AI customer support: deflection without the frustration
Support automation fails when it optimises for containment. What to measure instead, and how to design the handoff.
Everyone has been trapped in a support bot that would not let them reach a person. That experience is the product of one metric — containment rate — chosen because it is easy to measure and directly opposed to what customers want.
Containment vs resolution
What it counts
- Containment rate
- Conversations that didn't reach a human
- Resolution rate
- Issues actually solved
Counts a giving-up customer as
- Containment rate
- A success
- Resolution rate
- A failure
Optimising for it produces
- Containment rate
- A bot that hides the exit
- Resolution rate
- A bot people choose to use
Follow-up metric
- Containment rate
- None
- Resolution rate
- Repeat contact within 7 days
| Containment rate | Resolution rate | |
|---|---|---|
| What it counts | Conversations that didn't reach a human | Issues actually solved |
| Counts a giving-up customer as | A success | A failure |
| Optimising for it produces | A bot that hides the exit | A bot people choose to use |
| Follow-up metric | None | Repeat contact within 7 days |
Measure resolution, not containment
A conversation the bot 'contained' because the customer gave up is a failure recorded as a success. Track whether the issue was actually resolved, whether they came back within a week, and satisfaction on the automated path specifically.
Make the human path obvious
Counter-intuitively, an easy route to a person increases the share of people who stay with the bot. Hiding it produces adversarial customers who type 'agent' into every field. Offer it plainly, and earn the stay.
Transfer context, always
The single most infuriating moment in automated support is being asked to repeat everything to the person you were escalated to. The handoff should carry the transcript, what the bot understood, what it tried and what the account state is.
Ground it in your actual policies
A support assistant answering from general knowledge will invent a returns policy. Ground every answer in your published policy documents and account data, cite what it used, and refuse when the answer is not there.
The handoff decides the experience
Carry across
- The full transcript
- What the bot understood the issue to be
- What it already tried
- Account state and recent activity
Never
- Ask them to repeat everything
- Drop them into a queue with no context
- Hide the route to a human
- Count a customer giving up as a success
Start with the boring tickets
Where is my order, how do I reset this, what is my balance — high-volume, low-ambiguity, self-contained. Get those genuinely right before touching the emotive cases, where a clumsy automated response does real damage.