For any other service you run, you can say what it is doing, what it costs and who owns it. For most AI in production, nobody can. Not because AI is special, but because it arrived through a pilot, on a credit card, with an API key, and skipped the golden path and the governance that every other workload gets by default.
This talk makes the case that AI is a workload, and that we have already solved every one of its problems once: shared bills nobody could attribute, dependencies that changed under us, services nobody owned. The disciplines are old. The speed is new.
Event
Observability & SRE Summit 2026, co-located with the Platform Engineering & Developer Experience Summit, runs 21 to 22 October 2026 at Hilton London Syon Park. This session is in Track A, room SYON 2, on Thursday 22 October, 15:00 to 15:30. Full agenda: iqpc.com/events-observability-sre-summit.
The core argument
A model provider swapping the model behind the same name is a third-party API changing under you. One API key with one token bill is the shared cloud account we untangled with tagging and showback. A confident, plausible, wrong answer is an HTTP 200 with the wrong content. An agent nobody owns is a microservice nobody owns, except this one holds real credentials. Same problems, same discipline, less time to apply it.
What the talk covers
See it. Fast, available and wrong is still an outage. Keep the golden signals, add tokens and saturation, and split errors in two: the system errors you already alert on, and the quality errors that matter most. Score outcomes on a sample of live traffic, per model version and per agent version, on the same OpenTelemetry traces as everything else.
Cost it. Cloud FinOps, on fast-forward. Tag every model call, show every team its own spend, and manage cost per successful task rather than price per token: sometimes the small model wins, sometimes the large one does the job in one shot and costs less, and only your evaluations can tell you which. Put a gateway in front of the providers as the allocation layer AI billing doesn’t give you.
Own it. You can’t own what you don’t understand. Own the agent, not the prompt: catalogue it, version it as one unit, re-test on every model change you didn’t choose, review outcomes rather than changes, and give it least privilege. Treat your provider as a dependency, and buy only as much model-agnosticism as the risk is worth.
Where it is genuinely new. Correctness becomes a distribution, cost becomes a behaviour, change arrives without a deploy, and agents act instead of only answering. Each one needs something this audience already builds: a signal, a grain, a detector, an identity.
Who this is for
SREs, observability and platform leads who get paged when AI features misbehave, and the engineering leaders who have to say what those features cost and who owns them.
Sources
Every claim in the talk links to its source on the slides. The main ones:
- OpenAI, OpenAI – Hugging Face incident technical report (August 2026)
- LangChain, State of Agent Engineering
- FinOps Foundation, State of FinOps 2026 and Token Economics
- OpenTelemetry, GenAI semantic conventions
- Google, Monitoring Distributed Systems (SRE book) and the 2025 DORA report
- OLG Hamm, press release on the chatbot judgment, 12 May 2026 (German)