Let an AI write to your CRM — book jobs, add customers, update line items — and the failure mode isn't dramatic. It's quiet: duplicates. Give an agent only the power to create, and every little hiccup gets papered over with a brand-new record. Here are the three guardrails we put on every write-capable agent.
1. Match on a strong key, never a name
We once watched a loose last-name match book the wrong "Riley" — right last name, wrong human, wrong address, and yes, we texted the wrong person. Now agents identify and de-duplicate on a strong key: phone, email, or ID. Only an exact match auto-acts; anything fuzzy gets read back and confirmed first. A different address than expected is treated as a red flag that it's a different person.
2. Ship the whole toolkit, not just "create"
If an agent can create a record but can't find, update, remove, or reassign one, it "fixes" problems by making duplicates — or by claiming it did something it has no tool to do. So every write agent gets the full set, and it only reports success when a tool actually returns it.
3. Notifications off by default, nothing auto-filled
Booking a job shouldn't secretly text the customer or assign a technician nobody named. Anything that messages a third party is opt-in ("…and text them"), and anything the caller didn't specify stays blank instead of being guessed.
The owner takeaway
The question to ask a vendor isn't "can your AI update our CRM?" — of course it can. It's "how do you keep it from creating duplicates or editing the wrong record?" If the answer is fuzzy, plan on cleaning up a mess.
Want us to pressure-test how an agent would touch your systems? Grab a free workflow audit.