How do you approve or prohibit AI providers?
An approved AI providers list works only when it's enforced on the devices and agents people use. Record each decision against the provider or client that Arcjet observes, enforce it on managed devices, and count any use after the decision as a violation attributed to a person where Arcjet can identify one. Prohibiting the ChatGPT app while approving the OpenAI API is one decision.
A provider decision is an approval or prohibition recorded against an AI provider, such as OpenAI, or against one client of it, such as the ChatGPT app. It's the third of the four records that AI governance frameworks ask for, after the AI inventory and a named owner for each AI system. For how the four fit together, see what is AI governance.
Frameworks ask for this decision in their third-party sections. NIST AI RMF 1.0 (the National Institute of Standards and Technology AI Risk Management Framework) covers policies for third-party AI risk in GOVERN 6.1 and the monitoring of third-party resources in MANAGE 3.1. Arcjet's evidence page also maps provider decisions to ISO/IEC 42001.
Why isn't a written approved AI tools list enough?
A written list isn't enough because nothing connects it to what happens on a laptop. An AI acceptable use policy that names three approved tools says nothing about the fourth tool someone installs next week, and nobody can tell from the document whether the list holds. An auditor asking for GOVERN 6.1 evidence wants to see the decision and whether people followed it.
Prohibition without an alternative also moves use rather than stopping it. A person who needs to summarize a contract and finds the approved route closed uses a personal account on the same laptop, out of reach of your logs and contracts. The shadow AI policy template is built around this failure: each enforceable clause names the control that backs it and the record that shows whether it held. Provider decisions are the control and the record for its approved tools clause.
Approve first, then prohibit. Publish what's approved and how to request something new before any block takes effect, so that a person who meets a block has somewhere to go.
Should you decide at the provider or the client level?
Decide at the client level when one provider offers routes with different terms, and at the provider level when every route is the same to you. A provider is the company behind a model service. A client is one way to reach it, such as the ChatGPT app versus the OpenAI API. The following table shows when each level fits.
| Situation | Level | Example decision |
|---|---|---|
| You have no contract or data processing terms | Provider | Prohibit the provider |
| You've approved every route the provider offers | Provider | Approve the provider |
Your contract covers the API, but people also use the consumer app | Client | Approve the OpenAI API and prohibit the ChatGPT app |
You've approved a coding agent but not the provider's chat app | Client | Approve the agent's client and prohibit the chat app |
A client-level decision is one record with one owner and one review date.
A client isn't an account. The ChatGPT app signed in to your company workspace and the same app signed in to a personal account are the same client to Arcjet, because they use the same hostnames. To refuse personal accounts in an approved client, use the vendor tenant controls in block personal AI accounts on work laptops.
How does Arcjet enforce a provider decision?
Arcjet enforces a provider decision on managed devices through the Arcjet endpoint agent, the same agent that runs Discovery. When a person's device opens a new connection to a prohibited provider or client, the agent blocks it on the device. Blocked attempts appear in Arcjet with the person, device, app, and provider.
The agent identifies the signed app that opened each connection and the person behind the local account. Blocking a provider doesn't require reading HTTPS content, so the agent doesn't read prompts. That's why a decision can separate a desktop app from an API client: they're different signed apps.
Your own applications and coding agents need a second layer, because a provider decision on a device doesn't cover a server or a cloud session. For the AI systems you build, Guard policies decide which models and tools an agent may call at runtime. For coding agents, the coding-agent.model-allowlist and egress allowlist starter policies enforce the same choices in the hooks that Claude Code, GitHub Copilot, Cursor, OpenAI Codex, and Muse Code fire. For details, see restrict models in coding agents and the Arcjet coding agents guide.
What counts as a provider violation?
A violation is an observation of a prohibited provider or client at or after the time of the decision, attributed to a person. Use before the decision isn't a violation, because the person had no rule to break. That keeps the count honest on the day you prohibit something that half the company already uses.
Arcjet counts observations, not connections. An observation is one person, device, account, app, and destination, so a person whose ChatGPT app opened 300 connections in an afternoon counts as one observation, not 300. The count tells you how many people and devices still reach a prohibited service, which is the number you act on.
The following cases don't count as violations:
- Use observed before the decision time.
- Use of an approved provider or client, or one with no decision yet.
- Use on a device without the endpoint agent, which Arcjet doesn't observe.
