Before you let an AI agent change your CRM, fix four things: duplicates, stages that mean different things to different people, wrong owners, and missing activity. Agents act on whatever the record says, so bad data that used to cause bad reports now causes bad actions. Then roll out read-only first, and widen access only when the change log shows the agent getting it right.
This matters more after 2026's shift to agent-run CRMs. Salesforce's Headless 360 and HubSpot's refounding both point to a CRM that agents work from, and what a headless CRM is explains the idea.
The eight checks
| Check | Why agents need it | How to test it |
|---|---|---|
| One record per company | Duplicates split history, so an agent sees half the story | Search for your ten biggest customers and count the records |
| One record per person | Duplicate contacts get duplicate emails drafted | Sort contacts by email and look for repeats |
| Stages that mean one thing | An agent moves deals by rule, and vague stages break the rule | Ask three reps what "proposal" means; compare answers |
| Required fields filled | Agents cannot reason about a deal with no value or close date | Filter open deals missing value or close date |
| Owners correct | Work goes to whoever the record names | Filter records owned by people who have left |
| Activity logged | Agents brief and prioritise from history | Pick five live deals and check the last activity is real |
| Stale records closed | Agents chase what looks open | Filter deals with no activity in 90 days |
| Naming consistent | Agents match names literally | Look for the same company spelt three ways |
Use the agent to find problems, not fix them
Connect Claude with view-only access first and ask it to find the problems in the table above. It is good at this, and it changes nothing. "Which open deals have no activity in 90 days?" and "Which companies appear more than once?" are exactly the kind of questions an MCP connector answers well.
Have a person approve fixes in batches, or fix them in bulk with an import.
Roll out in four steps
- Read-only for everyone. Let people ask questions for two weeks.
- One job, one team. Allow changes for one task, such as logging calls, for one team.
- Check the log weekly. Every change should show where it came from. If your CRM's connector cannot show that, stop here.
- Widen slowly. Add jobs and teams only when the log shows the agent doing the right thing.
What your CRM's connector needs to support this
- Read-only and read-write levels you can set per workspace.
- Actions that use each person's own permissions.
- A log of every change with the app it came from.
- A way to undo changes.
- An administrator switch to turn it off immediately.
Empiraa Signal is a CRM and more, and its MCP connector does all five: View only or Full access per workspace, each person's own role, every change logged with its source, 46 of 51 kinds of change undoable for seven days, and an off switch that takes effect immediately.
Related reading: B2B data decay: a working guide to prospect list hygiene, can Claude run your CRM, and what does a salesperson do when the CRM runs itself.



