Skip to main content

Questions to Ask an AI Vendor About Human Approval and Access Controls

The short answer

Ask what the AI can read, what it can change and which actions need approval. Test denied actions, changed requests and revoked access before production use.

Illustration for the article: Questions to Ask an AI Vendor About Human Approval and Access Controls
The short answerAsk what the AI can read, what it can change and which actions need approval. Test denied actions, changed requests and revoked access before production use.

The short answer

Ask what the AI can read, what it can change and which actions need approval. Test denied actions, changed requests and revoked access before production use.

What should we ask an AI vendor about human approval and access controls before connecting company data? We should get clear answers about what the AI can access, which actions need human approval, how those controls are enforced, and how we can test and revoke access.

Key Takeaways

Question to askWhat a useful answer should establish
What task will the AI perform, and what data does it need?A narrow use case, named systems and fields, and a clear account of what the AI produces.
Can permissions be limited by system, record, field, user, or action?Access follows least privilege, with read-only access where the task does not require changes.
Can the AI prepare an action without executing it?Consequential actions wait behind a technically enforced approval gate.
Who receives company data, and how long is it retained?The vendor identifies model providers and subprocessors, retention periods, training use, and deletion steps.
How can we prove the controls work?Acceptance tests cover allowed and blocked actions, approvals, audit logs, and revocation.
How do we pause or remove access?A named owner can stop tool access, revoke credentials, and manage queued actions and emergency access.

Before connecting systems, we can use AI Integration Services as an example of a staged approach that investigates existing systems and tests integrations before go-live. The questions below help us assess whether any proposed setup fits our data boundaries and approval requirements.

Define the task and map the data it needs

Start with the work the AI is supposed to do, not with a list of systems the vendor can connect. Ask the vendor to name the task, the systems and data fields it needs, and the output it will produce; a narrow scope makes unnecessary access easier to identify.

  • Which steps will the AI perform? Ask for a workflow description from input to output, including where a person enters or reviews information.
  • Which systems and fields does each step use? Ask for a system-by-system data map that distinguishes information the AI reads, transforms, stores, or sends elsewhere.
  • Can access be limited by record, field, user, or business unit? Test the proposed limits against the task instead of assuming that a full-system connection is necessary.
  • What can the AI do with its output? Draw a firm line between recommendations or drafts and actions that change records, contact people outside the company, or trigger transactions.

Ask the vendor to show the data flow from the initial connection through to each output and destination. A connection can expose data beyond the first application if the workflow passes information to other systems or providers.

We should also separate stages that carry different levels of risk. An AI that drafts a customer reply needs a different level of authority from one that sends the reply, updates a customer record, or submits a payment instruction.

An IT specialist securing a server cabinet

Illustrative image; not a customer case study.

Scope permissions and give the connection its own identity

Ask for a permission matrix for every connected system. It should show the data types the AI can access, the actions it can take, and whether each connection is read-only or has write access.

Use least privilege as the test: grant only the permissions needed for the defined task. If the AI only retrieves information, read-only access is safer than write access; if it must update a record, limit that ability to the specific records and fields involved.

  • Which OAuth scopes or equivalent permissions are granted? Ask whether the vendor can narrow them and why each scope is needed.
  • Does the connection require administrator rights? Prefer specific scopes over broad administrator access when the workflow does not need system-wide control.
  • Does the AI use a separate non-human identity or service account? A dedicated identity makes actions easier to attribute and access easier to revoke than shared employee credentials.
  • How are credentials stored, rotated, and revoked? Ask what happens to active sessions, tokens, and queued work after revocation, since removing a credential alone may not end existing access.

For an implementation partner, our AI Integration Services work investigates existing systems and builds integrations in stages, with testing before go-live. The service page describes encrypted connections, access controls, and audit logging, features that are relevant when we assess how an integration will handle permissions and trace activity.

Ask the vendor to demonstrate the permission boundaries in the connected systems themselves, not only in an AI configuration screen. A restriction that exists only in a prompt or written policy does not replace a permission enforced by the application or connection.

Make approval gates specific and technically enforceable

A human-in-the-loop policy is useful only when it identifies the actions that require review and prevents those actions from running before approval. Ask the vendor to list the exact actions that need an approval gate, such as sending a message, changing a record, or submitting a transaction.

For each action, ask whether the AI can prepare it without having permission to execute it. This separation lets a person review a proposed action while limiting the AI’s authority to carry it out.

  • What will the approver see? The review screen should show the exact proposed action, the affected records, the destination, and the data to be sent.
  • What technically blocks execution? Ask whether the system prevents the action until an authorized person approves it, rather than relying only on a vendor policy or assurance.
  • What happens if the action changes after approval? Edits to the message, destination, records, or transaction details should not inherit approval for a different action.
  • How do retries and queued steps behave? Confirm that a retry cannot silently execute a changed action under an earlier approval.
  • How are failures and exceptions handled? In a multi-step workflow, an incomplete or uncertain step should be flagged for resolution, not treated as approved by default.

For workflows connecting several applications, AI Integration Services Cape Town connects ERP, CRM, POS, accounting, and legacy applications with automated data validation, exception handling, real-time synchronization, and audit trails. Those workflow features make it especially important to ask how approval applies at each step and how the system handles a failed or changed step.

Approval screens should make decisions reviewable, not merely offer an approve button. If a request bundles several actions together, ask whether the approver can inspect and approve them separately.

Trace where company data goes and how long it remains

Ask the vendor to identify the model provider, subprocessors, and any other parties that receive company data. For each party, ask what information it receives and what role it plays in the workflow, so we can see the full data flow.

Ask separately about prompts, uploaded content, outputs, and logs. These data types may have different retention periods, and we need to know whether any of them are used to train or improve models.

