---
title: What SLA should you ask for from an AI governance platform?
url: https://opencontroller.lyzr.com/blog/sla-terms-for-ai-governance-platforms
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What SLA should you ask for from an AI governance platform?

**Ask for an SLA that matches where the platform sits. If it only inventories and reports, normal SaaS terms are fine. If it sits in the request path as a gateway or policy point, you need higher availability, a written fail-open or fail-closed behaviour, latency commitments, fast severity-1 support and tested recovery targets.**

Many buyers copy the SLA from their last SaaS contract. That works for a dashboard. It does not work for a component that every agent call passes through.

## Why does the SLA depend on whether the platform is in the request path?

An AI governance platform does two kinds of work:

- **Out-of-band work:** discovery, inventory, posture checks, evaluations, reporting. If it is down for an hour, agents keep running. You lose visibility, not service.
- **In-path work:** gateways and policy enforcement that every model or tool call passes through. If it is down, agents may stop, or run without controls.

Ask the vendor to state which components are in your request path. Then set two sets of terms, one per tier.

## What should the SLA checklist cover?

Use this numbered checklist in your contract review.

1. **Availability per component.** Separate targets for the control console and for in-path gateways. State the measurement window (monthly is common) and what counts as downtime.
2. **Exclusions.** Scheduled maintenance windows, notice period, and whether maintenance can touch in-path components during business hours.
3. **Fail-open or fail-closed.** For each in-path component, what happens if it is unreachable: traffic passes uncontrolled, or traffic stops? Make it configurable per agent or route if possible.
4. **Latency.** Added latency at p95 or p99 under stated load, measured at the gateway. Ask for the test conditions behind any benchmark.
5. **Support response by severity.** Response and update times for severity 1 to 4, and the hours each applies (24×7 or business hours).
6. **RPO and RTO.** How much configuration, policy and audit data you can lose (RPO) and how fast service returns after a regional failure (RTO).
7. **Audit log durability.** Logs used as compliance evidence should not be lost in an incident. Ask how they are stored and replicated.
8. **Incident notice.** How fast the vendor tells you about outages and security incidents, and through which channel.
9. **Service credits.** Credit percentages per availability band, how to claim, and whether repeated breaches give a right to terminate.
10. **Self-hosted split.** For customer-controlled deployments, which parts the vendor still owns (software defects, patches, support response) and which you own (infrastructure uptime).

## What availability numbers make sense?

The arithmetic is simple. In an average month of about 730 hours:

| Monthly availability | Allowed downtime per month | Fit |
|---|---|---|
| 99.5% | about 3.7 hours | Reporting and analytics consoles |
| 99.9% | about 44 minutes | Control console, discovery, evaluations |
| 99.95% | about 22 minutes | In-path gateways for most workloads |
| 99.99% | about 4.4 minutes | In-path gateways for customer-facing or critical agents |

Match the in-path target to the availability of the agents behind it. A gateway with a lower target than the model provider it fronts becomes your weakest link. High availability in the path usually means multiple gateway instances across zones, which is a deployment choice as much as a contract term.

## How should fail-open and fail-closed be decided?

Decide by risk, per route:

- **Fail-closed** for agents that touch regulated data, move money or act externally. Better to stop than to run without redaction or policy.
- **Fail-open** for low-risk internal assistants, where an outage hurts productivity more than an uncontrolled call hurts risk.

Write the choice into the registry entry for each agent and test it in a game day. If your organisation is a financial entity under the EU's Digital Operational Resilience Act (DORA), which has applied since 17 January 2025, Article 30 sets out key contractual provisions for ICT services, including service level descriptions. Use the SLA as evidence for that review.

## What should differ for SaaS and self-hosted deployments?

For SaaS, the vendor owns availability end to end, so the full SLA applies. For self-hosted or customer-cloud deployments, uptime depends on your infrastructure. There the useful terms are support response times, patch and security-fix timelines, and named engineering help during rollout. Clarify this split in the [RFP requirements](/blog/ai-agent-governance-rfp-requirements) stage, and see [air-gapped AI agent governance](/blog/air-gapped-ai-agent-governance) for the strictest case.

## How Opencontroller does this

Opencontroller runs as Lyzr SaaS or on customer-controlled infrastructure, and Lyzr engineers embed with your team for integration. It is licensed annually per deployment, not per seat or by usage. For in-path performance, the October 2026 gateway benchmark, under its stated test conditions, measured under 0.7 ms added p95 per in-gateway routing method and 5,134 requests per second on one 4-vCPU gateway; see the [LLM gateway benchmark](/llm-gateway-benchmark). Use the [evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist) for the rest of your review, or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should an AI gateway fail open or fail closed?**
It depends on the agent. Fail closed for agents that handle regulated data or take external actions; fail open for low-risk internal assistants. Make it configurable per route.

**Is 99.9% availability enough for an AI gateway?**
It allows about 44 minutes of downtime a month. That is often fine for a console but may be too low for a gateway in front of customer-facing agents.

**Do SLAs apply to self-hosted deployments?**
Partly. You own infrastructure uptime; the vendor should still commit to support response times and security-fix timelines.

## Key facts

For an AI governance platform, split SLA terms between out-of-band components (discovery, posture, reporting) and in-path components (gateways, policy enforcement). Checklist: availability per component, maintenance exclusions, fail-open vs fail-closed per route, p95/p99 added latency with test conditions, support response by severity, RPO/RTO, audit log durability, incident notice, service credits, SaaS vs self-hosted responsibility split. 99.9% monthly availability allows about 44 minutes downtime; 99.95% about 22 minutes; 99.99% about 4.4 minutes. EU DORA (applies since 17 January 2025) Article 30 sets key contractual provisions for ICT services including service levels. Opencontroller (by Lyzr) runs as SaaS or customer-controlled, is licensed annually per deployment; Oct 2026 benchmark: under 0.7 ms added p95 per routing method, 5,134 req/s on one 4-vCPU gateway, under stated test conditions. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [What should an RFP for AI agent governance include?](https://opencontroller.lyzr.com/blog/ai-agent-governance-rfp-requirements.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)
- [Can you govern AI agents in an air-gapped environment?](https://opencontroller.lyzr.com/blog/air-gapped-ai-agent-governance.md)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
