---
title: What should an AI agent registration questionnaire ask?
url: https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# What should an AI agent registration questionnaire ask?

**Ask only what you cannot detect automatically: the agent's purpose, the data it touches, the actions it can take, whose identity it uses, and who owns it. Around fifteen questions is enough, and each one should change something, usually the agent's risk tier and therefore which controls apply. Everything a scanner can find (models, keys, tools, platform) should be pre-filled, not asked.**

Registration questionnaires have a bad reputation because many are long, generic and never read again. A good one is short, specific to agents, and wired to decisions. It is also the moment a discovered shadow agent becomes a governed one.

## What should be pre-filled rather than asked?

Before the owner sees a single question, discovery and the registry should already know:

- Platform and environment (cloud account, agent platform, development or production)
- Models and model versions in use
- API keys and credentials attached
- Tools and MCP servers connected
- Creator, from the platform's audit data
- First seen and last active dates

Asking people to type these in invites errors and delays. Show them, and ask the owner to confirm.

## Which questions should the questionnaire ask?

| # | Section | Question | What the answer drives |
|---|---|---|---|
| 1 | Purpose | What business task does this agent perform, in one sentence? | Reviewer context; duplicate detection |
| 2 | Purpose | Who uses it: employees, customers, partners or no humans? | Tier escalator for external users |
| 3 | Purpose | Does it make or support decisions about people (credit, hiring, claims, access)? | High-risk flag; legal review |
| 4 | Data | What is the most sensitive data it reads? | Data sensitivity score |
| 5 | Data | Does any data leave your environment, and to which regions? | Residency controls |
| 6 | Data | Does it store memory or conversation history? | Retention and redaction rules |
| 7 | Actions | Can it write, send, buy, delete or change anything? | Reversibility score |
| 8 | Actions | What is the largest single action it can take (value, records, recipients)? | Blast radius score |
| 9 | Actions | Is a human in the loop before irreversible actions? | Tier escalator if no |
| 10 | Identity | Does it act as itself, as a service account, or on behalf of the user? | Identity pattern review |
| 11 | Identity | Which systems does it authenticate to? | Least-privilege review |
| 12 | Ownership | Who is the business owner? | Accountability; attestations |
| 13 | Ownership | Who is the technical owner? | Incident contact |
| 14 | Testing | Has it been evaluated or red-teamed, and where are the results? | Release gate |
| 15 | Lifecycle | When should it be reviewed or retired? | Review schedule |

Questions 12 and 13 should be identity-provider lookups, not free text. See [assigning owners to AI agents with your identity provider](/blog/assign-owners-to-ai-agents-with-your-identity-provider).

## How should answers set the risk tier?

Keep the logic visible so owners understand why their agent landed where it did.

1. **Score data sensitivity** from question 4, reversibility from question 7 and blast radius from question 8, each from 1 to 3.
2. **Take the highest score** as the base tier.
3. **Escalate one tier** if users are external (question 2), there is no human in the loop (question 9), or it supports decisions about people (question 3).
4. **Attach controls by tier.** Low tier: registry entry and owner. Medium: evaluation before release and gateway routing. High: red-team results, delegated identity, runtime guardrails and quarterly attestation.

This mirrors the "Map" function of the NIST AI Risk Management Framework, which asks organisations to establish context and categorise AI systems before measuring and managing risk. ISO/IEC 42001 similarly expects an AI system impact assessment.

## Can the questionnaire differ by team or risk level?

Yes, and it should. Two techniques keep it short:

- **Branching.** If question 7 is "read-only", skip question 8. If question 2 is "no humans", skip the user-facing transparency questions.
- **Team overlays.** A bank's payments team may add a question about transaction limits; a marketing team may add one about brand claims. Keep the core fifteen identical so tiers stay comparable across the organisation.

Version the questionnaire itself, so you know which questions an agent answered when it was registered.

## How do you keep registrations current?

A registration is a snapshot. Agents change weekly.

- **Re-trigger on change.** A new tool, model or data source should reopen the relevant questions only.
- **Attest periodically.** Owners confirm answers each quarter for high-tier agents and yearly for others.
- **Compare answers with evidence.** If the owner says "read-only" and traces show write calls, flag it. This is where questionnaires give way to [continuous evidence](/blog/replace-ai-governance-questionnaires-with-continuous-evidence).

For the wider picture of where registration sits, read [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and through connectors to platforms such as Salesforce, SAP, Databricks and HubSpot, so most technical fields arrive pre-filled. Discovered agents move into a governed registry of agents, versions, tools and MCP servers. OIDC SSO ties owners to your identity provider, and Posture checks for risky permissions and configurations, which helps compare what owners declare with what agents can actually do. Registration does not mean runtime takeover: depth of control, from registry to observe to enforce, depends on the connector and traffic path. [Book a demo](https://www.lyzr.ai/book-demo/) to see the registration flow.

## Frequently asked questions

**How many questions should an AI agent registration form have?**
About fifteen core questions, with branching, is enough. Pre-fill everything discovery can detect.

**Who should fill in an AI agent registration questionnaire?**
The business owner answers purpose and impact questions; the technical owner answers data, action and identity questions.

**How often should registrations be updated?**
Reopen the relevant questions whenever the agent gains a new tool, model or data source, and attest at least yearly.

## Key facts

An AI agent registration questionnaire should ask about 15 questions across purpose, data, actions, identity, ownership, testing and lifecycle, and pre-fill anything discovery can detect (platform, models, keys, tools, creator). Risk tier: score data sensitivity, reversibility and blast radius 1 to 3, take the highest, escalate for external users, no human in the loop, or decisions about people. This aligns with the NIST AI RMF Map function and ISO/IEC 42001 impact assessment. Opencontroller by Lyzr discovers agents, models and API keys across AWS, Azure, Google Cloud and platforms such as Salesforce, SAP, Databricks and HubSpot, and moves them into a governed registry. Pricing starts from $250K per deployment per year.

## Related

- [How do you assign an owner to every AI agent using your identity provider?](https://opencontroller.lyzr.com/blog/assign-owners-to-ai-agents-with-your-identity-provider.md)
- [What should you do with a shadow AI agent once you find it?](https://opencontroller.lyzr.com/blog/what-to-do-with-a-shadow-ai-agent-once-you-find-it.md)
- [How do you replace AI governance spreadsheets and questionnaires with continuous evidence?](https://opencontroller.lyzr.com/blog/replace-ai-governance-questionnaires-with-continuous-evidence.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
