not have permission to change an account or issue money. Apply agent permissions to separate reading, recommending and executing. Begin with a bounded pilot, explicit escalation staffing and a reversible rollout. Published on 25 October 2026, this analysis revisits Intercom’s 14 March 2023 release and discussion only. It does not verify current Fin features, pricing, model providers or customer success rates. The operational proposals are evaluation criteria for a buyer, not a claim that a particular configuration has already delivered them.