Skip to content
forgeo
Data quality

Clean your CRM before you connect AI to it

What assistants and agents do with duplicates, empty fields and stale records, and a checklist to run before you plug them in.

By Quentin Tristan3 min read

AI assistants and agents are moving into the CRM: they summarise accounts, draft follow-ups, score leads and answer questions about the pipeline. They all have one thing in common. They work from the records they can read, and they trust them. An AI connected to a messy CRM doesn't fix the mess; it repeats it, faster and with more confidence.

What AI does with bad data

A person who knows the accounts can spot that a record looks wrong. An assistant reading the CRM can't tell, unless the data says so. In practice:

  • Duplicates split the context. Ask for a summary of a customer that exists twice, and you get half the history: the emails on one record, the deals on the other.
  • Empty fields become blind spots. An assistant asked for "all SaaS companies over 200 employees" skips every company whose industry or size is empty, and nothing in its answer says so.
  • Stale records look current. A contact who left the company two years ago still has a title and an email. An agent drafting outreach will happily use both.
  • Missing associations break the links. A deal with no contact, or a company with nobody attached, can't be connected to the conversations around it.
  • Inconsistent values defeat filters. "France", "FR" and "france" are three different countries to a query.

Why it matters more with agents

Assistants that only read can mislead you. Agents that also write can spread the problem: update the wrong copy of a duplicate, assign a lead to an owner who left, enrol a bounced address in a sequence. Before giving any tool write access to your CRM, ask two questions:

  • Can I review what it will change before it changes it?
  • Can I see, and undo, what it changed afterwards?

Forgeo was built around both: it proposes values and merges, nothing is written to HubSpot until an owner or admin approves it, every write is logged, and field updates can be undone batch by batch.

A checklist before you plug AI in

  1. Merge duplicates, so each person and each company has one record and one history. See how to find and merge duplicate contacts in HubSpot.
  2. Give every record an owner. It's what lets an agent route work to the right person.
  3. Fill the fields your questions depend on: industry, size, country, lifecycle stage, job title. Start with what your CRM already knows, then enrich.
  4. Close or update stale deals, so pipeline answers reflect reality. See how to clean a stale deal pipeline.
  5. Normalise dropdown values and replace free text with options where you can.
  6. Repair associations between contacts, companies and deals.
  7. Measure the result, and keep measuring: the HubSpot data audit checklist gives you the checks.

Give AI the right access

How an assistant reaches your CRM matters as much as the data it finds there. The Model Context Protocol, an open standard introduced by Anthropic in November 2024, lets assistants connect to tools and data through a common interface (Anthropic: introducing the Model Context Protocol). Whatever the connection, apply the same rules as for a new team member: the narrowest access that does the job, read before write, and a log of every change.

Forgeo's own MCP integration is coming soon. It will let you ask your assistant about your CRM's data quality, duplicates and records. The cleanup it depends on can start today.

Keep reading