Enterprise AI security: how to use AI without losing control of your data

AI security is not solved by telling staff to be careful. It comes from deciding which tools are approved, what information they may access, what actions they may take and who is responsible when something goes wrong.

That matters whether your team uses Microsoft 365 Copilot, ChatGPT Business, an AI feature inside existing software or a custom workflow connected to company systems. Each option has different controls, but the same basic questions apply.

The aim is not to block useful work. It is to make useful work repeatable, visible and proportionate to the risk.

Start with the business task

Do not begin with a model or licence. Begin with the job the business wants to improve.

Write down:

• what starts the workflow
• which people and systems are involved
• what information the AI needs
• what the AI may draft, classify, retrieve or recommend
• what a person must check
• what success and failure look like

A tool that drafts an internal meeting summary does not need the same controls as a system that can send customer messages, approve refunds or update financial records. Security should match the consequence of a mistake.

Our guide to AI business process mapping explains how to define the work before choosing the technology.

Know where the data goes

Create a short data map for every approved AI use. It should identify the information entered, the service receiving it, where outputs are stored, who can view them and how long records are retained.

Check the supplier's current business terms and admin controls rather than assuming every version of a product works in the same way. For example, OpenAI states that inputs and outputs from its business products and API platform are not used to train its models by default. Microsoft says Microsoft 365 Copilot operates within the Microsoft 365 service boundary and respects the signed-in user's existing permissions. Those statements are useful, but they do not replace your own configuration review.

The ICO's AI and data protection risk toolkit can help organisations assess risks to people's rights and freedoms. The ICO notes that this guidance is under review following changes in UK data law, so check the live version when making decisions.

Use identity and least privilege

Give every person and integration only the access needed for the approved task.

In practice, that means:

• using named business accounts rather than shared consumer logins
• requiring multi-factor authentication where supported
• removing access when a person changes role or leaves
• limiting integrations to the smallest useful set of permissions
• separating test and production systems
• avoiding long-lived keys in prompts, documents or shared notes
• reviewing connected apps and service accounts regularly

If an AI workflow can take actions, such as sending an email or changing a record, add an approval step until the business has evidence that the action is reliable and appropriately controlled.

Fix oversharing before Microsoft Copilot rollout

Microsoft 365 Copilot works with information the signed-in user is already allowed to access. That is useful, but it also means old SharePoint permissions, broad groups and casually shared folders can become more visible through search and summarisation.

Microsoft's current security guidance for Microsoft 365 Copilot recommends a secure and governed data foundation, including work to identify and reduce oversharing. Before a wider rollout:

• review sensitive SharePoint and OneDrive locations
• check public and organisation-wide links
• confirm group membership and site ownership
• apply labels and retention rules where appropriate
• choose a small pilot group
• record the approved use cases

Blue Canvas provides Microsoft Copilot training in Northern Ireland with practical rollout, governance and workflow support.

Treat external content as untrusted

AI systems can be influenced by instructions hidden in websites, documents, messages or other material. This is commonly described as prompt injection.

The UK's National Cyber Security Centre guidance for secure AI development recommends protecting AI-related assets, documenting data and prompts, managing suppliers and controlling what information systems can access.

For a business workflow, sensible controls include:

• limiting which sources the system may retrieve
• keeping tools and actions on an allow list
• validating important outputs before they are used
• preventing generated text from being executed as code by default
• testing with misleading and hostile inputs
• recording tool calls and approval decisions

No prompt can replace these system controls.

Give staff a simple rulebook

Staff need answers they can use during the working day. A practical rulebook should state:

• which AI tools are approved
• which information must never be entered
• when personal or confidential data needs extra review
• which outputs require human approval
• how customer-facing AI use should be disclosed
• how to report an error or suspected data leak

Use examples from real roles. A salesperson, finance manager and operations lead will each face different decisions. Our AI policy template for business provides a starting structure.

Monitor the workflow, not just the tool

Keep a simple register of AI systems and workflows. For each one, record the owner, supplier, purpose, data, connected systems, reviewers, known limitations and next review date.

Monitor:

• unusual access or activity
• failed or rejected outputs
• human overrides
• complaints and corrections
• changes to the supplier or model
• staff workarounds and unapproved tools

The register should change when the workflow changes. A pilot that starts as internal drafting may need a fresh review before it sends messages or updates customer records.

Prepare an incident response

Decide in advance what happens if confidential data is entered into the wrong tool, a workflow takes an incorrect action or a supplier account is compromised.

The response may include stopping the workflow, revoking access, preserving relevant logs, notifying the system owner, checking legal or regulatory duties and correcting affected records. The exact response depends on the incident, but ownership should never be unclear.

A practical first security review

For each current AI use, ask:

1. Is the tool approved?
2. Does it use a business account with suitable controls?
3. What information can it see?
4. What actions can it take?
5. Who checks the output?
6. Where are prompts, files and outputs stored?
7. What would happen if it failed or exposed information?
8. Who owns the decision to continue, change or stop it?

If those answers are documented and tested, the business has a workable foundation. If they are not, pause the highest-risk uses and fix the basics first.

Blue Canvas helps UK and Irish businesses map AI workflows, configure controls, train teams and deploy practical systems across Microsoft Copilot, ChatGPT and custom software.

Sources

• NCSC guidelines for secure AI system development
• ICO AI and data protection risk toolkit
• Microsoft 365 Copilot security
• OpenAI business data privacy, security and compliance

Book a free 15-minute call