Pick an indicator, a target and a window: 99.9% of requests succeeding over 30 days, or p95 latency under 800ms. The failure the target allows is the error budget: at 99.9% of requests, 0.1% may fail, for example 1,000 of 1,000,000 in the window; for a time-based target of 99.9% over 30 days, it is about 43 minutes of full outage. Publishing the SLO tells users what to expect. A service level agreement (SLA) is a contract that sets consequences for missing the SLOs it contains, and the target you run against internally should be tighter than anything you promise, so you hear about trouble before you breach it (Google SRE book, Service Level Objectives).
Page on burn rate, not on single errors: a burn of 14.4x for an hour spends 2% of a 30-day budget and is worth waking someone for (Google SRE workbook, Alerting on SLOs). In a customer deployment, dependencies you do not run set a ceiling: 3 in series at 99.9% each allow at most about 99.7%, so either exclude them in writing or set the SLO below that. For a model-backed feature, add a quality indicator measured on a sample of responses, such as the share a grader or reviewer marks correct, because a wrong answer is still HTTP 200.
OpenAI’s Financial Services posting in New York, as of September 2026, asks the FDE to own reliability, observability and on-call readiness for the production AI systems it ships. Source 1Forward Deployed Engineer (FDE), Financial Services- NYCPublisherOpenAI (Ashby)Source typecompany job posting In behavioral questions, an SLO gives a story its measure: in an incident story, name the SLO that was breached and how much of the budget the incident burned; in a story about pushing back on sales, a spent error budget is the objective reason to pause a feature for reliability work (Google SRE workbook, error budget policy). Say who agreed the SLO with the customer, because a target nobody signed up for settles no argument.