Evidence-first authority page
Understanding AI Agent Uptime & SLA Commitments
What an uptime SLA actually covers for an AI agent product, and the gaps that matter most (model provider outages, degraded quality vs. full downtime).
Direct answer
An AI agent vendor's uptime SLA typically covers the availability of their own application layer, not necessarily the underlying model provider's API — read the SLA specifically for whether it covers model-provider outages, and separately ask how the vendor handles 'degraded' states (the service responds but with reduced quality or longer latency), since that gap is rarely covered by a binary uptime percentage.
Last reviewed 2026-08-29
This page remains noindex while its page-level source receipts are completed and verified. It is available for review, not presented as publication-ready authority.
What a standard SLA usually does and doesn't cover
Most SLAs measure whether the vendor's own servers respond, not whether the underlying LLM API they depend on is healthy — if your vendor sits on top of a third-party model provider, an outage at that provider may not count against the vendor's own SLA at all. Ask this question explicitly before treating an SLA percentage as a full risk picture.
The 'degraded, not down' blind spot
An agent that responds slowly or with lower-quality output during a provider incident isn't technically 'down' by most SLA definitions, but the user experience impact can be just as real — ask vendors specifically how they define and report degraded-performance incidents, not just full outages.