Deflection rate — the percentage of contacts that never reached a human — is the number every AI support vendor puts on the slide, including, historically, us. It is easy to compute, it moves in the right direction when things go well, and it converts cleanly into money saved.
It is also, taken alone, an actively dangerous target, because of what it rewards when it is under pressure.
What a deflection target optimises for
Suppose your only goal is fewer human contacts. The cheapest ways to achieve that, in rough order of effort:
- 1Hide the "talk to a human" option.
- 2Add friction before escalation — a form, a confirmation, a "are you sure?" step.
- 3Have the bot answer everything, confidently, whether or not it knows.
- 4Actually answer questions well.
Only the fourth is what you wanted. The first three are faster, and a team under quarterly pressure will find them. Not through malice — through the ordinary gravity of a metric.
The better target: resolution
Resolution asks a different question. Not "did we avoid a human?" but "did the customer end up with what they needed?" A conversation that is escalated in thirty seconds to the right agent, with full context, and solved in one reply is a success under resolution and a failure under deflection.
That inversion is the whole point. It is also the version that matches what your customers would say if you asked them.
Measuring resolution without a survey
Surveys have a response rate of about four per cent and a strong bias toward the furious. You can approximate resolution from behaviour instead:
- No repeat contact within seven days on the same topic — the strongest single signal.
- The customer completed the action they came to ask about (upgraded, exported the file, connected the integration).
- Explicit positive feedback, when offered, weighted lightly because it is sparse.
- No downstream negative event — no refund request, no cancellation, no one-star review mentioning support.
None of these is perfect. Together they are considerably harder to game than a deflection count, which is the property you are actually buying.
Keep deflection — demote it
Deflection is still a fine efficiency metric. It tells you what the bot costs and what it saves, and finance will reasonably want it. The mistake is letting a cost metric run the product.
A workable hierarchy:
| Tier | Metric | Who watches it |
|---|---|---|
| North star | Resolution rate | Whoever owns the customer experience |
| Guardrail | Repeat contacts, escalation latency | Support leadership, weekly |
| Efficiency | Deflection rate, cost per conversation | Finance, monthly |
A bot that says "I am not sure, let me get someone who is" fifty times a day is doing its job. Do not build an incentive that trains that sentence out of it.
The version customers notice
Here is the thing that does not show up in any dashboard: people can tell. They can tell within two exchanges whether a bot is trying to help them or trying to make them go away. The second kind produces a specific, memorable irritation that attaches itself to your brand rather than to your chat widget.
Optimise for the customer leaving with their problem solved. The cost savings arrive anyway — they just arrive as a consequence rather than as a target.
- Strategy
- Metrics
- Customer experience
