Before connecting an AI assistant to your CRM, decide which records it may read, what it may draft, when a person approves changes, and how to stop it safely.

An AI assistant can have limited access to your CRM when it has a defined job, a named owner and permissions that match the job. Start with a narrow read-only test or a copy of example records. Give it the ability to change records or contact customers only after you have tested the consequences, agreed an approval point and proved you can switch it off.

The question is easy to underestimate. A CRM connection sounds like one permission, but it can expose contact details, private notes, deal values, support history and the automations attached to each field. If the assistant changes a status, it might start an email sequence or alter a report. Your team will see the result in the CRM; the customer may see it in their inbox.

The ICO opened a call for evidence on agentic AI in October 2026. It asks about data security, transparency, accountability, automated decisions, fairness and lawful processing. It is a consultation that will inform future guidance, not a new exemption or a finished rulebook. The NCSC’s interim advice on agentic AI makes an operational point: the more autonomy an agent has, the more you need to constrain, observe and stop it. Those are good questions to answer before entering an API key.

What job needs CRM access?

Pick one task. “Help the sales team” is too broad to tell you what access is necessary. “Summarise the last three service interactions for a named account before a call” is testable. So is “draft a follow-up from a meeting note, for a person to check”. Neither requires the assistant to edit the customer’s master record or send an email.

I would write the task on one page before opening a connector:

DecisionExample for a meeting preparation assistant
TriggerA team member requests a summary for one named account.
Records readThat account’s recent activities and agreed contact details.
Records excludedOther accounts, payment details, HR notes and unrestricted exports.
OutputA draft briefing in a review queue.
Action rightsNo CRM writes, email sends or automatic follow-ups.
Owner and stop routeSales operations owner can disable the connection and revoke its credential.

That example is a starting boundary, not a claim that every CRM offers field-level permissions. Some connectors expose only broad access. If the real permission grant is wider than the task, consider a controlled intermediary that selects fields, a separate test environment, or a manual export of synthetic examples. Do not describe a broad token as restricted because the prompt tells the assistant to behave.

Which rights should it have first?

Separate read, draft and write. They are different business decisions even if the software bundles them together.

Read access means the assistant may fetch specific records to answer a defined question. You still need to decide which people can ask, which records it can retrieve, how much it can retrieve at once and whether the results leave your systems. A tool that can search every contact and export every note has a wider job than a meeting preparation assistant needs.

Drafting creates a proposed note, task or message for review. Keep the proposed output separate from verified CRM facts. The reviewer should see the source record, what the assistant inferred and where the text would go. A plausible sentence is not proof that a customer agreed to a price or that a colleague promised a date.

Writing changes the operational record. A contact update, deal-stage change or new task can trigger another system. Map those downstream effects before enabling each write action. For example, if moving a deal to “qualified” sends a sequence, the approval is for both the field change and the resulting communication. A generic “approve AI actions” button hides that distinction.

For a first live pilot, I would permit a short list of reads and draft outputs. A person remains responsible for publishing a note, moving a deal or sending a message. Later, a low-impact write could be considered one operation at a time, with an explicit reviewer, a rollback method and a test of the CRM automation it might trigger. There is no prize for granting every permission on day one.

What customer information may it process?

Look at the data flow, not merely the CRM login. Identify what is copied into the assistant, sent to a model provider, stored in prompts or logs, and passed to another tool. A support note may include health, financial or other sensitive information even if the field name looks ordinary. A narrow query can still pull an unexpectedly detailed free-text note.

The ICO’s AI and data protection guidance covers the existing principles and risk-assessment resources. Its controller and processor contract guidance explains the need for appropriate terms when a processor handles personal data on your behalf. The roles depend on what each party actually does; a supplier’s marketing label does not settle them. That contract page also says some guidance is under review following the Data (Use and Access) Act, so check the current position for your arrangement.

Ask the supplier to show the terms for the plan and connector you will use. Which parties receive the data? Can the provider use it for another purpose? Where are copies retained, for how long, and who can delete them? How will you handle a person’s access or correction request if their details have entered prompts and logs? What happens when the contract ends? The answers belong in the purchase decision and the data-protection review, before live records are connected.

