The customer insists the AI must be fully autonomous, with no humans reviewing outputs. You think that's risky. How do you handle it?

30-second answerSay your answer out loud first, then reveal.

Understand why they want autonomy (cost, speed, scale) and respect the goal. Make the risk concrete with their data: error rates from evals, the cost of a wrong action, and regulatory exposure. Then propose graduated autonomy: start with human review, measure agreement between AI and humans, and expand autonomy to the case types where measured accuracy and risk justify it, with guardrails, monitoring and rollback. Document the decision and the risks if they still choose otherwise.

Conversation approach

Acknowledge: "Your goal is to handle 10x volume without growing the team. That's achievable."
Quantify the risk: "On the eval set, the AI is wrong on 6% of refund decisions. At 5,000 a day, that's 300 wrong decisions daily. What does one wrong decision cost?"
Propose a path:

Phase	Autonomy	Gate to next phase
1 (weeks 1–4)	AI recommends; humans approve all	AI–human agreement ≥ 97% on case type A
2	Auto-approve type A under ₹2,000; humans review the rest	Weekly audits show < 1% error
3	Expand to type B and higher thresholds	Same criteria

4. Guardrails regardless: limits, anomaly alerts, sampling audits, kill switch, reversibility where possible.

If they still insist: escalate internally (your manager, legal or compliance if needed), get the decision in writing, and implement the strongest guardrails possible. Never quietly build something you believe is unsafe.

Interview signal: customer empathy plus spine. You push back with data and offer a path, not just "no".

PreviousYour POC reaches 70% accuracy, but the customer expected 95%. What do you do?

This is what real progress feels like.

The customer insists the AI must be fully autonomous, with no humans reviewing outputs. You think that's risky. How do you handle it?

Conversation approach

Related