"We bought the enterprise licenses and turned the tools on. Isn't that the governed version?" It is the question we hear most often from teams running Microsoft 365 Copilot, Google Gemini, or ChatGPT Enterprise, and the honest answer is no. An enterprise license gives you an admin console. Governance is what you do with that console, every time the vendor changes what is behind it.
Here is the direct answer. AI platform governance means knowing which models, features, and data paths are actually enabled in your tenant today, who can change them, and how you will notice when the vendor ships something new. Since late summer 2026, Microsoft, Google, and OpenAI have each changed their admin surfaces in different directions, and at least one new capability shipped with no admin control at all.
Microsoft now uses Anthropic's Claude models inside Copilot, and the control for that sits in a setting most small-business admins have never opened. According to Microsoft's documentation on AI providers operating as subprocessors, Anthropic models can be used across Microsoft Copilot, Researcher, Copilot Studio, Power Platform, and Copilot in the Microsoft 365 apps. The setting is on by default for most commercial cloud customers. In the EU, EFTA, and UK it defaults to no users, and Microsoft added a separate setting there in April 2026 for Copilot in Microsoft 365 apps with Anthropic models.
The default is not the only thing worth knowing. The same page states that Anthropic models are currently excluded from the EU Data Boundary and, where applicable, from in-country processing commitments. For a U.S. company with European clients or contractual data-residency language, that is a compliance-relevant fact, not a model preference. Changing the setting requires the AI Administrator or Global Administrator role, and Microsoft's own Copilot admin guidance recommends using the least-privileged role and limiting Global Administrator to emergency scenarios. In many 50-person companies, nobody holds the narrower AI Administrator role, so the setting simply goes unreviewed.
Nothing in Word, Excel, or the Copilot chat window tells an employee which model produced the answer they just accepted. The only place that decision is visible is the admin center. That is the pattern this article is about: the decision that matters is made once, in a console, and then becomes invisible to everyone using the tool.
It matters because most governance programs assume every capability has a switch an admin can find and turn off. On September 2, 2026, Google extended custom instructions for Gemini from Docs into Ask Gemini in Drive and Chat and the Gemini side panel in Slides, Sheets, and Gmail. The Google Workspace Updates announcement is explicit: "There is no admin control for this feature."
Custom instructions are a reasonable personalization feature, and this is not a security flaw. But it means a user can set standing instructions that shape every Gemini response across five apps, and there is no tenant-wide policy to restrict it and nothing in the console to audit. A team that defines governance as "we reviewed the admin console" will miss this entirely, because there is nothing in the console to review. The control has to live in policy and training instead.
OpenAI has gone the other way, adding admin surface area faster than most small teams can track it. The ChatGPT Enterprise and Edu release notes show a steady run of changes from September onward. On September 11, changes made through the Codex Policies and Configurations screen started appearing in workspace audit logs. On September 17, tenant-wide SCIM was extended to API Platform access, so identity-provider group membership now grants API organization and project access without an invitation. On September 24, OpenAI added External access controls over whether ChatGPT Sites can use members' connected apps and whether applications can access ChatGPT Ads, both off by default during the admin preview.
The defaults do not all point the same way, and that is the real lesson. The same release notes announced ChatGPT for Word on September 17 with a note that, starting October 1, 2026, Word access would be enabled by default, with Microsoft 365 admins still needing to allow the add-in. In mid-October, OpenAI consolidated app permissions onto a redesigned External access page and retired administrator-managed sync connectors in favor of real-time Google Drive, SharePoint, and Teams apps. A setting you reviewed in August may now live on a different page, govern a different connector, or have changed its default.
| Platform | Recent change | Admin control status |
|---|---|---|
| Microsoft 365 Copilot | Anthropic models available across Copilot experiences as a Microsoft subprocessor | Toggle exists; on by default in most commercial tenants, off by default in EU, EFTA, and UK |
| Google Workspace Gemini | Custom instructions extended to Drive, Chat, Slides, Sheets, and Gmail | No admin control |
| ChatGPT Enterprise | SCIM for API Platform, Codex audit logging, External access controls, ChatGPT for Word | New controls added; some off by default, Word access on by default from October 1 |
Because nothing requires them to. Each vendor runs its own release cadence and its own philosophy about what deserves an admin control. Google treats custom instructions as a user preference. Microsoft treats model provider choice as a data-processing decision and puts it behind a named admin role with regional defaults. OpenAI is still deciding which controls belong at the global-admin level, which is why so many of its additions arrive labeled as previews.
None of those approaches is wrong in isolation. The problem for a small or mid-sized business is that all three often run in the same company, governed by three different philosophies, and nobody has reconciled them into one picture. An employee who drafts a proposal in Word, summarizes a thread in Gmail, and asks ChatGPT to outline a contract clause in the same afternoon has crossed three governance models without seeing a seam. The same is true of the data each tool can reach through connectors, which we break down in our guide to approving AI connectors.
The common belief is that AI platform governance is a setup task: pick the right settings at rollout, document them, move on. Our view is that a configuration is only accurate on the day you checked it. Settings move on the vendor's schedule, new capabilities arrive with defaults you did not choose, and some arrive with no control at all. The teams that stay ahead treat the admin console the way they treat patching: a recurring operational task with an owner and a cadence.
This is also why vendor and subprocessor changes deserve the same scrutiny as any third-party risk. When Microsoft adds Anthropic as a subprocessor, your data path now includes your vendor's vendor, a dynamic we cover in the real AI attack surface is your vendor's vendor. It is the same discipline behind treating AI governance as an operational discipline rather than a policy document.
Start from the assumption that your tenant does not match the vendor's default documentation, because defaults vary by region, plan, and rollout date, and settings get changed by whoever held admin rights last. For each AI platform in use, a quarterly review should answer the questions below, plus an out-of-cycle check whenever a vendor's release notes mention a new default.
For Microsoft tenants, that review pairs naturally with your data controls, since Copilot can only surface what users already have access to. Our walkthrough on safeguarding sensitive data in Microsoft Copilot covers that side, and securing Copilot and AI agents before they become shadow IT covers the agent layer. Underneath both is basic data governance: if permissions are messy, every AI tool inherits the mess.
If you want a structured starting point, our AI readiness assessment walks a team through which AI capabilities are live in their environment versus simply assumed to be off.
It means employees are making AI decisions without the information the admin console holds. Nothing in Word, Gmail, or ChatGPT discloses which model answered, whether a setting defaults on in your region, or whether a capability is governed at all. The issue is not that employees are reckless. It is that the interface gives them no way to see what has been configured behind it.
That makes platform literacy a workforce skill, not just an IT task. A team member who understands that platforms separate user preferences from admin controls, and that the two do not always line up, will ask the right question before adopting a new feature rather than after something goes wrong.
When we take on a client's AI platforms, we begin by mapping every tenant setting that touches models, connectors, and data processing across Microsoft 365, Google Workspace, and ChatGPT Enterprise, then compare that map with what leadership believes is configured. The gaps are usually small individually and significant together: a subprocessor setting nobody reviewed, a Global Administrator account doing routine work, an add-in enabled by default. We assign least-privileged admin roles, document the intended state, and track vendor release notes against it so new defaults are caught and decided on, not inherited.
If you cannot say with confidence what is enabled in your AI platforms today, that is the first question to answer, and it needs an owner and a schedule rather than a one-time cleanup. Pair that with people who understand why the settings matter.
If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. AI University gives your people role-based AI training with security built into the material, not bolted on afterward.
If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.