Blog · AI agents · Controls
Give every AI agent an ID badge: identity, permissions and audit trails
16 October 2026 · 4 min read · By LeadAI
AI agents are moving from experiments into everyday work. In recent weeks Cisco has added agents that can be invited into Webex meetings and spaces to carry out multi-step tasks, Workday has made an agentic assistant its main interface for HR, finance and IT, and vendors such as ServiceNow and RSA have launched tools to verify and govern agent identities. Agents are becoming digital coworkers.
Coworkers have badges. They have a name, a manager, a set of doors they can open and a record of where they went. Most agents have none of these. This article gives you a simple badge to put on every agent, and a way to start in 30 days.
The seven-field badge
| Field | The question it answers | Example |
|---|---|---|
| 1. Name and purpose | What is this agent for, in one sentence? | "Drafts first replies to billing queries." |
| 2. Owner | Which named person is accountable? | Head of customer support |
| 3. Autonomy level | Does it suggest, act with approval, or act alone? | Acts with approval above a set value |
| 4. Access | Which systems, data and actions, and for how long? | Read billing records; cannot issue refunds |
| 5. Credentials | Does it have its own identity that can be revoked? | Own service identity, short-lived credentials |
| 6. Record | What did it do, when and on whose request? | Action log kept and reviewed monthly |
| 7. Off switch | How is it stopped, by whom, and when was that last tested? | Disabled by IT on request; tested in September |
If you cannot fill in a field, that is a finding. Write it down and decide what to do about it.
Why autonomy level matters
An agent that can only read is a very different thing from one that can send, buy or delete. Sort every agent into one of three groups and treat the third with the most care:
- Suggests. A person decides and acts.
- Acts with approval. The agent prepares the action and a person confirms it.
- Acts alone. The agent completes the action without a person in the loop.
Moving an agent up a level should be a decision that someone signs off, not a side effect of a product update.
Three common mistakes
1. Shared accounts and borrowed logins. An agent running under a shared service account, or under one person's login, makes it impossible to tell who did what. Give each agent its own identity and the minimum access its task needs.
2. Relying on the kill switch alone. Being able to stop an agent is essential, but it only helps after something has gone wrong. If the agent already has broad access, the switch contains the damage rather than preventing it. Limit access first, then keep the switch tested.
3. Registering agents only at launch. Teams build agents with low-code tools and switch on agent features in software you already pay for. Treat your agent register like your AI inventory: a declared list plus a discovered list, compared regularly. Our guide to building an AI inventory explains how.
Where ISO/IEC 42001 fits
ISO/IEC 42001 is a management standard, not a technical control. It will not configure an agent for you. What it does is make you show that each AI system has an owner, that risks were assessed, that operation is controlled and monitored, and that incidents are handled. The badge gives you much of that evidence in one place.
If you already run or plan an AI management system, add agent-specific controls inside it rather than starting a separate programme. The register, the roles, the review cycle and the audit trail are the same ones you need anyway.
A 30-day start
- Week 1: Find. List every agent you know of. Ask teams, check vendor settings, and look at API keys and cloud accounts.
- Week 2: Own and classify. Name an owner for each agent and set its autonomy level.
- Week 3: Reduce. Cut access to what each task needs and move agents onto their own credentials.
- Week 4: Test. Switch one agent off on purpose and time it. Review a sample of its action log and fix what is missing.
Regulators are starting to treat the organisation that gives an agent instructions and access as the one that answers for it. Our note on the FTC investigation into rogue agents explains why.
To see how an inventory, owners and evidence fit together, take the free AI Readiness Check or follow the LeadAI Assure journey.
This article gives general information and is not a substitute for ISO/IEC 42001:2023 or professional advice.
Sources
- AI Agents News, week of 17 October 2026: https://aiagentstore.ai/ai-agent-news/this-week
- AI Agents News Brief, 16 October 2026: https://aiagentsdirectory.com/news/ai-agents-news-brief-october-16-2026
- Zuplo, what regulators expect of AI agents: https://zuplo.com/blog/what-regulators-expect-ai-agents
- NHI Management Group, AI agent kill switches and identity governance: https://nhimg.org/articles/ai-agent-kill-switches-expose-a-deeper-nhi-governance-gap/