---
title: How do you run a shadow AI amnesty for unapproved agents?
url: https://opencontroller.lyzr.com/blog/run-a-shadow-ai-amnesty
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How do you run a shadow AI amnesty for unapproved agents?

**Announce a fixed window, usually four to eight weeks, in which teams can declare any AI agent, app or API key they built or bought without approval, with no penalty and a fast review. Keep the form short, triage what comes in, and publish what happens after the deadline. Back it with automated discovery so you can tell what was declared from what exists.**

Most shadow AI is not malicious. Teams built something useful because the approved path was slow or unclear. If the first contact from security is a shutdown notice, the next agent will be hidden better. An amnesty turns discovery into a deal: tell us what you have, and we will help you keep it safely.

## Why run an amnesty instead of just scanning?

Scanning finds infrastructure. It does not find purpose, owner or business value. An amnesty gets teams to tell you:

- What the agent is for and who relies on it.
- What data it touches, in their own words.
- Which tools and accounts it uses that a scan might miss, such as personal SaaS accounts.

Scanning is still essential. It is how you check the amnesty was honest. See [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents) for the discovery side.

## How do you run it, step by step?

1. **Get a sponsor.** A CIO, CISO or COO announces the amnesty. It must be clear that declaring is safe.
2. **Set the window.** Four to eight weeks, with a firm end date. Avoid quarter-end and holiday periods.
3. **Define scope.** Agents, assistants, automations with AI steps, model API keys, AI SaaS subscriptions and coding agents on laptops.
4. **Publish the form.** Ten fields or fewer (see below). Make it reachable from the intranet and chat.
5. **Publish review service levels.** For example: acknowledgement within 2 working days, decision within 10.
6. **Run discovery in parallel.** Scan cloud accounts, agent platforms, expense data and devices.
7. **Triage weekly.** Sort declarations into tiers and send decisions.
8. **Chase the gaps.** Contact owners of discovered but undeclared assets before the deadline. Many simply did not see the announcement.
9. **Close and enforce.** After the deadline, undeclared assets follow the normal path, which may include blocking.
10. **Report.** Share counts, decisions and lessons with leadership and the teams.

## What should the registration form ask?

Keep it short. Every extra field cuts responses.

| Field | Why it matters |
|---|---|
| Agent or tool name | Identification |
| Owner and backup | Accountability |
| What it does, in one sentence | Purpose and risk |
| Platform or model provider | Supply chain |
| Data it reads or writes | Data classification |
| Actions it can take | Write tools raise risk |
| Who uses it, roughly how many | Business impact |
| How it authenticates | Personal keys, shared accounts or proper identity |
| Cost and who pays | Finance and procurement |
| Anything you are worried about | Often the most useful field |

## How do you triage what comes in?

Use four outcomes, and say which applies within the published service level:

1. **Approve.** Low risk, proper identity, approved provider. Register it and move on.
2. **Approve with conditions.** Useful but needs fixes: move from a personal key to a workload identity, add logging, narrow scopes. Give a deadline.
3. **Migrate.** The use case is fine but the tool is not. Move it to an approved model, platform or gateway.
4. **Retire.** No owner, no users, or a risk you will not accept. Retire it with the owner's involvement.

Most declarations should land in the first three. If you retire more than a small share, teams will hear that declaring is risky. Our guide on [what to do with a shadow AI agent once you find it](/blog/what-to-do-with-a-shadow-ai-agent-once-you-find-it) covers each outcome in more detail.

## What happens after the deadline?

Be explicit before the amnesty starts. A common rule: anything discovered after the deadline is suspended first and reviewed second. That is not punishment of people; it is the normal control path. Keep the registration form open permanently as the standard intake, so the amnesty becomes a habit rather than a one-off. Agents hosted on third-party platforms can still be registered; see [registering AI agents you don't host](/blog/register-ai-agents-you-dont-host).

## Which frameworks expect an inventory?

NIST AI RMF GOVERN 1.6 calls for mechanisms to inventory AI systems, resourced by risk. ISO/IEC 42001 expects you to know the AI systems in scope of your management system. An amnesty is one of the fastest ways to build that first inventory.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and through connectors to Salesforce, SAP, Databricks, HubSpot and Claude Enterprise and OpenAI Enterprise workspaces. Declarations and discoveries land in one governed registry, so gaps are easy to spot. After the deadline, per-agent Block and Quarantine controls act through the connected platform. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How long should a shadow AI amnesty last?**
Four to eight weeks is common. Shorter windows miss people; longer ones lose urgency.

**Should declared shadow AI agents be shut down?**
Rarely. Most should be approved with conditions or migrated. Retire only those with no owner, no use or unacceptable risk.

**How do you find agents that were not declared during the amnesty?**
Run automated discovery and compare the results with declarations. Undeclared agents follow normal enforcement after the deadline.

## Key facts

A shadow AI amnesty is a time-boxed programme (typically four to eight weeks) in which teams declare unapproved AI agents, apps and API keys without penalty, in exchange for a fast review. It works best with a short registration form (ten fields or fewer), triage into tiers (approve, approve with conditions, migrate, retire), published service levels for review, and enforcement after the deadline backed by automated discovery. NIST AI RMF GOVERN 1.6 calls for mechanisms to inventory AI systems. Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and connected platforms, which lets an amnesty be checked against what actually exists, and provides per-agent Block and Quarantine controls for enforcement. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [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 register an AI agent your team doesn't host?](https://opencontroller.lyzr.com/blog/register-ai-agents-you-dont-host.md)
- [How do you find shadow AI subscriptions in expense reports?](https://opencontroller.lyzr.com/blog/find-ai-subscriptions-in-expense-reports.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
