AI is useful in ops when it drafts, ranks, or summarizes — and dangerous when it silently invents facts in the CRM. Use it where speed matters and errors are cheap to catch.
1. Good fits (start here)
- Summarize long email threads or call notes into a short deal brief
- Draft first-pass outreach or follow-ups for a human to edit
- Score or rank leads with an explainable model — not a black box destiny
- Classify tickets / intents to route work faster
2. Bad fits (avoid or heavily gate)
- Writing definitive CRM fields (ARR, legal entity, closed-won) without review
- Auto-sending customer-facing messages with no human approval on day one
- Replacing process design — AI won’t fix a broken handoff
- Training on data you don’t have rights to use
3. Guardrails that belong in the design
- AI writes to “suggestion” properties or notes; humans promote to system-of-record fields
- Log prompt version + model for anything that influenced a decision
- PII minimization — send only the fields the model needs
- Fallback: if AI fails, workflow continues without it
4. Measure like an ops team
- Baseline time for the task before AI
- Track edit rate — how often humans change the draft
- Track error types that reach the customer or CRM
- Kill or redesign if edit rate stays near 100%
5. Rollout pattern
- Pilot with one team and one workflow
- Document “AI may / may not” in the playbook
- Train on judgment — when to trust, when to rewrite
- Expand only after two weeks of calm metrics
AI earns its place when it removes busywork and leaves CRM truth under human control. That’s the bar.