Violations appear in the Discovery lists and in the weekly owner digest, which tells each provider's owner about violations on their providers. A connection from Unknown user, such as a shared or system account, still counts, but has no person to follow up with. Map those accounts to people when you can.
How do you handle exceptions to a provider decision?
Handle exceptions by narrowing the decision first and by keeping an exception register second. Many requests for an exception are really requests for a different client. A data science team that needs a model provider's API for an approved project doesn't need the consumer app, so approve the API client and leave the app prohibited.
When a person needs a prohibited client for a limited time, record the exception in your register with a business reason, an approver, and an end date. Violations from that person still appear in Arcjet, which is what you want: compare them against the register, and anything not covered by an open exception needs a conversation. When an exception expires and violations continue, the record shows it.
Avoid approving a provider broadly to make one exception go away. A broad approval applies to everyone, has no end date, and is easy to forget at the next review.
How do you roll out provider decisions, step by step?
The following procedure starts from what you observe and stages enforcement so that blocks don't surprise anyone.
- List the providers people use. Start from the providers and clients in your AI inventory, sorted by how many people use each one.
- Approve what you already support. Approve each provider or client you have terms for, and publish the approved list with a way to request others.
- Prohibit what you can't support. Prohibit providers with no terms, and prohibit specific clients where only another route is approved, such as the ChatGPT app when the OpenAI API is approved.
- Give each decision an owner. Assign an owner who must be a member of your Arcjet team, and write why the decision was made in its purpose.
- Enforce in stages. Roll out enforcement to a pilot group first, read what was blocked, then widen it to the fleet.
- Follow up on violations. Read violations by person, check them against your exception register, and talk to the people behind them.
- Review each decision every 180 days. Confirm the decision still matches your contracts and use, and mark it reviewed. For what a review covers, see how to run periodic AI system reviews.
Only team administrators record decisions. Other members of the team read them. Every change to a decision, its owner, purpose, or review is written to the WorkOS audit log with the fields that changed, and the purpose text itself is left out.
Where Arcjet fits
Arcjet records provider decisions on the providers and clients that Discovery observes, enforces them on managed devices through the endpoint agent, and counts later use as violations attributed to a person. Guard policies and coding-agent hooks carry the same choices into the agents you build and run.
The decisions, their owners, and their violations supply evidence for NIST AI RMF GOVERN 6.1 and MANAGE 2.4 on the evidence page, which also maps them to ISO/IEC 42001. They don't cover supplier assessments, contract terms, or intellectual property risk, which stay with your procurement and legal teams. They also don't separate personal from company accounts in the same client, which needs the vendor tenant controls.
After the decisions are in place, set up reviews so they don't go stale. See how to run periodic AI system reviews, and for the full governance picture, see what is AI governance.
Frequently asked questions
Does a provider violation count every connection?
No. Arcjet counts observations, not connections. An observation is one person, device, account, app, and destination, so an app that opens hundreds of connections to a prohibited provider in a day counts once. Only observations at or after the decision time count.
Can Arcjet block a personal ChatGPT account but allow the company workspace?
Not with a provider decision. A personal account and a company workspace in the ChatGPT app use the same hostnames, so they're the same client to Arcjet. Use the vendor's tenant restriction on a TLS-inspecting proxy, such as the ChatGPT-Allowed-Workspace-Id header, for that split.
Does a provider decision apply to my own applications and servers?
Device enforcement covers managed devices with the Arcjet endpoint agent. For applications and agents you build, use Guard policies to decide which models and tools they may call at runtime. For coding agents, use the model allowlist and egress allowlist starter policies in Arcjet hooks.
Who can approve or prohibit a provider in Arcjet?
Team administrators record provider decisions, and other team members read them. Every change to a decision, its owner, purpose, or review is written to the WorkOS audit log with the fields that changed and the new owner. The purpose text isn't logged.
Which frameworks ask for decisions about AI providers?
NIST AI RMF 1.0 covers policies for third-party AI risk in GOVERN 6.1 and monitoring of third-party resources in MANAGE 3.1. In Arcjet, provider decisions and violations supply evidence for GOVERN 6.1 and MANAGE 2.4, and Arcjet's evidence page also maps them to ISO/IEC 42001. They don't cover supplier assessments or contract review.
AI runtime security in your code
Protect your AI agent workflows with Arcjet
Get allow, deny, and redact on agent actions before the side effect.