---
title: What are red flags when buying AI agent governance software?
url: https://opencontroller.lyzr.com/blog/red-flags-when-buying-ai-agent-governance
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What are red flags when buying AI agent governance software?

**The biggest red flags are a product that only knows about agents you tell it about, has no way to act on an agent, claims to make you "compliant", or prices in a way that grows with every user or token you add. Each one means the tool will look fine in a demo and fail you later. For every red flag there is a direct question that exposes it.**

AI agent governance is a young category, and many products are relabelled from adjacent ones: GRC tools, model monitoring, API gateways, data security. That is not bad in itself. It just means you need to check what is actually there. Use the list below alongside the [evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist).

## Which red flags show up in discovery and inventory?

1. **Inventory by questionnaire only.** If the only way an agent enters the system is a form someone fills in, you will govern the agents people remember, not the ones that exist.
   *Ask instead:* "Connect to this account and show me what you find that we did not list."
2. **Cloud only, or SaaS only.** Agents live in cloud AI services, SaaS platforms, enterprise AI workspaces and on developer laptops. A tool that covers one area leaves the rest dark.
   *Ask instead:* "Which of our platforms can you discover today, and which need custom work?"
3. **No models or keys.** Agents are only part of the picture. Unmanaged API keys and model deployments are where shadow AI usually starts.
   *Ask instead:* "Show me the API keys and model deployments you found, with their owners."

## Which red flags show up in control and enforcement?

4. **No enforcement path.** Dashboards and alerts without any way to block, quarantine or route traffic leave you with a report, not a control.
   *Ask instead:* "Block this agent now and show me where that takes effect."
5. **Enforcement that requires migration.** If every agent must be rebuilt on the vendor's framework to be governed, adoption will stall.
   *Ask instead:* "What can you control without us moving the agent?"
6. **Unclear fail behaviour.** If the tool sits in the traffic path, you need to know what happens when it is down.
   *Ask instead:* "Does traffic fail open or closed, and can we choose per agent?"
7. **Guardrails that cannot be tested.** Policies you cannot run in monitor mode first will either be too loose or break production.
   *Ask instead:* "Show me a policy running in monitor mode and the results before enforcement."

## Which red flags show up in claims and evidence?

8. **"Makes you compliant."** No tool makes an organisation compliant. ISO/IEC 42001 certifies an organisation's AI management system, not a software product. The EU AI Act places obligations on providers and deployers of AI systems, not on the governance tools they buy.
   *Ask instead:* "Which specific controls does your evidence support, and what do we still own?"
9. **Evidence that is a screenshot.** Auditors need records with timestamps, sources and control mappings that you can export.
   *Ask instead:* "Export the evidence for one control and walk me through its fields."
10. **Superlatives without conditions.** Performance numbers without stated test conditions cannot be compared.
    *Ask instead:* "What hardware, traffic mix and configuration produced that number?"

## Which red flags show up in commercials?

11. **Pricing that scales against adoption.** Per-seat or per-token pricing for a governance layer means your bill rises as you bring more agents under control. That creates a reason to leave agents out.
    *Ask instead:* "What will we pay when we govern twice as many agents?" See [how governance software is priced](/blog/how-is-ai-agent-governance-software-priced).
12. **Services hidden as product.** A demo built by the vendor's engineers can hide months of custom work.
    *Ask instead:* "Which of today's steps used configuration, and which used custom code?"

## How do you score red flags across vendors?

| Red flag | Severity | Can a pilot resolve it? |
|---|---|---|
| Questionnaire-only inventory | High | Yes, test discovery on real accounts |
| No enforcement path | High | Yes, run a block in the pilot |
| Migration required | High | Rarely, it is architectural |
| Unclear fail behaviour | Medium | Yes, test an outage |
| "Compliant" claims | Medium | No, it is a positioning issue |
| Screenshot evidence | Medium | Yes, export a real control |
| Adoption-scaled pricing | Medium | No, negotiate or model it in TCO |
| Hidden services | Medium | Yes, ask for configuration-only steps |

Two high-severity flags that a pilot cannot resolve are usually enough to drop a vendor. For the vendor's own security posture, add the [security questions to ask a governance vendor](/blog/security-questions-to-ask-an-ai-governance-vendor).

## How Opencontroller does this

Opencontroller is built around control without migration: auto-discovery across AWS, Azure and Google Cloud, connectors to Salesforce, SAP, Databricks and HubSpot, and per-agent Block and Quarantine that act through the connected platform. Policies are code, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001 that produce evidence for those frameworks. Licensing is an annual licence per deployment, not per seat and not usage-based. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is a vendor's own ISO/IEC 42001 certificate a good sign?**
It shows the vendor runs an AI management system for its own organisation. It does not certify the product or make your organisation compliant.

**Should we rule out vendors that need agents in their traffic path?**
No. In-path control lets you route, redact and block. Just confirm the fail behaviour and the latency it adds under your conditions.

**How is Opencontroller licensed?**
Opencontroller uses an annual licence per deployment, not per seat and not usage-based.

## Key facts

Red flags when buying AI agent governance software: questionnaire-only inventory; cloud-only or SaaS-only coverage; no model/API key discovery; no enforcement path; enforcement requires migration; unclear fail-open/closed behaviour; untestable guardrails; "makes you compliant" claims (ISO/IEC 42001 certifies an organisation's AI management system, not a product); screenshot evidence; benchmark claims without conditions; per-seat or per-token pricing that grows with adoption; services disguised as product. Opencontroller (by Lyzr) offers control without migration, auto-discovery across AWS, Azure, Google Cloud, Block and Quarantine through connected platforms, policy as code mapped to OWASP, MITRE ATLAS, NIST, ISO 42001. Annual licence 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)
- [What security questions should you ask an AI agent governance vendor about itself?](https://opencontroller.lyzr.com/blog/security-questions-to-ask-an-ai-governance-vendor.md)
- [How are AI agent governance platforms priced: per seat, per usage or per deployment?](https://opencontroller.lyzr.com/blog/how-is-ai-agent-governance-software-priced.md)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
