AI agents in customer service: from distribution to servicing

The difference between answering and resolving

The first generation of automated assistants in airlines quickly hit a recognisable ceiling: they could explain conditions, locate information and answer frequent questions, but they could not do anything. When the passenger asked to change a date, claim for a bag or recover an amount, the assistant handed over.

That pattern produces a peculiar outcome: the airline automates the cheap end of servicing and leaves the expensive end untouched. Cost does not fall meaningfully and customer perception worsens, because the passenger has spent time in a conversation that brought them no closer to a solution.

The current discussion about autonomous agents makes sense precisely here, in the stretch that had been left out.

Four requirements to move from answer to resolution

Available actions with a defined scope

An agent can only resolve what the system permits it to execute. Before the conversational model comes the decision about which operations it may perform on its own — reissue, reassign, refund within a limit, open a baggage file — and which require approval. That list is a commercial decision, not a technical one.

Full context on the first attempt

The most frequent cause of poor service experience is not a lack of automation but a loss of context. If the agent does not hold the real state of the booking, the history of the incident and what was promised earlier, it will repeat questions the passenger already answered and forfeit the only advantage it offered.

Escalation with an intact handover

Escalating to a person is part of the design, not its failure. What determines perception is whether the human agent receives everything that happened or starts from zero. A handover that forces the passenger to reassemble the story cancels out the prior work and usually produces the very complaint it was meant to avoid.

Economic limits and traceability

When an agent can refund, compensate or rebook, it needs an explicit limit and a record of what it decided and why. Without those two elements, the organisation must choose between accepting a risk it cannot bound or restricting the agent so heavily that it reverts to being an answer engine.

How to deploy it in phases

1. Inventory contact reasons by cost. Identify the five that consume the most human time.

2. Define the action catalogue. What the agent may execute, within what limit and with what evidence.

3. Resolve context before conversation. Booking state, history and prior incidents.

4. Design the handover. What information the person receives, and in what format.

5. Expand by outcome, not by calendar. Add actions only once the previous ones resolve reliably.

Metrics that separate resolution from conversation

• First-contact resolution without human intervention: cases closed by the agent.

• Repeat contact within 72 hours: an indicator of apparent resolution.

• Context available at handover: escalations where the human agent repeats no questions.

• Cost per case resolved: against the same reason handled by a person.

• Out-of-limit exceptions: decisions that required additional approval.

Mistakes that recreate the earlier ceiling

• Starting with the conversational assistant rather than the action catalogue.

• Measuring conversations handled instead of cases resolved.

• Escalating without passing on the full history.

• Allowing compensation without a limit or a record.

If the project is described internally as an improvement to the chat, it will probably automate only the stretch that was already cheap.

Conclusion

The value of an agent in servicing comes not from its ability to converse but from its ability to act within clear limits, with full context and leaving a trail. Deciding first what it may do — and what it may not — turns automation into a real reduction in cost and waiting, rather than one more step before speaking to a person.

Consumer Services Hub - Strategic ecommerce consultancy for B2C service companies

consumerserviceshub.com