What an AI layer should own
- First response: acknowledging within seconds, even if the answer takes an hour.
- Repeat questions: order status, opening hours, return window, sizing, compatibility.
- Qualification: three questions that let a human skip the intake when they take over.
- Drafting: writing the reply a human reviews and sends, which cuts handling time in half.
What it must never own
- Money decisions: refunds above a threshold, compensation, disputed charges.
- Anything involving safety, medical or legal claims.
- Angry customers: the moment sentiment turns hostile, escalate. A bot that keeps apologising to a furious customer makes it worse.
- Commitments with deadlines: “it will arrive Friday” is a promise, not an answer.
| Signal | Action |
|---|---|
| Second repeat of the same question | Escalate — the bot has already failed once |
| Mentions of refund, lawyer, review, chargeback | Escalate immediately |
| Message length tripling | Escalate — long messages mean frustration |
| Profanity or all caps | Escalate, do not keep apologising |
The handoff is the product
Customers do not hate bots. They hate repeating themselves. So the single most valuable thing you can build is a handoff summary: what the customer said, what was tried, what is still unresolved. A human who opens with “I can see you have already tried X and it did not work” wins the conversation instantly.
Measuring it properly
| Metric | Why it matters |
|---|---|
| Deflection rate | Volume handled without a human — the obvious one |
| Repeat contact rate within 7 days | The one that catches confident wrong answers |
| Handoff quality score | Whether the human had to re-ask anything |
| CSAT on bot-only threads | Compare against human-only, not against last year |
Deflection alone will make you automate too much. Pair it with repeat contact and handoff quality, or the numbers will look great while your churn quietly rises.
FAQ
Q1Should we tell customers they are talking to a bot?
Yes, once, at the start. Concealing it works until someone asks directly, and then you have lost more trust than you saved.
Q2How much can AI actually deflect?
Depends on your question mix, not on the model. Catalogues with stable answers see high deflection; bespoke B2B work sees very little. Measure your own repeat-question share before setting a target.
Q3What about multilingual support?
This is where AI is strongest — consistent tone across languages at a cost no human team can match. Keep the escalation rules identical in every language.
Based on desensitised conversation samples from real private-domain and customer-service work, plus public communication research. Each page is revised as reader feedback comes in. How we work