The honest range
In an HR SaaS build we delivered, the AI layer resolved 67% of tickets autonomously and first response went from around 24 hours to about 12 seconds. Customer satisfaction rose from 72% to 94%. Those are real numbers from one product with good documentation and a high proportion of repetitive questions.
They are not universal. A product with sparse documentation and complex, account-specific problems will land lower — 30–40% is a fair expectation there. The variable that predicts the outcome is not the AI, it is whether the answers exist in writing anywhere.
Deflection is the wrong headline metric
Deflection is easy to game: make escalation hard and it goes up while your customers get angrier. Track it, but judge the system on a different set.
- Resolution rate — the customer's problem actually went away, measured by no re-contact within seven days.
- Satisfaction on escalated tickets — the ones that went to a human. This is the number that reveals a bad system.
- Time to human when escalation happens. If the AI stage adds five minutes before a person appears, it made things worse.
- Re-contact rate, which catches confidently wrong answers that deflection counts as wins.
Escalation design decides everything
The single biggest failure in AI support is not wrong answers, it is trapping people. A customer who wants a human and cannot reach one has a worse experience than they would have had with a slow email reply.
Rules that have served our clients well: escalate on the second failed attempt, escalate immediately on any signal of frustration or a complaint, escalate anything about money or cancellation, and always show a visible route to a person.
When escalation happens, hand over the full context. Making the customer repeat everything to a human is the moment goodwill is lost, and it is entirely avoidable.
Answer only from your own material
Support answers must come from your documentation, policies and account data — never from the model's general knowledge. A model that invents a refund policy creates a commitment you may have to honour.
Practically that means retrieval with a strict instruction to say “I don't know, let me get someone” when the material does not cover it. That refusal behaviour needs testing as carefully as the answers.
It will expose your documentation
Every AI support project turns into a documentation project by about week three. The system cannot answer what was never written down, and the gaps become visible immediately and specifically.
This is a genuine benefit disguised as a problem. The list of questions the AI could not answer is the most useful documentation backlog most support teams have ever had.
A rollout that does not risk your customers
- Shadow mode. The AI drafts, agents send. Two to four weeks. You learn the real accuracy without exposing anyone.
- Assisted. Agents accept or edit suggestions. Speed improves, quality stays owned by people.
- Autonomous on a narrow class. Password resets, order status, opening hours. Boring and safe.
- Widen by evidence, one category at a time, watching re-contact rate rather than deflection.