Raphael Friedmann
← Understanding AI

Choosing an AI provider: a checklist for IT

What to settle before signing with an AI provider: DPA and EU hosting, SSO and audit logs, model retirement, limits and SLAs, pricing, certificates, an exit plan and a test on your own work.

A sales demo shows you the model. IT has to live with everything around it: the contract, the admin console, the rate limits and the day the model you built on gets switched off. This is the checklist I work through before a provider gets company data. Vendor details are as of November 2026 and linked to the vendor’s own pages; they change often, so re-check them on the day you sign.

Evaluation in six steps
1must-havesDPA, EU region, SSO2paper checkterms, certificates3shortlisttwo or three vendors4own-work testyour tasks, your data5pilota few teams, real work6contractexit plan, notice periodsa missing must-have ends it for that vendor, at any step
The paper checks are cheap and come first. The test on your own work is the expensive step, so only the shortlist gets there.

1. Contract and data location#

Ask for the data processing agreement (DPA) before anything else, and check that it is part of the contract you sign, not a separate page that can change. Anthropic’s Commercial Terms incorporate its DPA by reference, and customers in the EEA, Switzerland or the UK contract with Anthropic Ireland. Azure’s model hosting is governed by the Microsoft Products and Services DPA. Why this matters under the GDPR is in the lesson on the law.

Then ask where prompts are processed and stored, and what it takes to keep them in the EU. The answers differ a lot:

2. Retention and training#

All four large providers keep business and API data out of training by default. The questions that remain are how long prompts sit in abuse monitoring logs, who can read them, and whether zero data retention is available for your use case and with which exclusions. Get the answers in the contract, not from a help page.

3. Identity, admin and audit#

  • SSO and SCIM. Single sign-on is the minimum; SCIM provisioning is what removes a leaver’s access automatically. On OpenAI’s and Anthropic’s pricing pages, SCIM is an Enterprise feature and not part of ChatGPT Business or Claude Team.
  • Audit logs. Ask what is logged, how you export it to your SIEM and how long the vendor keeps it. OpenAI’s compliance logs are kept 30 days, so you need to pull them continuously. Claude’s audit log export contains identifiers, not chat content.
  • Admin controls. Can you restrict models, connectors and features per group, and set retention for the workspace yourself?

4. Models and their lifecycle#

The model you test today will be retired. What matters is how much warning you get:

  • OpenAI: at least 6 months for generally available models, 3 months for specialized variants, preview models as little as 2 weeks.
  • Anthropic: at least 60 days’ notice for publicly released models.
  • Azure Foundry: generally available models get an 18-month lifecycle (12 months for some partner models), notice at least 60 days before retirement, and retirement dates cannot be extended. Standard deployments are upgraded automatically, provisioned ones are not.

Pin model versions in production, and keep the test set from step 4 so you can re-run it on the successor. Also ask whether you can use other vendors’ models on the same platform: Azure, Bedrock and Google Cloud host several model families.

5. Limits and SLA#

Rate limits on the OpenAI and Anthropic APIs rise in tiers with your credit purchases or usage history, so a new account starts low. Ask what limit you get on day one and whether reserved capacity exists: Anthropic sells a Priority Tier with committed tokens per minute, Azure sells provisioned throughput. For availability, Microsoft states 99.9% for Azure OpenAI. Whatever the number, you want it in a contract with service credits; a status page is not an SLA.

6. Pricing model#

Chat products are sold per seat, APIs per token. Seats are predictable and also bill the people who never open the tool. Tokens bill actual use and need budget limits and monitoring. Look for multipliers on top of the list price: the residency surcharge above, Anthropic’s 1.1x for US-only inference, and higher rates for long prompts on some models. If the volume is high and steady, self-hosting becomes a fair comparison.

7. Certifications#

ISO/IEC 27001 and a SOC 2 Type II report are the baseline; ISO/IEC 42001 covers the management of AI systems. What the vendors state:

Read the scope. A certificate covers named products, and “aligned with ISO 27001” is not a certificate.

8. Exit plan#

Decide how you would leave before you arrive. Can you export chats, files and custom assistants in a usable format? Is your integration written against a proprietary API or a common one? The OpenAI-style chat completions API has become the shared interface: Google offers an OpenAI-compatible endpoint for Gemini, and Anthropic offers an OpenAI SDK compatibility layer, which it describes as mainly for testing and comparing models, with features such as prompt caching missing. Self-hosted servers speak it too: imp, the inference engine I build, serves /v1/chat/completions and Anthropic’s /v1/messages, so moving an application to a local model is a change of base URL. Keep prompts, test sets and integration code in your own repository, not only in a vendor’s assistant builder.

9. Test on your own work#

Benchmarks rank models on someone else’s tasks. Before you decide, run the shortlist on a set of real tasks from your own work and grade the results the way you grade a colleague’s; how to set that up has its own lesson. It is also the test set you re-run when a model is retired.

The short version#

Question Why it matters Red flag
Is a DPA part of the contract? GDPR requires a processor contract DPA only as a changeable web page, or none
Where are prompts processed and stored? Transfers outside the EU need a legal basis “Global” routing only, no region in writing
How long are logs kept, who reads them? “No training” still allows storage and review No period stated, no ZDR option at all
SSO and SCIM included? Leavers must lose access automatically SCIM only in a tier you will not buy
Can you export audit logs? Incidents and eDiscovery need them No API, or logs expire before you can pull them
What is the deprecation notice? Every model gets retired No published policy, or weeks of notice for GA models
What are your limits and SLA? A pilot can work and production fail No SLA, limits only after spend history
Per seat or per token? Decides who pays for idle users or heavy use Multipliers not on the price list
Which certificates, for which product? Scope decides what is covered “Aligned with” instead of certified
How do you leave? Lock-in is a cost you pay later No export, proprietary API only

Bottom line#

Most of this checklist is paperwork, and that is the point: DPA, region, SSO, audit logs and deprecation policy can be checked in a day and rule vendors out cheaply. Spend the real effort on a test with your own tasks for the two or three that remain, sign with an exit plan in hand, and keep your integration on an interface that more than one provider, including your own hardware, can serve.