AI Features Your Customers Use Directly

The tolerance is different

An internal tool that gets something wrong costs an apology in a meeting. A customer-facing one that gets something wrong may create a commitment you have to honour, or a screenshot that circulates.

So we build them to a different standard, and we usually recommend proving the capability internally first.

The five guardrails

1. Strict grounding — answers only from your own material, with refusal rather than inference
2. Visible escalation to a person, always available, never buried
3. Disclosure that it is an AI assistant, at the start
4. Rate limiting and abuse handling, because it will be probed
5. Logging of every interaction, for dispute resolution and improvement

It must be able to say it does not know

A model that invents a refund policy has created a commitment. We test refusal behaviour as carefully as we test answers, because it is the property that keeps customer-facing AI safe.

Escalation design decides satisfaction

The biggest failure in customer-facing AI is not wrong answers, it is trapping people. Escalate on the second failed attempt, on any sign of frustration, on anything about money or cancellation, and always show a route to a person.

When it escalates, hand over the full context. Making the customer repeat everything is where goodwill is lost, and it is entirely avoidable.

What to measure

• Resolution rate — no re-contact within seven days
• Satisfaction on escalated cases, which reveals a bad system faster than anything
• Time to human when escalation happens
• Re-contact rate, which catches confidently wrong answers that deflection counts as wins

Frequently asked questions