• CISO
  • CIO
  • CFO
7 min readBy Rullan Scott TechnologiesLast verified

What Happens When a Software Vendor Turns On a New AI Feature?

THE SHORT ANSWER

When an existing SaaS vendor enables a new AI assistant, your original security review no longer reflects reality. Before leaving the feature active, an assigned owner must verify three things: what data it can access, what actions it can take on that data, and who can turn it off at the tenant level.

On this page

A year ago, your company approved a standard help desk platform. It passed procurement and a security review, and support agents used it to read and answer tickets. Nothing about it was unusual.

Around the six-month mark, the vendor rolled out an AI assistant inside the same subscription. There was no new contract and no new purchase order. The assistant can summarize tickets, read attachments and draft replies. The application still sits on your approved vendor list, unchanged, and nobody has asked what the assistant can now see or do.

Day 1

What you approved

Help Desk SaaS
Support Tickets

Human-only access, covered by the original procurement review.

Day 180

Vendor enables AI

Help Desk SaaS + AI Assistant
  • Access

    Customer Records & Attachments

  • Action

    Auto-Drafting & Data Processing

  • Control

    No Tenant-Level Kill Switch

The same approved application, before and after the vendor enables AI.

The assistant does not invent new permissions. Microsoft documents that Copilot "can only summarize or reference content that the user is authorized to access." That protects you only if access is set correctly. AI features make files that are already shared too broadly far easier to surface, because anyone can now ask for them in plain language.

The Three Checks to Run Before Leaving a Vendor AI Feature Active

An approval reflects what you reviewed at the time. When an update changes what the software can access or do, the risk you approved is no longer the risk you have. Run these three checks as soon as the feature appears, and mark each one as you go.

  1. Check 1: Access

    What data can it access?

    In the admin console, confirm whether the assistant works under each user's own permissions across tickets, attachments and connected drives, or queries tenant-wide records through a separate service account. Pass only if you can show the setting, not the vendor's description of it.

  2. Check 2: Action

    What can it do with that data?

    Confirm whether the feature only summarizes for a person to read, or can draft and send customer replies, update records or trigger workflows in other systems. Then check the vendor's data terms for whether prompts and outputs are fed into vendor or model provider training.

  3. Check 3: Control

    Who can turn it off?

    Confirm that your own IT admin has a tenant-level switch to disable the feature or limit it to a pilot group, and whether it ships on or off by default. Test the switch in a pilot, and check whether turning it off removes any features you pay for.

The three checks at a glance
CheckWhat It CatchesPrimary OwnerEvidence to Capture
What data can it access?An assistant that searches attachments, workspaces or connected apps beyond the record a user has open, or that runs under a service accountIT, as application administratorAdmin console settings showing the data scope and any service account the feature uses
What can it do with that data?Replies sent without review, records changed or workflows triggered on its own, and prompts used for model trainingBusiness owner of the applicationThe vendor's written description of the feature's actions and data terms, plus which actions need human approval and where they are logged
Who can turn it off?Features that ship on by default with no tenant-level switch, or that can only be disabled through a vendor support ticketIT, with Security reviewing the exposureA tested disable or pilot-group restriction, with the named person who can use it and the date it was tested

Why This Isn't Hypothetical: Recent Vendor Rollouts

  • Vendor Announcement (Atlassian)

    Rovo AI added to existing Jira and Confluence subscriptions

    Atlassian announced that Rovo AI search, chat and agents were "included in your existing subscription" for Jira, Confluence and Jira Service Management cloud customers. The rollout started with Premium and Enterprise plans, with Standard customers to follow.

  • Admin Notice (Microsoft 365 Message Center, MC1152323)

    Microsoft 365 Copilot app installed automatically on Windows devices

    Microsoft resumed automatic installation of the Microsoft 365 Copilot app on eligible Windows devices with the Microsoft 365 desktop apps, rolling out from mid-June to mid-July 2026. Microsoft said the installation was "enabled by default" unless administrators opted out. The change did not apply in the European Economic Area.

