Which AI tools already handle sensitive information across your organisation, and what can they do with it?
A colleague might paste a customer email into a public chatbot to draft a reply. A supplier might switch on an AI feature that summarises case notes in software you already use. An internal team might pilot an agent that can update live records. Each route needs a decision about what data can go in, where the output can go and who can intervene.
In Deloitte’s 2026 survey of 25,000 UK workers, 31% of those using generative AI at work said they did so without their employer’s knowledge. For an IT director, the practical question is which of those uses touch information the organisation needs to protect. Deloitte UK’s GenAI Workforce Survey
Existing permissions are part of the answer when staff use Microsoft 365 Copilot. We covered those in Article 03 on Copilot and existing permissions.
Find where AI is being used
Start with a workflow that handles sensitive information, such as summarising a customer case for a handover. Ask the people doing the work to show you which tool they use, what they give it and where they save the result. Use a dummy record for the walkthrough. Then check what the available reporting and application settings can confirm.
Public AI tools. Check managed device or network reporting for visits to AI services. A visit gives you a lead to follow up. Establish whether someone pasted text, uploaded a file or used the service for a task that involved no customer information. This is one place to look for shadow AI, where staff use AI outside the organisation’s approval process.
AI features in business software. Ask the application owner to check which features are enabled, who can use them and which records they can process. Ask procurement to check the terms that apply to the feature. Browsing reports may show visits to a case management system. Check the supplier’s usage records or speak to the team to establish whether anyone used the built-in summary feature.
Agents and custom AI. Review the relevant admin inventories and ask delivery teams about pilots and live deployments. Include agents that run inside another system and applications that call a model through an API. Record which systems each agent can reach and where it has permission to change a record or send a message.
For organisations using Microsoft security tools, Defender for Cloud Apps can identify generative AI services in the traffic it monitors. Microsoft Purview can detect certain interactions, such as sensitive information pasted into supported AI sites, where the relevant policies and coverage are in place. Record the devices, browsers and period covered. Include any gaps, such as unmanaged devices or supplier features with limited usage reporting. See Microsoft’s guidance on discovering AI app use.

If discovery involves monitoring staff, agree what information you need to collect, who can see it and how you’ll explain the monitoring to them. See the ICO’s guidance on monitoring workers.
What to do when the evidence is incomplete
Suppose the network log shows that members of a customer operations team have visited a public chatbot. The team manager can ask whether staff were drafting general correspondence or working with live case notes. Security can check whether the available browser or endpoint reporting covers the tool, device and time window. Together they can establish whether the evidence supports a routine policy discussion or a possible data incident that needs assessment.
Record the finding at the level the evidence supports. For example, “The network log shows visits from managed devices. We still need to establish what, if anything, staff entered into the tool.” Record who will follow up and which evidence they can check. If a supplier can show that its summary feature is enabled but gives you limited usage reporting, ask the application owner and the team to establish how it is used in practice.
Share the outcome with the person who can act on it. An application owner can restrict an enabled feature. The team manager can give staff an approved way to complete the task. Security or information governance can assess a credible concern about live data sharing through the incident process.
Decide what staff can use with live data
Keep one record per use. Include the task, tool and account type, the data it handles, where the output goes, the business owner and the approval decision. Record which findings you’ve verified and which still need checking. Give any temporary exception a review date and a named owner.
For a team handling customer cases, a workable rule could allow public tools for generic templates and require a tool approved for customer information when staff work on live cases. Before approving a supplier’s summary feature, check who processes the case notes, what’s retained, and whether the supplier can use the notes for other purposes, including model training. If there’s credible evidence that live case notes were shared through an unapproved service, send the finding through your incident process for assessment.
The business owner can explain why the task matters and what data it needs. Security can check the technical controls, and information governance can assess the handling of customer records. Name who can approve the use and publish the rule where staff do the work. Give staff a route to ask for an approved tool when they find a recurring job they want AI to handle.
What one approval record could look like
A customer operations team wants to use the AI feature in its case management system to draft handover summaries. Here’s what its approval record could look like while the request is being assessed.
Scope. Draft a handover summary for one type of customer case. The caseworker checks the draft against the notes before saving it to the same record.
Current decision. Test with dummy case records only. The customer operations manager owns the request. The authorised approver will consider live use once the checks are complete.
Evidence to collect. Confirm which services receive the case notes, how long they retain them and whether they can reuse them. Have the application owner test that only the intended team can access the feature and read the saved summary.
When to review again. Reopen the decision if the team wants to include attachments or use the feature for another type of case, or if the supplier changes how it handles the notes.
With that record, a caseworker knows which records they can use today. The approver can see what still needs checking before customer data goes into the feature.
Test the controls before allowing live use
Suppose the review clears the supplier’s summary feature for one type of case. The application owner should test the limits with two accounts. One belongs to a caseworker in the approved team. The other belongs to someone outside it. Check which accounts can start a summary, which records they can select, and who can read the saved summary. If the tool can’t enforce the agreed limits, the approver needs to decide whether a narrower configuration or a different tool will support the workflow.
Give caseworkers a rule they can use during a handover. For example, “Use the built-in summary for the approved case type. Check dates, customer commitments and next actions against the case notes before saving it.” Put that instruction in the handover procedure and show team leads how to raise an error. For the first review, sample saved summaries against the source notes and check whether the agreed access settings are still in place.
Put someone in charge of each agent
Give every agent that can change a business record a named owner and an administrator who can suspend it. Record which identity the agent uses, what it can read, and which actions need human approval. An agent could suggest a category for a caseworker to apply. Before giving it permission to update the case record itself, check that someone can trace the change, suspend the agent and correct an affected record.

Reopen the approval before giving the agent permission to update case records itself.
Microsoft Agent 365 can help with the agents it covers. The agent registry in the Microsoft 365 admin centre can flag agents without owners and show entries outside Agent 365 management. Compare it with the Power Platform inventory and with Azure or supplier records where teams run agents there. Confirm with the business owner what the agent has been approved to do. See Microsoft’s agent registry guidance and Power Platform inventory guidance.
Review changes that affect the original approval
Put new or changed uses, ownerless agents and expiring exceptions on the regular review agenda. Reopen an approval when a supplier changes how it handles data, when a team wants to use a feature on more sensitive records, or when an agent gains permission to take a new action. Moving from suggesting a category to updating a case record triggers that review.
Record the decision from each review, who will act and when someone will check that the action is complete. If a team drops a pilot, give the application owner responsibility for removing its access. Ask the pilot owner to confirm what information needs to be kept and who can see it before closing the approval record.
Pick one workflow where an AI tool handles sensitive information or can change a live record. Can you show what data it receives, who approved the use and who has the authority to pause it?
Find out which AI uses need attention first
If staff are already using AI with customer or employee records, or an agent is about to gain access to a live system, tell us which team or workflow you want to look at first.
Servent’s AI Governance and Data Readiness Review brings together your available reporting and evidence from the people who own the work. You get a prioritised list of findings and a plan that shows what to investigate or change first, with indicative effort.
Gateway to AI is an eight-part Servent series on what organisations need to develop, deploy and operate enterprise AI.
Previous in the series · Copilot governance starts with the permissions you already have
More from the blog:



