subinternal Talk to us

ResearchGovernance

Who’s watching your AI?

The vendor runs the model, but your company answers for what it does with it. What each vendor gives you to watch and limit assistants and agents, what the standards ask, and the setup we would start with.

An engraving of an empty control room: a curved desk of dials, gauges and levers, a wall of brass indicator lamps, a blank chart and a pendulum clock, an open logbook and pen beside a brass lamp, and tall arched windows onto power lines crossing a valley.
ControlOpenAIAnthropicMicrosoftGoogle
Audit logWho did what, and whenAPI audit logsChatGPT’s not confirmed by us180 days, exportedIDs only, no chat textPurview, automaticLists the files Copilot readAdmin console and APISays when an agent acted
RetentionHow long chats are keptSet per projectAPI projects, by the Admin API30 days or moreKept indefinitely until setTwo sets of termsSome Claude models keep dataVault rules and holdsFor the Gemini app
Spend limitsWhat it may costHard monthly capRequests stop at the capOrg, group, memberWith pooled group budgetsUser, agent, environmentNew services join by defaultProject and account capsOverage is an admin switch
Who uses which modelWho gets whatAllowlist per projectAPI projectsPer roleEfficient mode for cheaper modelsClaude by user or groupOn by default outside the EUUsage reportsBy feature, app and user
Warning before a changeModel retirementsNone statedAliases move; pin a snapshotAt least 60 daysUsage export finds old modelsAt least 60 daysPreview models 30, then upgradedEarliest dates onlyOld names can be rerouted
What each vendor gives a company to oversee its AI, read from the vendor’s own pages in November 2026 and cited below: OpenAI’s API, Claude Enterprise and the Claude API, Microsoft 365 Copilot and Azure, Google Workspace and the Gemini API. A white mark is offered as standard, a gray one with a condition worth knowing, a dashed outline not offered or not confirmed by us.

Who is accountable for what an assistant does?

The company that puts it to work. A vendor runs the model, but your company decides who may use it, what data it can reach, what it may do on its own, and how long its conversations are kept. Every control in the table above is a setting your admins choose; none of them is on in a useful way until someone owns it.

The law reads the same way. Under the EU AI Act, a company that uses a high-risk AI system has duties of its own: it must use the system as its maker instructs, assign human oversight to a person in the company, and tell workers before the system is used on them.1 Most assistants at work are not high-risk systems, but the shape is the lesson: someone inside the company is answerable, by name.

What changes without anyone asking?

The model, the defaults and the bill. Vendors retire models on their own schedules, and the warning varies widely. Anthropic gives at least 60 days’ notice before retiring a publicly released model, by email, though Amazon Bedrock and Google’s own platform set their own dates for the same models.2 Microsoft promises at least 60 days for generally available models on Azure and 30 for previews, which are then force-upgraded, and says plainly that retirement dates can’t be extended.3 Google lists only the earliest possible shutdown dates. On October 8, 2026 it deprecated gemini‑3.7‑flash and sent its requests to gemini‑3.8‑flash the same day, and it gave Veo 2.0 and 3.0 fifteen days.4, 5

Googlegemini-3.7-flash0routed to 3.8 Flash the day it was deprecatedGoogleVeo 2.0 and 3.015June 15 to June 30, 2026AzurePreview models30at least; then force-upgradedAnthropicPublic models60at least, by emailAzureGA models60at least; dates can’t be extended
Days of warning before a model is retired or replaced. A solid bar is the least a vendor promises; an outlined one is a single case from Google’s changelog. From each vendor’s pages, read November 2026.

Names move too. OpenAI repointed its chat‑latest alias at least five times between May and October 2026, and since July its gpt‑5.6 alias routes to gpt‑5.6‑sol.6 A team that calls an alias is running whatever model the vendor points it at today.

And defaults change. Microsoft turns Anthropic’s models on by default for most commercial tenants outside the EU, EFTA and UK.7 Its Copilot spending policies apply to new metered services and agents automatically unless an admin turns that off.8 None of this is hidden; all of it is published. The risk is that nobody in the company is reading it.

What do the vendors give you to watch with?

More than most companies use. Each of the four now has audit logs, spend limits and some control over which models people use, though they differ in the details.

OpenAI’s API has an Admin API with an audit log of user actions and configuration changes, a model allowlist or denylist for each project, and data retention set project by project.9 Since July 22, 2026 a hard monthly spend cap makes requests fail once tracked spend reaches it.6

Claude Enterprise exports the past 180 days of audit logs: sign-ins, invitations, settings, file uploads and conversations created or deleted, by ID, without the text of any chat.10 Retention can be set to 30 days or more; until it is, data is kept indefinitely.11 Since October 2026, admins set spend limits for the organization, a group or a member, give a group a pooled monthly budget, and choose which models each role can use.12 On the API, a usage export broken down by key and model shows where an old model is still called.2

Microsoft 365 Copilot logs each interaction in Purview automatically once auditing is on, as part of Audit Standard, with the files, sites and emails Copilot read to answer, any agent involved, and the model’s provider.13 Admins choose which users or groups may use Anthropic’s models, and models that make Anthropic keep data stay off until an admin opts in.7 A cost dashboard sets limits by user, agent and environment, with read-only roles for finance.8

Google Workspace records Gemini activity as audit events in the Admin console and the Reports API, including whether an agent took the action.14 Usage reports show Gemini use by feature, app and user.15 Since June 2026, Vault applies retention rules and litigation holds to the Gemini app, and they override user and console settings.16 The Gemini API has spend caps by project and by billing account, and Gemini Enterprise bills usage over its quota only after an admin turns that on; otherwise the features stop until the quota resets.5, 17

What do the controls leave out?

