Most AI agent deployments that get quietly switched off were not switched off because the model was weak. They were switched off because one conversation went badly in a way that was avoidable at design time.
Four decisions prevent nearly all of it.
Decision one: disclose
The agent should say it is an agent, in the first message, without being asked.
The argument against is that disclosure reduces engagement. It does, slightly. The argument for is that customers who discover they were misled respond far worse than customers told plainly — and in a growing number of jurisdictions, disclosure is a legal requirement rather than a courtesy.
A named human persona with a stock photo is the version that ages worst.
Decision two: constrain the scope
Ground the agent in your documented answers and constrain it explicitly from improvising on anything with commercial or legal weight.
What to let an agent handle
The right column is where a wrong answer costs real money.
Safe to handle
- Documented product and service questions
- Qualification against defined criteria
- Availability and appointment booking
- Routing to the right team
- Collecting context before a human call
Route to a person
- Pricing beyond published figures
- Contract or commitment questions
- Anything with legal or compliance weight
- Complaints and escalations
- Anything the documentation does not cover
Decision three: build the escape route first
Design the handoff before designing the conversation flow. When the agent cannot answer, it should pass to a person with the full transcript attached, not apologise in a loop or re-ask a question already answered.
Looping is the single behaviour that turns a mildly useful agent into a memorably bad experience, and it happens when the escape route was an afterthought.
Test the handoff before launch by deliberately asking three questions the agent cannot answer. If any of them loops, it is not ready.
Decision four: measure the right thing
Containment rate — the share of conversations handled without a human — is the metric vendors lead with. On its own it rewards an agent that refuses to hand off.
Measure qualified appointments booked, handoff quality, and whether customers who met the agent convert at a rate comparable to those who did not. A busy agent that books nothing is a cost.
Where the value actually is
Not during working hours, if your team already answers quickly. During working hours a good human beats a good agent on almost every measure.
The value is concentrated in evenings, weekends and timezone gaps — the Friday 9pm enquiry that would otherwise wait until Monday, by which point two competitors have replied.
Scoping the deployment to that window rather than to everything is what makes the economics work and keeps the failure surface small.
The first month
Read real transcripts weekly. Not samples the dashboard surfaces — actual conversations, including the ones that ended without a booking.
Every deployment we have seen needed scope adjustments in the first month, and every one of those adjustments came from reading a transcript rather than from a metric.