Home / Resources / AI Usage Control (AIUC) requirements
Checklist
AI Usage Control (AIUC) requirements: a buyer's checklist
Eight questions to put to any AI Usage Control (AIUC) tool before you buy one. They start from Gartner's public description of the market and are tied to the OWASP risk lists. Use them to ask each vendor to show, not tell.
Version 1 · Published 17 November 2026 · Updated 17 November 2026
What is inside this checklist
- What Gartner says AI usage control tools do
- Eight requirements, each as a question with what a good answer looks like
- A map from OWASP risks to the requirement that covers each one
- How AIBound answers its own checklist
- How to run the checklist in a vendor test
Who this checklist is for
- CISOs and security leads who are choosing a tool for the AI their employees run.
- Security engineers who run the vendor test on a real device.
- GRC and procurement teams who need the same questions asked of every vendor.
What do AI usage control tools do?
Gartner's Market Overview for AI Usage Control, published on 20 August 2026, describes three jobs. These tools discover how employees use third-party AI applications, assess the risk, and let security teams enforce policy.
Gartner's full research is behind a paywall. This page uses only the public abstract, and the eight requirements below are AIBound's reading of it. They are not a Gartner list.
What should you ask a vendor?
Each requirement is a question. Ask the vendor to show the answer on one real employee device, not on a slide.
| Requirement | Question to ask | What a good answer includes |
|---|---|---|
| 1. Find | Does it find AI on the device, not only in traffic or the cloud? | Agents, MCP servers, skills, models and apps found by reading the device, including ones started from a terminal. |
| 2. Identify | Does it link each AI to a person and an account? | The person, the account the AI runs on, and a flag when it is a personal account. |
| 3. Reach | Does it show what each AI can reach? | The systems, files and credentials each AI is connected to, per person. |
| 4. Grade | Does it grade each AI on evidence? | A grade for every AI, with the evidence behind it and a recommendation. |
| 5. Read | Does it read skills and sub-agents before they run? | A review of each skill at install, with the exact line behind each finding. |
| 6. Enforce | Can it allow, ask or deny in real time, inside the agent? | A rule set by you, applied where the AI runs, that the user cannot switch off. |
| 7. Prove | Does it keep evidence an auditor can use? | Decisions and findings kept as records, mapped to the EU AI Act, NIST AI RMF and ISO/IEC 42001. |
| 8. Deploy | Can IT roll it out like any other app, with no other security product first? | A read-only client pushed by IT, with no new agent to run and no suite to buy beforehand. |
Which OWASP risks does each requirement cover?
OWASP publishes the OWASP Top 10 for LLM Applications and the OWASP MCP Top 10. The table shows which requirement helps with some of the main risks in each list. The OWASP MCP Top 10 is a beta release at the time of writing.
| OWASP risk | What it is | Requirements that cover it |
|---|---|---|
| MCP09 Shadow MCP Servers | An MCP server nobody approved or listed | 1. Find, 6. Enforce |
| MCP03 Tool Poisoning | A tool or skill that hides harmful instructions | 5. Read, 6. Enforce |
| MCP02 Privilege Escalation via Scope Creep | An AI that slowly gains more access than it was given | 3. Reach, 6. Enforce |
| MCP07 Insufficient Authentication and Authorization | An AI that acts without a proper check on who it is | 2. Identify, 3. Reach |
| MCP08 Lack of Audit and Telemetry | No record of what an AI did | 7. Prove |
| LLM03 Supply Chain | A risky model, skill or package brought in from outside | 4. Grade, 5. Read |
| LLM06 Excessive Agency | An AI that can do more than its job needs | 3. Reach, 6. Enforce |
How does AIBound answer its own checklist?
AIBound is an Enterprise AI Security Platform. It answers the eight requirements with five products that work from one read-only client on the employee device.
- Find and identify: AISeeker finds every AI agent, personal agent, MCP server, skill, model, app and prompt your employees run.
- Reach: AgentTracer traces each AI from the person, through the account it runs on, to the systems and data it can reach.
- Read and grade: SkillGuard reads every skill like an attacker, grades it from A to F and blocks bad ones at install.
- Enforce: AgentHarness enforces policy inside the agent, with allow, ask or deny.
- Prove: AIAssure turns each decision into evidence for the EU AI Act, NIST AI RMF and ISO/IEC 42001.
IT pushes the AIBound Secure Client like any other app. Install takes 15 minutes, and it needs no other security product first.
How do you use this checklist in a vendor test?
Run the same test for every vendor. A fair test is short and uses a real device.
- Pick one employee device that runs a coding agent, an MCP server and a few skills.
- Ask each vendor to find everything on that device, and compare their lists.
- Ask for the grade and the evidence behind three of the items.
- Set a rule for one risky skill and watch what happens when you install it.
- Ask for the record an auditor would see for that decision.
Check each vendor's answers against its own documentation, and ask for the date of the page you are relying on.
Other resources
How AIBound grades AI from A to FFAQ
Frequently asked questions
What is AI Usage Control (AIUC)?
AI Usage Control (AIUC) is the practice of seeing and controlling how employees and their AI agents use AI at work. It finds what is running, judges the risk and applies your rules on the employee device.
Is this checklist a Gartner list?
No. Gartner's public abstract describes three jobs for these tools: discover AI use, assess risk and enforce policy. The eight requirements here are AIBound's reading of that description, plus the OWASP risk lists.
Does a tool have to meet all eight requirements?
Not on day one. Many teams start with finding AI on the device, then grading it, then enforcing rules. Ask every vendor about all eight, so you know what you are buying and what you are not.
How is AI Usage Control (AIUC) different from AI governance?
Governance sets the rules and keeps the records. AI Usage Control (AIUC) is what finds the AI in use and applies the rules on the device. Governance runs better when AI Usage Control (AIUC) supplies the evidence.
Sources referenced
Where the facts on this page come from. The facts about AIBound come from AIBound's own product pages.
- Source 1. Gartner. Market Overview for AI Usage Control (abstract). Checked 17 November 2026.
- Source 2. OWASP Foundation. OWASP MCP Top 10. Checked 17 November 2026.
- Source 3. OWASP Gen AI Security Project. OWASP Top 10 for LLM Applications. Checked 17 November 2026.