AIUC-1, published autonomy levels and per-customer certification answer three questions. A third-party standard asks whether a vendor's agents meet a common bar. An autonomy scale describes how independently an agent works. Per-customer certification asks how far this agent may act for this customer, on evidence from that customer's own cases. Only the third decides what an agent may do.
What does each one require, in plain words?
AIUC-1 is a third-party standard for AI agents from the Artificial Intelligence Underwriting Company. For a buyer, a standard of this kind is a common reference point for judging an agent vendor.
Autonomy levels. The Cloud Security Alliance has published autonomy levels for AI agents, running from 0 to 5. A scale like this gives everyone shared words for how much an agent does without a person. Our explainers on what autonomy levels are and how they work cover the idea.
Per-customer certification is narrower and stricter. It requires that an agent's capabilities (read a ledger, draft a refund, release a payment) are tested on real cases from the deploying customer's own operation, against a pass floor, and that the result expires. The authority an agent receives follows from that evidence: one certified depth, checked against the depth each action needs. We set out what it means to certify an agent in our guide.
Who does each apply to?
A third-party agent standard is aimed at vendors. A vendor seeks it once, for its product, and shows it to many buyers. It speaks to how the vendor builds and runs agents across its customer base.
Autonomy levels apply to anyone describing an agent: vendors in marketing, buyers in requirements, risk teams in policies. They are a vocabulary, not an assessment.
Per-customer certification applies to the deploying organisation. It is done for each customer, because the cases, data, policies and limits differ. A refund agent that passes at one retailer may fail at another whose refund rules have more exceptions. Our article on why certification is per customer, not per vendor makes the full case.
What does this mean when the agent acts?
When an agent only answers questions, a vendor-level standard and a broad autonomy label may be enough to judge it. When it takes actions, both fall short in specific ways.
An autonomy level describes the agent, not the action. An agent at a given level might be safe to update a customer's address and unsafe to change their credit limit. A single level hides that difference. We argue in autonomy levels are not enough: certify the agent that authority has to attach to each action.
A vendor standard describes the vendor, not your operation. It can show that a vendor tests, logs and governs its agents well. It cannot show that the agent handles your edge cases, your product terms or your customers' language. That evidence only comes from replaying the agent on your own cases.
Neither expires on its own clock. An agent's behaviour can drift with a model update, a new source or a policy change. Per-customer certificates lapse after a fixed period, and a run that clears no depth drops the agent to depth 1 the same day.
The three fit together as layers. The vendor standard is a reason to shortlist a vendor. The autonomy scale is a way to describe ambition. Per-customer certification is what grants the authority at the moment of action.
How do you combine them?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Vendor meets a recognised bar | Ask for third-party standard results or equivalent assurance during procurement | Vendor's certificate or report, with scope reviewed |
| Shared language for ambition | Describe each agent's target autonomy using a published scale | Agent register entry with target level |
| Each capability tested on your cases | Replay each capability against real cases from your operation, scored against a floor | Test run results per capability |
| Authority matches evidence | The agent's depth set by the test result, up to its role's ceiling | Depth record per agent, with source |
| Authority expires | Certificates lapse on a fixed cycle; failed runs reduce depth immediately | Certificate dates and downgrade records |
| Enforcement at action time | Each action checked against current certificate before it executes | Verdict record naming the deciding bound |
Management system standards add a further layer around all of this. Our piece on ISO/IEC 42001 and agent certification explains how an AI management system and agent certificates support each other.
What will a supervisor or auditor ask?
- What authority does each agent have, action by action?
- What evidence supports that authority, and was it drawn from your own cases?
- When was the agent last tested, and when does its certificate expire?
- What happens when a test fails?
- How do vendor assurances feed into your decision, and what did you test yourselves?
What does none of this cover?
None of these is a law. A third-party standard, an autonomy scale and a per-customer certificate can all support compliance with regulation, but none replaces it. They do not settle data protection duties, conduct rules or sector requirements such as model risk management in banking. And none makes a decision lawful that the law reserves for a person, such as filing a suspicious transaction report.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Pryme certifies each agent for each customer. Attest replays the agent against real cases from the customer's operation and scores every answer against seven gates. The evidence sets the depth, up to the role's ceiling, and no plan limits it. Certificates lapse after 90 days.
Questions people ask
Does passing a third-party agent standard mean we can skip our own testing?
No. A vendor standard tells you about the vendor's practices across customers. Only testing on your own cases tells you whether the agent can do your work within your rules.
Can we map our depths to a published autonomy scale?
You can describe them side by side for communication. The difference is that depth is earned on evidence and checked against the depth each action needs, while a scale usually describes the agent as a whole.
Who should own agent certification inside a firm?
The business owner of the process the agent works in, with risk or compliance reviewing the floors. Certification should not sit only with the technology team that deployed the agent.