AI spend · Feature
AI experiments need an end date, an owner and a cost centre
AI pilots rarely announce the moment they stop being experiments. FinOps Foundation guidance published October 30, 2026, argues for giving them a defined lifecycle, owner and cost allocation. Recent GitHub data shows why the question is timely: agents now start a substantial share of pull requests.
By Nythrex team4 min read
A team starts with a small test: one model, a sandbox and a few API calls. Then someone adds an evaluation run, another model, a logging service and a retry loop. The work may still be called a pilot, but its costs now arrive through several systems and people, often without one owner or a scheduled decision to stop.
That drift matters as agents enter more engineering workflows. GitHub’s October 28 Octoverse report says more than 40% of pull requests on GitHub started with an agent’s first draft by September 2026, up from 17% in March. That is a measure of activity on GitHub, not proof that any team has improved delivery or saved money. The practical question is how to let teams learn while making experimental spend visible, bounded and temporary.
- Share of GitHub pull requests whose first draft was started with an agent by September 2026; GitHub Octoverse, October 2026.
- >40%
- Agent-started pull requests on GitHub in September 2026, according to GitHub Octoverse, October 2026.
- 110 million
- Publication date of the FinOps Foundation’s guidance on experimental AI allocation, governance and lifecycle management.
- October 30, 2026
An experiment can quietly become a service
The FinOps Foundation’s October 30 guidance describes a problem that a monthly invoice can hide: experimental AI may span model calls, compute, storage, data preparation, software and staff time. Its proposed starting definition is operational: an experiment has an end date, does not need a production service-level agreement and does not serve customers at scale. That gives teams a testable boundary for deciding whether work is still exploratory.
The same guidance recommends tagging experiments with an owner, experiment identifier, lifecycle stage, cost centre and end date, then agreeing how to handle costs that cannot be directly tagged. This is practitioner guidance, not a measured guarantee of savings. GitHub’s report supplies a separate reason to revisit controls: by September, agent-started pull requests were more likely to change multiple files than pull requests without detected agent involvement (about 87% versus 79%). GitHub notes that its agent signal detects only some agent activity. Neither statistic tells an individual team what its experiment costs or whether its output is useful.
A pilot without a named owner and an end date is a recurring budget decision nobody has made.
The budget is bigger than model calls
For engineering leaders, the risk is not just an unexpectedly large API bill. An experiment can also consume evaluation and review time, create storage and telemetry costs, leave idle infrastructure running, or duplicate a tool another team already buys. These costs may sit in separate budgets. If nobody connects them to the same hypothesis and owner, a promising demo can look inexpensive while the organization carries its full operating burden elsewhere.
The decision is not simply whether to approve an experiment. Decide in advance what evidence would justify continuing it, what budget and duration it receives, and what happens when either limit is reached. Keep a customer-facing trial distinct from an internal productivity test; their risks, success criteria and owners may differ. A fixed token estimate alone will not capture the full cost of building, testing and maintaining a capability.
Set the rules before the first run
A team lead can start with one active pilot this week. Write down its hypothesis, a measurable result that would support it, the accountable owner, the spending limit and the end date. Attach a consistent experiment tag or cost-centre label to provider usage and related infrastructure; record staff time and shared services separately where they cannot be attributed directly. Make clear who can approve an extension.
Next, decide what the limits do. Alerts are useful, but an alert that no one sees is not a hard cap. Where the provider and environment support it, set usage quotas or rate limits and agree whether a threshold should pause the workload or require human approval. Set an expiry for sandbox resources and API credentials, with an explicit renewal step rather than an indefinite default.
Review actual spend against the hypothesis on a regular schedule. At the end date, make one of three decisions: stop and clean up compute, storage, endpoints and keys; extend with a revised hypothesis and approved budget; or graduate the work into a production budget with an owner, operating expectations and ongoing cost tracking. A pilot that passes its demo but has no production owner has not yet passed a business decision.
Sources
Frequently asked questions
How should an engineering team define an AI experiment?
The FinOps Foundation’s October 2026 guidance suggests a defined end date, no production service-level agreement and no service to customers at scale. Teams can use those conditions as a shared boundary, with separate treatment for higher-stakes customer-facing trials.
What costs should an AI pilot budget include?
Include model and API usage, compute, storage, data preparation, software and observability, plus staff effort where practical. Track shared costs with an agreed allocation rule and avoid counting the same work twice.
What should happen when an AI experiment reaches its end date?
Stop and decommission it, approve a time-bounded extension with a revised hypothesis, or move it into production with an accountable owner and operating budget. Make the decision explicit so a pilot does not become an unmanaged recurring expense.
Read next
Stop guessing.
Start seeing.
No card and no subscription: paid plans have not launched yet.