  • How long does each data type remain? Request a retention schedule for inputs, outputs, logs, and any copies created during processing.
  • How does deletion work across the chain? Ask how the vendor and relevant subprocessors delete data, including how deletion applies to backups.
  • Who among vendor staff can access company data? Ask what controls limit operational access and whether access is approved and recorded.

Ask for a clear description of how the vendor handles a deletion request when data has moved between providers or systems. A process that addresses only the primary application leaves other copies and processing locations unclear.

Test permissions and approval paths before launch

Use acceptance tests in our own environment to check what the AI can and cannot do. A vendor explanation is not a substitute for testing the actual connection, permission settings, and workflow.

  1. Test an allowed action. Confirm the AI can perform the intended read or write within the approved scope.
  2. Test a denied action. Try an out-of-scope record, field, or action and verify that the system blocks it.
  3. Test the approval gate. Prepare an action and attempt to execute it without approval; then approve it and verify that only the displayed action proceeds.
  4. Change the proposed action. Edit the content, destination, or affected records after approval and verify that the changed action requires fresh review.
  5. Test untrusted content. In a controlled environment, include connected content with malicious instructions, including prompt injection attempts, and check that it cannot grant tool permissions or trigger an unapproved action.
  6. Inspect the audit trail. Confirm logs record the AI identity, action, affected records, time, result, approver, and approval details.

These tests help expose excessive agency, where an AI system can do more than its assigned task requires. They also test whether connected content can influence the workflow without overriding the permissions and approvals we set.

Ask for staged implementation and testing before go-live, then agree how the team will monitor activity after launch. Reassess permissions and approval rules when the task, connected systems, or workflow changes.

Plan revocation, exceptions, and oversight

Agree on an access revocation procedure before the AI has production access. It should cover credential removal as well as active sessions, tokens, and queued actions, and we should test it during acceptance.

  • Is there a kill switch or equivalent pause control? Name who can use it and how they can quickly stop the AI’s tool access during a suspected incident.
  • How does emergency access work? Require a named approver, a limited duration, minimum necessary permissions, and a recorded reason.
  • Who requests, configures, and approves high-risk actions? Separate those roles where practical, so one error or compromised account cannot both authorize and execute an action.
  • How often will oversight take place? Set recurring reviews of permissions, service identities, approval rules, and audit logs.

Exceptions should have a defined purpose and end point. If emergency access becomes a routine way to get work done, it has become a bypass rather than a controlled response.

For each review, assign an owner who can confirm that the connection still serves its original task and that its permissions remain proportionate. This keeps oversight tied to actual changes in the workflow instead of treating launch approval as permanent approval.

Frequently Asked Questions

Should human approvals expire if an action is not carried out promptly?

Yes, set an expiry that reflects the action’s sensitivity and how quickly its context can change. When approval expires, the workflow should require a fresh review rather than carrying the old decision forward.

How should we handle a pilot that needs to use sensitive data?

Use a representative test dataset with sensitive fields removed or substituted wherever it can still validate the workflow. If real sensitive records are essential, limit the pilot audience and duration, and define how the pilot data will be cleared at its end.

What should happen if the AI vendor changes its model provider or subprocessors?

Require advance notice that identifies the new provider, the data it will receive, and the planned change date. Set a review point before the change takes effect so we can assess its impact on our data-handling requirements.

Can vendor support staff access production data, and under what conditions?

Ask for support access to be time-limited, approved for a specific case, and attributable to an individual support identity. The vendor should be able to explain how it records the access and closes it when the support task ends.

What should we do when an AI workflow starts producing more actions than its original task requires?

Pause the affected tool permissions and review recent activity against the task definition. Treat a growing set of actions as a scope change that needs renewed approval, not as an automatic extension of the original access.

Who should own ongoing reviews of an AI connection?

Assign an operational owner who understands the workflow and a security or governance contact who can assess its permissions and data handling. Make sure both know how to raise a concern and who can pause the connection.

Conclusion

Before connecting company data, we should ask the AI vendor to define the task, map the data flow, justify every permission, and show how a human approval gate blocks consequential actions until review. We should also test denied actions, approval changes, audit trails, and revocation in our own environment.

That is the practical answer to what should we ask an AI vendor about human approval and access controls before connecting company data? We ask for narrow access, meaningful and enforceable approvals, clear data handling, and an oversight plan that remains active after launch.

A reference for the vendor discussion

OWASP’s Excessive Agency guidance explains why an AI application needs limited functionality, permissions and autonomy. Use it alongside the acceptance tests above when assessing a proposal.

Put the next step into practice

For the related implementation detail, read Can a Business AI Assistant Work with Microsoft 365 and Google Workspace While Respecting User Permissions?. Bring one workflow, the systems it touches and a few representative examples to a discovery conversation.

TagsAI GovernanceAccess ControlsHuman Approval

Keep exploring

The short answer

Ask what the AI can read, what it can change and which actions need approval. Test denied actions, changed requests and revoked access before production use.

What each chapter added

  1. Start with the work the AI is supposed to do, not with a list of systems the vendor can connect.
  2. Ask for a permission matrix for every connected system.
  3. A human-in-the-loop policy is useful only when it identifies the actions that require review and prevents those actions from running before approval.
  4. Ask the vendor to identify the model provider, subprocessors, and any other parties that receive company data.
  5. Use acceptance tests in our own environment to check what the AI can and cannot do.
  6. Agree on an access revocation procedure before the AI has production access.
  7. OWASP’s Excessive Agency guidance explains why an AI application needs limited functionality, permissions and autonomy.
  8. Put the next step into practice

Transform Your Business with AI

Discover how AI can drive growth and efficiency: book a free consultation today.

If you want this applied to your own business, talk to the people who wrote it.Loxly Atkinson, CEO & Co-Founder

Related service: Our AI Services