AI integration permissions: what to check before connecting accounts
Review an AI integration's identity, access to company data, allowed actions, denial tests and disconnection procedure.
Article language: English

In this article
“Connect account” is a short button label that can conceal a long list of capabilities: reading mail, retrieving files, creating messages or changing data. Before enabling the integration, find out what it can do, whose identity it uses and how you will take away its access. The scope should follow the task. An app connected to one folder should not need a metaphorical key to every cupboard in the company.
To establish that, examine the consent request, the configuration and the integration's behaviour in a test. A visible Connect button and a successful sign-in do not yet establish that the scope is right. The steps below describe an assessment by the company; they do not assume that every service uses the same permission model.
Describe the task before reviewing the consent screen
Write down why you need the integration, which materials it should read and what result it should produce. If the task is to search approved instructions, full access to the organisation's entire documentation needs separate justification. Also identify who will receive the result, what approval is needed and when the task ends. Those boundaries give you something concrete against which to assess the proposed permissions.
Walk through the process with the data owner and the person responsible for the system. Ask whether a narrower directory, a prepared collection or a read-only connection would achieve the goal. Record the path of information in an AI data-flow register. The permission assessment uses that map but also examines whether a particular action can be performed on a resource. The Polish guide to mapping an AI data flow (in Polish) describes how to map the path.
Check the application's identity and organisation
Before approving access, compare the app name, publisher, identifier and environment with the deployment documentation. Make sure you are signing in to the intended organisation and account. Similar service names, test and production environments, or several accounts belonging to one person can lead to a connection with a different scope from the one intended.
Record the business and technical owners, support route and configuration location in an integration record. Do not put secrets, passwords or full access tokens there. You need identifiers that help you find the connection, not credentials that would let someone take it over. If the supplier cannot identify the application covered by its instructions, resolve the discrepancy before granting permissions.
Distinguish the user's actions from the application's own access
Microsoft explains delegated and application access. The former involves acting on behalf of a signed-in user; the latter lets an application act as its own identity. The effect depends on the particular permission and resource configuration. Other services may use different names, so consult their documentation.
In practice, ask whether the task still works after an employee signs out, changes role or leaves the company. Establish who gave consent and how far it extends. Do not assume that removing an employee's licence automatically ends every application connection. That is a separate point to check in the environment and in the departure procedure.
Reading, writing and sending need separate decisions
Break broad access down into actions: viewing metadata, reading content, creating a draft, changing existing material, deleting and acting outside the organisation. For each permission, record which step needs it. A permission that might be convenient for a future idea is not necessarily needed for the current trial. You can assess an expansion when there is a defined reason for it.
If the system is meant to prepare a reply, see whether it can remain a draft. If it must move a file, agree on the destination and what happens when a name conflicts. These cases have different consequences and controls. The responsible AI-agent workflow discusses boundaries for autonomy, approval and publication in more detail. Consent to read should not be treated as blanket consent to act further.
Examine access inherited through groups and sharing
Check what the account can see because of group membership, sharing links and roles granted earlier. An integration may expose an existing overbroad-access problem even when it does not change a file's permissions. The assessment therefore needs to cover resources the user can reach indirectly. A list of top-level folders may not be enough.
In a practice example, a document-search assistant operates with an account belonging to a broad project group. The owner expects it to read one collection of instructions, but the account can also see archived material from other projects. First clarify the scope and choose the configuration. A text instruction saying “use only the instructions” is not a technical restriction on access to the other resources.
Test both allowed and prohibited actions
Use approved test data and several resources with explicitly different access settings. Test reading a resource inside the permitted scope, denial for one outside it and behaviour after a permission is revoked. If the product can write, test in a controlled location. Agree in advance how to reverse the change and recognise any additional effects.
Record the time, account, configuration and observation. The result should show what happened, not just say “integration works”. An error must be understandable to the user. If the tool continues to answer from an older copy after losing the source, record that behaviour for further assessment. Whether that answer remains fit for use is a separate decision for the process owner.

Check logs and visibility of actions
Find out which events the source service records and which the AI product records. Accepting the result may require knowing who started the task, which resource was used and what action occurred. Check whether that information is available on the plan in use and which roles can read it. Do not assume conversation history covers every action performed by the integration.
