---
title: Can you run an AI governance proof of concept on dummy data?
url: https://opencontroller.lyzr.com/blog/ai-governance-proof-of-concept-with-synthetic-data
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# Can you run an AI governance proof of concept on dummy data?

**Yes, but only for part of the evaluation. Dummy data can prove policy logic, approval workflows, dashboards and reports. It cannot prove the things that usually decide the purchase: whether discovery finds your real agents, whether traffic actually flows through the control point, and whether identities map to your people. Start with dummy data, then move to read-only connectors on real accounts as early as security allows.**

Regulated buyers, especially in insurance, banking and healthcare, often want a first proof of concept with no production data. That is reasonable. The risk is signing a contract on a demo that never touched the estate it has to govern.

## What can a dummy-data proof of concept prove?

| Capability | Proven on dummy data? | Why |
|---|---|---|
| Policy logic (rules fire on the right conditions) | Yes | Policies run the same on synthetic events |
| Approval and registration workflows | Yes | Workflow steps do not depend on real data |
| Framework mappings and reports | Yes | Report structure can be reviewed |
| Dashboards and usability | Yes | Admins can judge the console |
| Red-team and evaluation runs on a sample agent | Mostly | A realistic sample agent gives useful signal |
| Discovery coverage of your estate | No | Only real cloud accounts and platforms show what exists |
| Traffic paths and enforcement | No | Only real traffic shows what bypasses the control point |
| Identity mapping to employees | No | Needs your identity provider and real usernames |
| Cost attribution accuracy | No | Needs real bills and usage |
| Performance under your load | Partly | Synthetic load tests help; real traffic patterns differ |

The pattern: dummy data tests the product, real connectors test the fit.

## How should a staged proof of concept run?

Move one stage at a time, with security sign-off at each gate.

1. **Stage 0: synthetic sandbox.** Load sample agents, policies and events. Test workflows, policies and reports. Exit when policy owners agree the logic is right.
2. **Stage 1: read-only connector to a non-production account.** One cloud account or one agent platform tenant, with read-only permissions. Exit when discovery results match what the account owner expects.
3. **Stage 2: read-only discovery on production.** Same read-only permissions, wider scope. Exit when the inventory is reconciled against a known list, and the unknowns are explained.
4. **Stage 3: observe traffic for one agent.** Route one low-risk agent through the gateway in monitor mode. Exit when cost, latency and logs match expectations.
5. **Stage 4: enforce on one agent.** Turn on one blocking policy. Exit when it blocks what it should and nothing else.

Each stage adds evidence a buyer can show internally. Stage 2 often changes the conversation because shadow agents appear.

## What should the security team approve at each stage?

- **Permissions:** the exact read-only roles and scopes per connector, written down before access is granted.
- **Data handling:** what metadata the tool stores, where, and for how long. Discovery usually needs configuration metadata, not business records.
- **Prompt content:** whether prompts are logged at Stage 3, and whether redaction is on.
- **Deployment:** SaaS or your own infrastructure, and the network path.
- **Exit:** how data is deleted if the proof of concept ends.

## How long should each stage take?

Agree time boxes up front so a dummy-data stage does not run for months. Typical plans give Stage 0 one to two weeks and each later stage one to three weeks, depending on security reviews. Write success criteria for each stage before it starts, not after.

## How Opencontroller does this

Opencontroller works on a registry, observe, enforce progression: registering an agent is not the same as taking over its runtime, and enforcement depth depends on the connector and traffic path. It auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to agent platforms such as Salesforce, SAP, Databricks and HubSpot. It runs on Lyzr SaaS or customer-controlled infrastructure, and Lyzr engineers embed for integration. [Book a demo](https://www.lyzr.ai/book-demo/).

Read the pillar [AI agent governance platform evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist), then [how to run a pilot of an AI agent governance platform](/blog/how-to-run-a-pilot-of-an-ai-agent-governance-platform) and [questions to test AI agent discovery coverage](/blog/questions-to-test-ai-agent-discovery-coverage).

## Frequently asked questions

**Is read-only discovery on production risky?**
Read-only connectors typically read configuration metadata, not business records. Approve the exact roles and confirm what is stored before granting them.

**Can we skip straight to production traffic?**
You can, but staging lets security approve one step at a time, which is often faster overall than one large review.

**What is the most common reason a dummy-data proof of concept misleads?**
It shows policies working on events that were designed to trigger them, while real agents that bypass the control point stay invisible.

## Key facts

A dummy-data proof of concept for AI agent governance can prove policy logic, approval workflows, framework reports and usability. It cannot prove discovery coverage, real traffic paths and enforcement, identity mapping or cost attribution accuracy. Staged plan: Stage 0 synthetic sandbox; Stage 1 read-only connector to non-production; Stage 2 read-only production discovery; Stage 3 observe one agent in monitor mode; Stage 4 enforce on one agent. Opencontroller follows a registry, observe, enforce progression where depth depends on connector and traffic path; it runs on Lyzr SaaS or customer-controlled infrastructure with embedded Lyzr engineers. Pricing starts from $250K per deployment per year.

## Related

- [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 should acceptance testing cover for an AI agent governance platform?](https://opencontroller.lyzr.com/blog/acceptance-testing-for-ai-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)
