AI sales evals: the internal QA loop

What to eval

• Reply quality — human-scored on 100 samples per prompt change
• Grounding faithfulness — every claim cited
• Guardrail compliance — no violations
• Latency — P95 within budget

Cadence

On every material prompt change. On every model version bump. On every playbook update.

Who owns it

RevOps engineering, not sales. Sales sets the bar; ops enforces it.