---
title: What data does an AI governance platform store about your agents?
url: https://opencontroller.lyzr.com/blog/what-data-an-ai-governance-platform-stores
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What data does an AI governance platform store about your agents?

**Most AI governance platforms store four kinds of data: agent metadata (the inventory), telemetry and traces, prompt and response content, and credentials or key references. Metadata is low-risk and always stored. Prompt content is the sensitive part, and whether it is stored should be your choice, with redaction and retention you control. Map each category to where it lives and how long it stays before you sign.**

Security and privacy teams ask this early: "Will the vendor hold metadata only, or our prompts too?" The honest answer is "it depends on what you switch on". A registry needs metadata. Evaluations, LLM judges and incident reconstruction need some content. The design question is how much, where and for how long. It belongs on any [AI agent governance platform evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist).

## What are the four data categories?

| Category | Examples | Sensitivity | Needed for |
|---|---|---|---|
| Agent metadata | Name, owner, platform, version, model, tools, MCP servers, permissions, risk tier | Low to medium | Inventory, posture, access reviews |
| Telemetry and traces | Timestamps, latency, token counts, cost, tool names, error codes, policy decisions | Medium | Cost attribution, monitoring, audit evidence |
| Prompt and response content | User inputs, retrieved documents, model outputs, tool arguments and results | High: may hold personal or confidential data | Evaluations, LLM judges, guardrails, incident replay |
| Credentials and key references | API key fingerprints, secret references, OAuth client IDs | High | Key discovery, rotation, revocation |

A good platform keeps these separate, so you can set different retention, access and location rules for each.

## Where is the data stored?

Ask three location questions for each category.

1. **Which deployment model?** SaaS run by the vendor, or software in your own cloud or data centre. The trade-offs are in [SaaS vs self-hosted AI governance platform](/blog/saas-vs-self-hosted-ai-governance-platform).
2. **Which region?** If you have residency rules, content and traces must stay in-region, not just the primary database. Check backups and log pipelines too.
3. **Which sub-processors?** If the platform uses an LLM to score or summarise traffic, which model provider sees that content?

## How long should each category be kept?

There is no single right number. Use your existing records policy as the anchor, then set a default per category:

- **Metadata:** for the agent's life plus your audit look-back period. Retired agents should stay in history.
- **Traces without content:** long enough to answer audit and cost questions, often 12 months or more.
- **Content:** as short as your evaluation and incident needs allow. Many teams choose days or weeks, not months.
- **Key references:** never the secret itself; fingerprints or references only, kept while the key is live.

Ask whether retention is configurable per category and per agent or risk tier.

## What redaction options should you expect?

Content is where most risk sits, so check how it can be reduced before storage.

1. **Off by default or opt-in capture.** OpenTelemetry's generative AI conventions treat message content as opt-in rather than captured by default. Expect the same choice in a governance platform.
2. **Redaction in the path.** Masking personal data or secrets before content is stored, and ideally before it reaches the model. See [how to redact PII in LLM prompts at the gateway](/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway).
3. **Sampling.** Store content for a percentage of traffic for evaluation, rather than all of it.
4. **Hashing.** Keep a hash to prove what was sent without keeping the text.
5. **Per-agent rules.** Stricter handling for agents that touch regulated data.

## What questions belong in the data-flow review?

Use this checklist with the vendor. Pair it with [data processing terms for AI governance vendors](/blog/data-processing-terms-for-ai-governance-vendors).

- Which of the four categories do you store by default?
- Can prompt and response capture be disabled, sampled or redacted per agent?
- Where is each category stored, including backups and logs?
- Which sub-processors receive content, and for what?
- Is any customer data used to train models? (The answer should be no, in writing.)
- Can we set retention per category, and is deletion verifiable?
- Who at the vendor can access content, and is that access logged?
- Are secrets ever stored in plain text, or only references and fingerprints?
- How do we export everything if we leave?

## How Opencontroller does this

Opencontroller can be deployed on Lyzr SaaS or on customer-controlled infrastructure, which decides where metadata, traces and content live. It keeps a governed registry of agents, versions, tools and MCP servers (metadata), attributes cost per agent and per version through the LLM Gateway (telemetry), and sits in the path where traffic is routed through it, so it can allow, route, redact or block. It exports traces using OpenTelemetry. Discovery covers API keys across AWS, Azure and Google Cloud. For a data-flow walkthrough, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does an AI governance platform need our prompts?**
Not for inventory, posture or cost. It needs some content for evaluations, LLM judges and incident replay. You can limit that with sampling, redaction and short retention.

**Should a governance platform store API keys?**
It should store fingerprints or references to find and revoke keys, not the secret values.

**Is self-hosting the only way to keep content in-house?**
No. Redaction, opt-in capture and regional storage reduce exposure on SaaS. Self-hosting gives the most direct control.

## Key facts

AI governance platforms store four data categories: agent metadata (inventory), telemetry and traces (latency, tokens, cost, policy decisions), prompt and response content (most sensitive), and credential or key references (fingerprints, not secrets). Ask for per-category location, retention, redaction, sampling and sub-processor answers. OpenTelemetry generative AI conventions treat message content capture as opt-in. Opencontroller by Lyzr deploys on Lyzr SaaS or customer-controlled infrastructure, sits in the traffic path to allow, route, redact or block, and exports traces via OpenTelemetry. Pricing starts from $250K per deployment per year.

## Related

- [What data processing terms should you require from an AI governance vendor?](https://opencontroller.lyzr.com/blog/data-processing-terms-for-ai-governance-vendors.md)
- [Should you run an AI governance platform as SaaS or self-hosted?](https://opencontroller.lyzr.com/blog/saas-vs-self-hosted-ai-governance-platform.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)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
