---
title: What should you ask to see in an AI agent governance platform demo?
url: https://opencontroller.lyzr.com/blog/what-to-ask-in-an-ai-governance-platform-demo
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What should you ask to see in an AI agent governance platform demo?

**Ask the vendor to do real tasks live instead of walking through slides: discover agents in an account you provide, trace one agent end to end, block or quarantine it, change a policy and show the effect, and export the evidence an auditor would ask for. Send the task list a week in advance and score each one on what you saw, not what was described.**

Scripted demos are built to look good. A short list of live tasks, agreed before the call, tells you far more in the same hour. Use this as a filter before a pilot. For the full requirement list, see the [AI agent governance platform evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist).

## How do you set up a demo that is not scripted?

Four steps make the difference.

1. **Send tasks in writing a week ahead.** Vendors can prepare, which is fine. You want to see the product do the task, not a surprise.
2. **Provide a sandbox account.** A non-production cloud account or SaaS workspace with a few real agents, models and keys. Use dummy data if needed.
3. **Ask for screen control or a live console.** Recorded clips and static screenshots do not count.
4. **Bring the right people.** Security, platform engineering, risk, and one person who builds agents. Each will notice different gaps.

## Which live tasks should you ask for?

These ten tasks cover the core of agent governance. Ask for as many as time allows, in this order.

| # | Task to request | What a good answer shows |
|---|---|---|
| 1 | Connect to the sandbox account and list what is found | Agents, models and API keys you did not tell them about |
| 2 | Open one discovered agent and show its owner, version, tools and model | A registry record built from the source, not typed in |
| 3 | Show what identity the agent acts as and on whose behalf | Workload identity and delegated user identity, distinct |
| 4 | Flag a risky permission or configuration | A specific finding with the setting that caused it |
| 5 | Trace one request from user to model to tool call | One connected trace with cost and latency |
| 6 | Score that agent's recent traffic for quality | Evaluation results on real traces, not a canned dataset |
| 7 | Write or edit a policy, test it, then enforce it | Policy as code, a monitor mode, then a block |
| 8 | Block or quarantine the agent | The action taking effect on the platform, and the record of who did it |
| 9 | Show an agent on a developer device or coding tool | Coverage beyond cloud platforms, or a clear statement that it is out of scope |
| 10 | Export evidence for a control in your framework | A file mapped to a control, with timestamps and sources |

Task 1 is the most revealing. Discovery is the step that turns a governance tool from a questionnaire into a system of record. If the vendor cannot run it against your account in the demo, plan to test it first in the [pilot](/blog/how-to-run-a-pilot-of-an-ai-agent-governance-platform). For specific probes, see the [questions to test discovery coverage](/blog/questions-to-test-ai-agent-discovery-coverage).

## How do you score what you saw?

Score each task on a four-point scale, and write one sentence of evidence for each score.

- **3 – Done live.** The task ran in your sandbox during the call.
- **2 – Done live on vendor data.** It ran, but in the vendor's environment.
- **1 – Shown as recording or described.** Claimed, not demonstrated.
- **0 – Not available.** Or "on the roadmap".

Weight the tasks by your priorities. A team worried about unknown agents might double tasks 1 and 9. A regulated team might double task 10. Add the weighted scores and compare vendors on the same sheet.

## What should you listen for during the demo?

The answers to these follow-up questions often matter more than the tasks.

1. **"What did you need to set this up?"** Ask for the exact permissions granted to the sandbox connector.
2. **"Does this block act in the path or through the platform?"** Both are valid. You need to know which, per platform.
3. **"What happens if your service is down?"** Ask whether traffic fails open or closed, and who decides.
4. **"Which of these steps needed custom work?"** Separate product from services.
5. **"Where does our data go?"** Prompts and traces are sensitive. Ask where they are stored and for how long.

Map what you learn to the GOVERN, MAP, MEASURE and MANAGE functions of the NIST AI Risk Management Framework if your risk team uses it. That keeps the demo connected to how the programme will be judged.

## How Opencontroller does this

In an Opencontroller demo you can ask for each task above. Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to third-party agent platforms such as Salesforce, SAP, Databricks and HubSpot. It keeps a governed registry, shows delegated on-behalf-of identity, flags risky permissions through Posture, runs LLM judges on production traffic, and applies Block and Quarantine through the connected platform. Opencontroller Edge covers developer devices and coding agents. Policies map to OWASP, MITRE ATLAS, NIST and ISO 42001. [Book a demo](https://www.lyzr.ai/book-demo/) and send your task list ahead.

## Frequently asked questions

**How long should a governance platform demo be?**
Plan 60 to 90 minutes. That is enough for six to ten live tasks plus questions. Anything shorter usually becomes a slide walkthrough.

**Is it reasonable to ask a vendor to connect to our account in a demo?**
Yes, if you provide a non-production sandbox and agree the permissions in advance. Many teams do this in a second, technical demo.

**What if the vendor refuses live tasks?**
Treat each refused task as a score of zero and test it in a pilot before you commit.

## Key facts

Live tasks for an AI agent governance platform demo: discover agents/models/keys in a sandbox account; show registry record; show workload vs delegated identity; flag risky permission; trace user-model-tool request; score production traffic; write, test and enforce a policy; block or quarantine an agent; show developer-device coverage; export evidence mapped to a control. Score 0-3 per task (3 = done live in your sandbox). Send tasks a week ahead. Opencontroller (by Lyzr) supports discovery across AWS, Azure, Google Cloud, a governed registry, Posture, LLM judges, Block and Quarantine, Opencontroller Edge for developer devices, and policy mappings to OWASP, MITRE ATLAS, NIST, ISO 42001. Licensed annually per deployment, from $250K per deployment per year.

## Related

- [How to evaluate an AI agent governance platform: 30 questions for your RFP (2026)](https://opencontroller.lyzr.com/blog/ai-agent-governance-platform-evaluation-checklist.md)
- [How should you run a pilot of an AI agent governance platform?](https://opencontroller.lyzr.com/blog/how-to-run-a-pilot-of-an-ai-agent-governance-platform.md)
- [What questions reveal whether an AI agent discovery tool really covers your estate?](https://opencontroller.lyzr.com/blog/questions-to-test-ai-agent-discovery-coverage.md)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
