CRM Connectors for Superintelligent Agents: Fields to Expose First
Which CRM fields an AI agent should read first, which to keep hidden until a workflow proves the need, and how to set the first read-only scope.

The first CRM fields an AI agent should read are the ones that answer three questions: which account or contact this is, where it sits in the sales or service process, and what a person decided or promised most recently. Personal details, billing data, free-text notes and fields owned by other teams should stay out of reach until a named workflow shows it needs them. Start with a short, typed read scope, then widen it one field at a time, each change with a written reason.
Key takeaways
- Expose identifiers, owner, stage, last activity and next step before any other field.
- Keep personal contact data, billing fields and free-text notes hidden until a specific workflow needs them.
- Prefer a structured field, such as a stall reason with a short list of values, over asking the agent to interpret long notes.
- Add each new field through a recorded decision that names the workflow, the purpose and the person who approved it.
- Review the read scope on a fixed schedule, because CRM fields change as teams and processes change.
Why the first read scope decides how useful the agent becomes
A CRM is a record of commitments and relationships. An agent given too little sees a name and a stage and cannot draft anything useful. An agent given too much reads private history, repeats sensitive details in its outputs, and produces summaries that mix stale notes with current status. Nobody trusts those summaries for long. The first scope should be narrow enough to be safe and complete enough to finish one task from start to end.
The connector should also respect the permissions your CRM administrator already maintains. It should not become a second, wider permission system that quietly exposes more than the human users of the CRM can see. Authorization guidance from OWASP describes the principle clearly: access decisions belong where the data is requested, and they should be checked on every request rather than assumed from an earlier setup.
Which CRM fields to expose first
Begin with fields that are stable, owned by one team, and needed by almost every follow-up task. These fields tell the agent what the record is and what should happen next without revealing more than the task requires.
- Record identifiers, such as the CRM account or contact ID and any external reference that other systems use to find the same record
- Account or contact type, and the segment or territory it belongs to
- Owner or assigned seller, so the agent knows who is accountable for the next message
- Pipeline stage, together with the date the stage last changed
- Date of the last logged activity and the channel it came through
- Next step and its due date
- Status flags such as do not contact, opted out or on legal hold
Status flags deserve early attention because they constrain behavior. An agent that can see a do-not-contact flag will not draft outreach for that record, and that single field prevents more damage than most of the others combined. Expose it as a blocking condition that the workflow checks before any action, not as background information the model may or may not weigh.
Which fields should stay hidden until a workflow proves the need
Hidden does not mean deleted. It means that no connector reads the field by default, and that adding it requires a decision. The following fields are the usual candidates for staying hidden at the start.
- Personal contact details beyond the business channel the task needs
- Billing, payment and bank fields, which belong to finance workflows with their own limits
- Contract values, discount approvals and margin figures
- Free-text notes, call transcripts and attachments, which often contain remarks from people who never expected automated reading
- Fields owned by another team, such as support ticket details or legal review status
- Bulk export endpoints and any view that returns the whole table at once
Free text deserves particular caution. It is where sensitive information tends to accumulate, and it is also where instructions can enter the system unnoticed. A note written by an outside party, or pasted from an email, may contain text that tries to steer the agent. The OWASP guidance on prompt injection prevention treats untrusted content as data to be handled carefully, not as instructions to follow. A read scope that excludes free text removes much of that exposure before any filter is needed.
A step-by-step plan for choosing the first fields
The fastest way to decide is to work backward from the task. Use the sequence below with the business owner and the person who will review the agent's output.
- Write down the single workflow the agent will support, its expected output and the person who reviews that output.
- List every field a person would open to complete that workflow, then remove any field that is only a convenience.
- Mark each remaining field as stable, sensitive or owned by another team.
- Grant read access to the stable fields only, then test the agent on real records chosen by the reviewer, not by the agent.
- Compare each draft with the record it used, and log every field the output depended on.
- Add a sensitive field only when a failed draft shows a real gap, and record who approved the change and why.
The names and types of the fields the agent receives matter as much as the selection. A field called status with values that only make sense to the sales team invites errors. Typed values, plain names and clear error messages reduce the chance that the agent invents an interpretation. The design principles in Tool Schemas for AI Agents: Typed Fields and Clear Errors apply directly to CRM connectors, and the guidance from Anthropic on writing tools for agents makes the same point about descriptions and parameter naming.
What an illustrative follow-up agent reads
Illustrative scenario: a sales operations lead wants an agent that drafts follow-up summaries for deals that have stayed in the same stage too long. The first read scope includes the record identifiers, owner, stage, last activity date, next step and the do-not-contact flag. The agent produces a short summary and a suggested message, and the deal owner reviews both before anything is sent.
In the first weeks, the summaries repeatedly failed to explain why deals had stalled. The reasons lived in free-text notes, which the team had kept hidden on purpose. Opening the notes would have exposed private remarks and invited instructions from outside parties into the agent's context. The team chose a different fix. They added one structured stall reason field with a short list of values, which owners fill in at the moment a deal changes. The agent then had a reliable explanation from a field it was allowed to read, and the notes stayed closed. The example shows a general pattern: when the agent lacks a fact, the better question is usually whether a structured field can carry that fact, not whether the agent should read more text.
How to widen access without losing control
Reading, drafting and writing are three different permissions, and they should be granted at different times. An agent that can read a stage should not automatically be able to move a deal to another stage. Write access is a later decision, and it belongs to a separate review. The distinction is explained in Least privilege for AI agents: start with the smallest access, and it matters most in CRMs, where a single careless update can reach many people's records.
Each new field should come with a short change record. The record names the workflow that needs the field, the purpose of the read, the owner who approves it, and the date when the scope will be reviewed again. Removing a field should be just as easy as adding one. If a workflow is retired or a process changes, the field should leave the connector the same week, not wait for an annual cleanup. The NIST AI Risk Management Framework is useful here because it frames these decisions as ongoing risk management rather than a one-time setup.
How Be Human manages CRM connectors in practice
Be Human starts with one role and its workflow, maps the fields that role needs, and sets the read scope, the approval boundaries and the list of hidden fields with the reason for each. The managed operation then covers connector monitoring, recurring evaluations that compare drafts with the records they relied on, permission reviews whenever the CRM schema changes, and incident response if a field is read outside its agreed scope. Field changes are handled as controlled improvements, and each one goes ahead only after the business owner approves it. If you are not sure which CRM field your first agent should read, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Should an AI agent read the whole CRM on the first day? No. A narrow read scope makes it easier to test the agent, to explain its outputs and to notice when something goes wrong. Widen the scope only after a real workflow shows which additional field closes a specific gap.
Can the agent write back to CRM fields once it can read them? Not by default. Reading and writing are separate permissions with separate risks, and writes should go through an approval step that a named person controls until the workflow has a track record.
What if a sales team needs a hidden field for one task? Treat it as a request, not a configuration change. Record the purpose, the owner, the expiry date for the permission and the specific output the field will support, then grant access for that workflow only.
How often should the read scope be reviewed? Review it whenever the CRM schema changes, whenever a workflow changes and on a fixed calendar cycle agreed with the business owner. The review should check every exposed field against its current purpose and remove any field that no longer has one.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


