AI & AUTOMATION · 3 MIN READ

What Data Should an AI Integration Be Allowed to Access?

Define AI integration access by task, user, and field, with limited tools, permission checks, and a clear account of what leaves each system.

“Connect the assistant to our CRM” leaves an important question unanswered: which records, fields, and actions should it be able to access for this particular task?

Start with the task and work backwards to the minimum information needed. A system that drafts a shipping-status answer may need an order reference and tracking state, but not internal sales notes or every customer's account history.

01

Draw the data journey

List the source systems, the application doing the retrieval, the model provider, and the destination of the result. Identify what is stored at each stage, including execution logs and review screens.

This is an implementation inventory, not a substitute for reviewing contractual or legal obligations. Its practical purpose is to prevent overlooked data movement. A carefully limited model request is not enough if a debug log stores the entire source record elsewhere.

02

Use the user's actual permissions

In a hypothetical internal assistant, a regional manager should receive only the records that the application permits that manager to view. Do not retrieve the full organisation's data and ask the model to hide the rest.

Enforce identity and record access before assembling the context. Treat model output as a proposal that must pass normal application checks before any write. The model should not decide whether its own request is authorized.

03

Design narrow tools

Prefer an operation such as “look up delivery state for this permitted order” over a general tool that accepts arbitrary database queries. Narrow interfaces are easier to explain, test, and log.

OWASP's prompt-injection guidance supports limiting permissions and validating tool actions as part of a layered design. A stricter prompt alone is not a complete boundary against instructions embedded in material the assistant reads.

For an initial release, separate read-only tools from tools that change records. Review those capabilities independently; usefulness in answering questions does not prove readiness to perform updates.

04

Inspect nonessential fields

Build an explicit field list for each model request. Remove internal comments, secrets, and unrelated records. Preserve information needed to answer correctly; excessive removal can also make the assistant misleading. The correct scope comes from the task rather than a generic rule to send everything or nothing.

Use representative test records to inspect the actual request, output, and logs. Check error paths too: a failed connector may expose more information in a diagnostic message than a successful run does.

05

Make access changes visible

Document who owns the integration credentials and who can expand tool permissions. Review access when the assistant gains a new task, changes audience, or connects to another system.

A useful first deliverable is a one-page access map with allowed operations and deliberately excluded information. That gives the business and developer a concrete basis for deciding what the assistant may do before it is connected to live accounts.

06

Discuss the implementation

Explore AI integrations and AI agents versus traditional automation.

KEEP EXPLORING

Sources & further reading

Official documentation for the topics discussed. Check implementation details against the versions and services used by your project.

WRITTEN BY MARKO ARNAUTOVIC

Practical thinking.
Hands-on engineering.

I build AI automation, web platforms, and mobile applications for businesses. These notes share the decisions behind that work.

Meet the engineer behind ArnautCode

PUT THE IDEAS TO WORK

Need help with
the implementation?

Explore the approach and scope for ai integrations.

Explore the service
Chat on WhatsApp