Three things, mostly. First, content: Claude Enterprise’s audit log carries IDs, not the words of a chat, so reviewing what was said takes a separate export of the data.10 Second, tools outside the suite: Microsoft logs people’s use of AI apps it doesn’t make only on pay-as-you-go billing, kept 180 days.13 Third, terms that move with the model. Microsoft’s Data Protection Addendum does not cover Anthropic models labeled as keeping data. Anthropic may keep that content up to 30 days, and up to two years if flagged.7

Agents widen each gap. Google and Microsoft now mark which actions an agent took and which agent it was.14, 13 That tells you what happened. It does not decide in advance what an agent may do, which is a choice your company has to make and write down.

What do standards and laws ask?

A named owner, a written process, and records. NIST’s AI Risk Management Framework is a voluntary framework for managing AI risk, with a profile for generative AI published in July 2024, and NIST says the framework is being revised.18, 19 ISO/IEC 42001, published December 18, 2023, sets out requirements for an AI management system: establishing one, running it and improving it. It is written for any organization that provides or uses products built on AI.20

The EU AI Act is law, in stages. Its bans and its AI literacy duties have applied since February 2, 2025, and its rules for general-purpose models since August 2, 2025. Most of it applies from August 2, 2026, including the transparency duties in Article 50. After the Digital Omnibus took effect on July 27, 2026, rules for high-risk systems start December 2, 2027, and for those built into products August 2, 2028.1 This is a summary, not legal advice; a company that sells into the EU should read the Act with its counsel.

What does a sensible oversight setup look like?

Small, named and written down. This is our judgment, built from the controls above and our own work setting assistants up, not a standard’s wording. Start with one owner for AI across the company, and a register: each assistant and agent, what it can reach, who may use it, and the model behind it.

Then turn on what the vendor already gives you. Send the audit logs to wherever security already looks. Set retention on purpose rather than leaving the default. Put spend limits at every level the vendor allows, and turn off automatic coverage of new services where a person should approve them first. Choose which people get which models.

Pin what matters. Where an output feeds a process, call a dated model version rather than an alias, and on Azure choose when a deployment upgrades. Put the vendors’ retirement notices in someone’s inbox, and check the usage export each month for models nearing retirement.

Give agents less than people. An agent gets its own identity, the narrowest access that does the job, and a person’s yes before anything that spends money, sends a message or changes a record. Review the register, the logs and the spend each month. This maps onto the NIST and ISO approaches without claiming to meet either; certification is its own project.

This is work we do. We set up and run the assistants a company’s people use, Claude, ChatGPT, Copilot and Gemini, with these controls in place from the first day. Where data can’t leave the building, we build private AI that runs on your own machines. If you want to talk through yours, get in touch.

References

  1. European Commission, “Navigating the AI Act.” Read November 2026. digital‑strategy.ec.europa.eu
  2. Anthropic, “Model deprecations.” Read November 2026. platform.claude.com
  3. Microsoft, “Foundry Models lifecycle and support policy.” Updated July 2026. learn.microsoft.com
  4. Google, “Gemini API deprecations.” Read November 2026. ai.google.dev
  5. Google, “Gemini API changelog.” Read November 2026. ai.google.dev
  6. OpenAI, “API changelog.” Read November 2026. developers.openai.com
  7. Microsoft, “Anthropic models in Microsoft Online Services.” Updated October 2026. learn.microsoft.com
  8. Microsoft, “Usage-based billing and cost management for Copilot Credits.” Updated October 2026. learn.microsoft.com
  9. OpenAI, “Admin APIs.” Read November 2026. developers.openai.com
  10. Anthropic, “Access audit logs.” Read November 2026. support.claude.com
  11. Anthropic, “Configure custom data retention controls for Enterprise plans.” Read November 2026. support.claude.com
  12. Anthropic, “Claude Cowork and chat come together for Enterprise.” October 2026. claude.com
  13. Microsoft, “Audit logs for Copilot and AI applications.” Updated August 2026. learn.microsoft.com
  14. Google, “Gemini for Workspace log events.” Read November 2026. knowledge.workspace.google.com
  15. Google Workspace Updates, “View Gemini feature usage and threshold reports in the Admin console.” February 2026. workspaceupdates.googleblog.com
  16. Google Workspace Updates, “Google Vault now supports retention rules and litigation holds for the Gemini app.” June 2026. workspaceupdates.googleblog.com
  17. Google Cloud, “Quotas and pay-as-you-go usage.” Updated October 2026. docs.cloud.google.com
  18. NIST, “AI Risk Management Framework.” Read November 2026. nist.gov
  19. NIST, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile” (NIST AI 600‑1). July 2024. doi.org
  20. IEC Webstore, “ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system.” December 2023. webstore.iec.ch