The Lineup AI Operations

Running The Lineup alone means a lot of recurring checks. Are the provider feeds stale? Did the model jobs run and produce forecasts? Does growth work need attention? Is a business commitment slipping? I was answering those questions by moving between scripts, dashboards and notes. It worked, but it depended on me remembering to look in each place, and on nothing slipping between them.

Two things made simple automation risky. First, a finished automated check does not prove the underlying problem is fixed. A clean run can sit on top of an issue that is still there. Second, some of the obvious next steps, such as deploying a change, suppressing a model or contacting a customer, affect customers or production. Those steps need to stop for a decision, not happen because a script judged them a good idea. I wanted the checks to happen whether or not I remembered them, with the results in one place, and without giving any automated worker the power to act on production by itself.

I built four internal AI workers for the checks I kept doing by hand. One watches sports data and prices. One watches model jobs and forecasts. One watches customer growth. One reviews business priorities. Each runs on a schedule, or when I ask for a specific check.

They share one request and review system. When a worker finds a material issue, it records the finding, its current status and one proposed next action in a single review queue. It proposes one step, not a list of ideas, and it stops if that step could alter production. The review dashboard shows what each worker found, what it recommends and what is waiting on me. A single queue means I start each review from the same place, whichever worker raised the issue.

The system also keeps a check separate from a fix. A completed check is not treated as a completed fix: an issue is only marked complete when newer evidence supports it, so a problem that looks handled but isn't stays visible. Approval rules, reliability checks and automated tests keep incomplete or blocked work in view instead of letting it drop out of the queue.

The result is that the recurring checks now happen on a schedule, issues live in one review queue instead of scattered across tools, and production changes cannot happen without explicit approval. It is the same pattern I use for client work, applied to my own business first: the system inspects, drafts and recommends, and a named person approves anything with consequences.