Who's really running your AI?
In the last 12 months, seven of the nine enterprise systems we track each independently signed their own deal with an AI lab. Most companies running two or three of these platforms have never added up what that means.
Published
Your vendors picked your model for you
If your company runs SAP, Salesforce, ServiceNow, Snowflake, or Workday, an AI lab is already running inside at least one of them — and you didn't sign that contract. Each of these vendors independently decided which large language model to build its agent features on, over roughly the same twelve months, largely without coordinating with each other or announcing it as a single industry moment.
That's not a criticism of any one vendor's choice. It's a fact about your risk surface that's easy to miss when each announcement lands separately, months apart, from a different vendor's press office.
The landscape, system by system
| System | LLM posture | What changed |
|---|---|---|
| SAP | Anthropic named | Claude named “a primary reasoning and agentic capability” across the Autonomous Enterprise (Sapphire, May 2026) — alongside separate deals with Mistral, Cohere, and OpenAI via SAP’s EU AI Cloud. |
| Oracle | No lab partnership | Verified directly against Oracle’s own OCI model catalog: no Anthropic model appears anywhere. Oracle added Google Gemini to Agent Studio in July 2026; its own AI stays model-agnostic by vendor choice. |
| Salesforce | Anthropic named | “Claudeforce,” announced Aug 2026: Claude set as the default model across six Salesforce surfaces, including a new plugin that acts on live CRM records. |
| ServiceNow | Anthropic + OpenAI, both named | Signed multi-year deals with OpenAI and Anthropic one week apart (Jan 2026), then retired its own in-house model as the platform default in favor of both. |
| Snowflake | Anthropic + OpenAI, both named | Symmetric $200M partnerships with Anthropic and OpenAI; Claude specifically powers Snowflake’s own Intelligence/CoWork features. |
| Databricks | No single lab — hosts all of them | Signed deals with Anthropic, OpenAI, and others, but the product itself is built to host and switch between any model — including open-weight ones — rather than anchor to one. |
| Workday | Google named | Expanded Google partnership (May 2026): Gemini is now the default model for Workday’s Sana agent front door. |
| Microsoft | Anthropic + OpenAI, both deeply embedded | ~27% owner of OpenAI and a multi-billion-dollar investor in Anthropic at the same time; Claude is set as the default AI processor across most Microsoft 365 commercial tenants. |
| Palantir | No lab partnership (by design) | Stays strictly multi-model on its own platform. Its Anthropic relationship is a government-cloud distribution deal, not a product anchor. |
The question none of these vendors can answer about the others
Run SAP and Salesforce together, and you're touching Anthropic through two separate contracts you never signed with Anthropic. Add Snowflake, and it's a third. None of these platforms' own governance consoles see past their own walls — SAP's agent-governance tooling can't see what's running inside Salesforce, and Salesforce's audit trail can't see SAP.
Microsoft's Agent 365 is the one product that markets a cross-platform view, and it's a genuinely useful inventory of which agents exist and how they're configured across several of these systems. What it doesn't answer is which model lab sits under which vendor contract, on what data terms, or what your combined exposure looks like if one of those labs changes its terms — a different question, and one Microsoft has its own stake in either lab's answer to.
None of this means any of these AI-vendor choices are wrong. It means the aggregate view — what you're actually running, across every vendor, under whose contract — isn't something any single platform is positioned to hand you.
See it built, not just described
Standing Watch Assessment is the engagement this pattern comes from.
Standing Watch Assessment →