Service is where agents look most attractive on a spreadsheet and carry the most risk in practice, because every mistake happens in front of the person whose opinion of you is being formed.
Deflection is the wrong target
A decade of chatbots optimised for deflection — reducing tickets reaching humans. Customers learned to defeat them, and satisfaction fell while the deflection metric improved. Agents make a different outcome possible: retrieving an order, checking eligibility, processing a change and logging it. That is resolution, and it is worth considerably more than a deflected question.
Where the risk concentrates
- Regulated advice — anything touching financial, medical or legal guidance needs hard boundaries, not soft prompting.
- Commitments — an agent that can promise a refund or a date is making a contractual statement on your behalf.
- Emotional context — distress and complaint escalation require recognition thresholds set conservatively.
- Edge cases — the long tail is where confident wrong answers live, and where customers remember the outcome.
Why conventional deployments disappoint
Most service AI is bought as a cost reduction, so it is scoped to maximise containment. The escalation path becomes an afterthought, thresholds are set to keep containment high, and the customers who most needed a person are the ones who cannot reach one. Cost falls, retention falls further, and the second-order loss is invisible on the dashboard that justified the project.
Designing the escalation path before the capability is the difference. What may the agent never decide alone, what phrases trigger an immediate handover, and what does the human receive when it happens — those are the questions that determine whether this works.
What this means for your business
Scope for resolution rather than containment, set escalation thresholds conservatively, and measure satisfaction alongside cost. Our comparison of agents and chatbots covers the architectural difference this rests on.