At the same time, limit what goes into the log itself. Full document contents in logs create another place where data is processed, with its own access and retention needs. Record what is needed for diagnosis and acceptance. If information needed to reconstruct an important action is missing, document the limitation and agree on an additional control before expanding the scope.
Verify disconnection instead of trusting a button
Prepare a procedure for removing access both in the AI tool and at the data source. Microsoft documents how to review and revoke application consent. For the particular deployment, establish what the administrator does and what needs a separate step. Removing a visible connection card may have a different effect from revoking consent in the source system.
After the change, test access again against a prepared resource. Take active tasks, previously retrieved copies and the connection's refresh mechanism into account. Do not give a universal deadline after which everything is certain to stop working: the outcome depends on the service and its configuration. If access remains possible, identify its source before declaring the integration disconnected.
Separate ending access from deleting data
A disconnected integration may leave a history, search index, working files and exports. These need a separate plan. Record what must be retained, what must be deleted and how completion will be confirmed. Preventing retrieval of a new document does not answer what happens to material already processed. If prepared answers remain, also check the AI cache after an access change, so an old result is not still available to the wrong person. The Polish article on the freshness of cached AI answers (in Polish) develops this separate check.
If you suspect improper access, use the data-incident procedure so the right person takes responsibility for securing information and limiting the impact. Do not delete evidence on your own simply because it looks like a copy of an unnecessary conversation. The decision should account for the entire incident. See the Polish guide to retention and deletion of AI data (in Polish) for the wider plan. The Polish article on responding to an AI data incident (in Polish) can help prepare the company's own procedure.
Keep the integration record current
Record the consent scope, resources, owners, test date, disconnection method and next review date. A change in functionality, account owner or source directory may require another assessment. A new need to write should also go through a decision, even when the existing read access was approved. Record why the expansion is needed and who accepts the change. When ownership changes, connect the record to the handover of automation and removal of a departing employee's access.
Before purchase, link the record to the AI supplier assessment. In day-to-day work, the AI use policy points back to it. This keeps technical consent, the permitted task and staff instructions connected. A mismatch is a reason to review, especially when one team maintains the integration and another uses it. The Polish guide to a practical company AI policy (in Polish) explains how to prepare rules for a company.
How do you assess an integration made of several services?
Map each stage of the connection separately. A tool visible to an employee may send requests to an intermediary server, which then retrieves a document from the source system. The review owner needs to know where each component is authenticated and who maintains its configuration. One consent at the start does not explain every later capability. A simple list in the record can help: component, identity, resource, action and where the result is stored. If a server exposes tools through MCP, a communication protocol for applications, separately check the identity and descriptions of those tools and changes to them.
At every boundary, check whether the restriction carries through. If the user is entitled to one folder, establish whether the intermediary uses the same scope or its own broad account. Do not assume correct behaviour from an architecture slide alone. You need configuration evidence or a test that answers the question. Record unresolved scope before making real materials available.
What if the narrowest access is still too broad?
Some products do not allow consent to be tailored to the intended scope. The process owner should then consider another solution, a different way to perform the task or a limited trial in a separate environment. If the company knowingly accepts a wider capability, it needs a specific justification, additional controls and a person to make that decision. “The supplier requires it” describes a product limitation; it is not by itself a justification for accepting the risk.
Also check whether a prepared data collection creates a new maintenance problem. A copy of the instructions may need regular refreshes, removal of withdrawn versions and access revocation. Assign an owner to those activities when choosing the narrower solution. The permission assessment is then connected to running the process, rather than ending with a correct setting on launch day.
How should the permission review be accepted?
Acceptance should include a list of capabilities, evidence that access works within scope, evidence of denial outside scope and a tested disconnection method. Record any missing test as a limitation, together with the decision on whether the task can go live. The person taking over maintenance must be able to find the configuration and data owner without contacting the implementation author.
If your company has connections that nobody can describe, start with one process and its actual access. The Aeneia Audit offer helps define the scope of such a review.
Also check what happens when access is denied. The task should enter a managed exception queue instead of retrying with ever broader permissions or reporting a false success. Agree who may grant the missing permission and on what basis.







