How we taught an AI colleague to handle customer questions (and where we drew the line)

Description and use

The team implemented a four‑level escalation matrix to define when AI should answer, draft, route, or stay silent. This enabled a two‑week deployment and measurable gains in time and error rates.

Problem

Teams often ask if AI can answer a question instead of whether it should answer it autonomously, leading to ineffective AI deployments in customer contact.

Outcome

The structured escalation matrix shifted value from AI answers to clear rules about what AI should not handle, improving trust and enabling quick, measurable deployments.

Workflow

• 1. Identify a repeating weekly process (e.g., answering ten common questions, preparing quotes).
• 2. Define a four‑level escalation matrix: 1) AI answers directly, 2) AI drafts and human approves, 3) AI routes to a person, 4) AI stays out.
• 3. Measure baseline metrics: time spent per week, response time, error rate.
• 4. Deploy AI for the selected process within two weeks.
• 5. Monitor metrics and adjust escalation rules as needed.
• 6. Iterate to more complex processes once trust is established.

Evidence

Most teams that try AI in customer contact start with the wrong question. They ask "can the AI answer this?" instead of "should the AI answer this on its own?"

Source

The escalation matrix is the real product. The value is not in the answers the AI gives. It is in the rules for what it does not handle.

Source

Pick a process that costs time every single week. Answering the same ten questions, completing incoming requests, preparing quotes, following up on leads. If a process does not repeat weekly, it is not a good starting point.

Source

The first process should be live in two weeks. Not six months, not after a data project. Two weeks is enough to cover one workflow honestly.

Source

Pick three numbers before you start: time spent per week, response time, and error rate. If you cannot measure it, you cannot defend the project later.

Source