Do not assume that removing a name makes a note anonymous. A company, job title and unusual transaction can identify someone together. For a test, fabricated records are simpler. For a production workflow, minimise fields and check whether a data protection impact assessment is required for the specific processing. Ask your data protection adviser where the answer is unclear.

How do you test the connection before customers are affected?

Start in a test CRM or with fabricated records. Give the assistant a small set of ordinary and awkward tasks, then watch the records and connected automations. The aim is to discover the real boundary of the connection, including what happens when instructions or data are misleading.

I would test these cases:

  1. The wrong account has a similar name. Does the assistant ask for an identifier, or confidently combine two customers?
  2. A note contains an instruction to ignore the task. Does it treat the note as customer data rather than a command to export contacts or change its permissions?
  3. A record contains a sensitive free-text detail. Can the retrieval step exclude it, and is it absent from the output and logs?
  4. The CRM is unavailable or returns an incomplete record. Does the assistant stop with a clear exception rather than invent a complete briefing?
  5. The same request is submitted twice. Does it avoid duplicate notes, tasks and customer messages?
  6. A proposed change would start an automation. Can the reviewer see the later email, assignment or status change before approval?

Keep a simple result sheet: expected behaviour, actual behaviour, records touched, reviewer decision and fix required. A failed case is useful evidence. Correct the permission, integration or process and rerun the case. My Friday afternoon workflow test covers the value of testing interruptions and messy hand-offs rather than only a tidy demonstration.

Prompt instructions are part of the setup, but they cannot carry the whole security boundary. The NCSC recommends controlling the agent’s environment, observing its activity and keeping an emergency stop available. That translates into practical CRM checks: use a dedicated connection where possible, restrict reachable systems and destinations, log retrievals and actions, and test revocation. If the supplier cannot tell you what the assistant accessed, a live pilot is hard to supervise.

What should a person approve?

Approval should be tied to the effect, not to a vague confidence score. A draft internal summary might need ordinary review. A change to contact details, deal ownership, pricing, credit terms or customer communications warrants a more explicit check. Some uses affecting an individual may require additional legal safeguards, especially where a decision has a significant effect. That is a matter for the actual workflow and applicable law, not a rule you can settle by adding an “AI reviewed” tick box.

For each proposed action, show the reviewer the source record, the proposed edit, the destination, any automation it will trigger and the reason for the change. Let them correct or reject it. Name who resolves exceptions and what happens when no one is available. An approval queue that silently times out into execution is no approval at all.

The ICO’s detailed agentic AI call is seeking evidence because these systems can act across tools and purposes. Treat the guidance under development as a reason to keep a clear record of your own choices. It does not mean every SME needs an AI committee. It means somebody must be able to explain what the assistant was permitted to do and why.

What would make you stop the pilot?

Agree stop conditions before measuring productivity. I would pause for access to an excluded record, an unapproved write, an unexplained outbound connection, repeated wrong-account matches or a customer message sent without the agreed check. A prompt-injection attempt that changes the assistant’s behaviour is also a stop and investigate event, even if no damage is obvious yet.

Test the stop route. Disable the integration, revoke its credential, check scheduled jobs and watch whether any queued action still runs. Preserve the relevant logs and investigate which records were read or changed. The NCSC explicitly advises an emergency shutdown capability because stopping a visible process may leave its wider network and tool access intact.

If the pilot behaves well, measure complete tasks, review time, corrections, exceptions and customer impact together. An assistant that drafts quickly but sends reviewers back through every record may save little. The 30-day AI pilot scorecard gives a way to make that decision with actual cases and costs.

The decision I would make this week

Choose one CRM task and write its read, draft and write boundary. Ask the supplier for the exact connector permissions and data terms. Test six awkward cases with synthetic records, including the downstream automations. Put one person in charge of exceptions and the stop switch. Only then decide whether live read access is justified. Keep writes and customer contact as later, separately approved decisions.

If you want help scoping that first workflow, Ajairu’s AI governance service covers the controls and ownership around practical use. Contact me with the task and the CRM you use, and I can help you decide where the boundary should sit.

CRM Governance

Recommended Reads