The prerequisite nobody mentions

Connecting an assistant to your CRM makes it answer from your data. If that data is duplicated, stale or inconsistent, you have built a faster way to be confidently wrong. Data quality is the first project, not the second.

Giving an AI assistant access to CRM data is the highest-value connection for most B2B businesses, and the one with the most ways to get it wrong.

What becomes possible

QuestionCurrentlyConnected
"What is happening with this account?"Read several screensAnswered with a summary
"Draft a follow-up for this opportunity"Generic, or paste context manuallyGrounded in the real history
"Which deals have gone quiet?"A report someone buildsAsked in plain language
"Summarise this call into the CRM"Typed manuallyDrafted, written on approval
"What did we agree last quarter?"Search notes and emailRetrieved and summarised

Design the tools narrowly

The instinct is to give the assistant query access to the CRM. The design that works — and that survives a security review — enumerates operations:

  • Get account summary by identifier, returning specified fields.
  • Get recent activity for an account, date-bounded and limited.
  • Find opportunities matching defined criteria, capped in number.
  • Get contact details, filtered to what the workflow needs.
  • Draft a note — created unsent, pending human approval.
  • Update a named field, only after explicit confirmation.

Permissions must come from the requesting user, not a service account. CRM permission models encode real commercial boundaries — territories, ownership, restricted fields. A connector on a broad service account lets anyone who can ask a question reach data their role deliberately excludes them from.

Data quality determines the outcome

Data problemHow it manifests
Duplicate accountsSummaries missing half the history
Stale opportunity stagesConfident advice on dead deals
Inconsistent field usageAnswers that contradict each other
Notes in email, not CRMThe context is simply absent
Unowned recordsPermission logic produces odd results

An assistant reading your CRM is a very effective data quality audit. Teams routinely discover during this work that their pipeline reports have been wrong for a year.

Use the pilot as a diagnostic. Ask the assistant to summarise twenty accounts your team knows well, and have them mark what is wrong. The errors are almost never the assistant's — they are your data, made visible. That list is worth having regardless of what you do next.

Untrusted content in CRM records

CRM notes and descriptions contain text from customers, forms and imports. Treat all of it as data:

  • Retrieved content is summarised, never followed as instruction.
  • No consequential action happens without a human approving it explicitly.
  • Tools are scoped so a followed instruction cannot exceed the user's own permissions.
  • Bulk retrieval patterns are logged and reviewed.

A rollout that works

  1. Read-only, one team, a few accounts they know intimately.
  2. Have them mark every inaccuracy for two weeks.
  3. Fix the data problems that surfaced — this is real value, independent of the connector.
  4. Widen the user group once answers are reliable.
  5. Add drafting — notes and follow-ups created unsent.
  6. Add writes individually, each behind explicit approval and logged.

Measuring it

  • Time to prepare for a customer conversation — the clearest win.
  • CRM note completeness, which usually improves when drafting is easy.
  • Draft acceptance rate for generated summaries.
  • Data quality issues found and fixed, which is a genuine return.
  • Whether people still ask colleagues the questions the assistant can answer.

Connecting an assistant to your CRM? Tell us which CRM and what your team asks about most. See our connector service, CRM and helpdesk connectors, and custom MCP servers.

Frequently asked questions

Try it. Where your requirements match what it does, it is far cheaper than building. Custom work becomes worthwhile when you need specific operations, tighter permission mapping, or an audit trail the packaged option does not provide.
Only if you build it that way, and the first version should not. Read-only lets you learn what people actually ask for before granting anything the ability to change records. Writes come later, gated by explicit approval.
Then the connection amplifies the problem — answers will be confidently based on wrong data. Fix the data quality issues that matter first. This is frequently the more valuable project and it is discovered during connector work rather than before it.