Neither change needed a new purchase order, so neither would trigger a procurement review on its own.

How should you decide what to do with each change?

Match the response to what the update changes
If the update changes...ExampleWhat to do
What data the application can reachThe assistant can now read attachments across all ticketsReview before enabling widely. Limit it to a pilot group until the three checks are done
What the application can do on its ownThe assistant can now close tickets or email customersRequire human approval for those actions, and confirm the actions are logged
Where your data goesPrompts are now processed by a third-party model providerCheck the vendor's data terms and subprocessors before enabling
Only the interfaceA new summary button over data users already seeRecord the change and move on

How do you make this part of your normal process?

Start with five applications that already handle customer, employee or financial data. Give each one an internal owner. That owner watches release notes and vendor notices, then checks in the admin console what is actually enabled and who can use it.

Use one simple rule: if an update changes what data the application can reach, where that data goes, or what actions it can take, review the change when you find it. Then decide whether to approve it, limit it to a pilot, or turn it off where the controls allow until you have answers. When you need the vendor's answers on record, send them the request below.

Vendor AI feature: change review request

Send to the vendor's account or security contact when a new AI feature appears.

Subject: Written confirmation requested: [AI feature name] in [Application]

Hello [Vendor contact],

We see that [AI feature name] is now available in our [Application] subscription. Before we leave it active, please confirm the following in writing:

1. Data access: What data can the feature read? Does it act under each user's own permissions or under a separate service account? Can it reach attachments, other workspaces or connected apps?

2. Actions: Can the feature only summarize and draft for a person to review, or can it update records, send messages to customers or trigger workflows in other systems?

3. Data handling: Where are prompts, source data and responses processed and stored, and for how long? Are any third-party model providers or new subprocessors involved? Are prompts or outputs used to train any model?

4. Controls: Is the feature on or off by default? Can our administrators disable it at the tenant level or limit it to a pilot group, and does disabling it affect any features we pay for?

5. Logging: What record do we get of the actions the feature takes?

Please also tell us where you publish admin-facing change notices so our application owner can subscribe.

Thank you,
[Your name]
[Application owner, Company]

External Standards Referenced

NIST's Generative AI Profile treats vendor AI as something to review on acquisition and keep monitoring after deployment.

  • NIST AI 600-1, GV-6.1-009Due diligence for generative AI vendors
    "Update and integrate due diligence processes for GAI acquisition and procurement vendor assessments to include intellectual property, data privacy, security, and other risks."

    Published July 2024. Read the source (opens in a new tab)

  • NIST AI 600-1, GV-6.2-004Continuous monitoring of third-party generative AI
    "Establish policies and procedures for continuous monitoring of third-party GAI systems in deployment."

    Published July 2024. Read the source (opens in a new tab)

Questions we hear

Doesn't our vendor contract already cover this?

Contracts usually describe how data is processed at the time you sign. A new AI feature can add a model provider, a new data flow or new actions that the contract never contemplated. Check the data processing terms for subprocessors and AI providers, and ask the vendor to confirm in writing what changed.

Should we block all vendor AI features by default?

Blocking everything can push staff toward unapproved tools that you can't see at all. A better default is to limit any AI feature that touches sensitive data to a small pilot group, run the three checks, then decide. Features that only reorganize data users already see can usually be approved quickly.

Who should own this, IT or security?

The business owner of each application should be accountable for the decision. IT confirms what the feature can access and whether limits can be enforced. Security reviews the exposure. That split matches how most AI governance frameworks assign ownership, and it stops reviews from stalling between teams.

How often should someone check release notes?

For the five applications with the most sensitive data, at least monthly, and whenever the vendor sends an admin notice. Most enterprise vendors publish admin-facing change notices ahead of time. Subscribing the application owner to those notices is the cheapest early warning you can set up.

Sources

Take the eight-question Third-Party Risk assessment