Lights-Out is an operating model, not a feature
Every large enterprise now has AI. Very few have changed how they operate. The pattern is consistent: a copilot here, a chatbot there, a forecasting model in a corner of FP&A — each one genuinely useful, none of them changing who runs the process. The work is still owned by people; the AI assists. Adoption goes up, the operating model stays exactly where it was — and eventually someone asks why the P&L never showed it.
The answer is that assistance and autonomy are different economic propositions. An assistant makes a person ten or twenty percent faster at a task. Autonomy removes the person from the task entirely and gives them a different job: judgment, exceptions, and control. The first compounds slowly and erodes as the novelty fades. The second changes the cost structure of the function.
Lights-Out is not "more AI." It is a decision about who runs the process — and it has to be made deliberately, because no copilot rollout ever drifts into it.
What Lights-Out actually means
The term comes from manufacturing: a factory that runs with the lights off because no one needs to see. Applied to enterprise functions, it means processes that run themselves end to end — overnight, at quarter-end, through a Day-1 separation — and route only exceptions to people. The function doesn't go dark; the people move up. They stop executing and start governing.
That distinction is why Lights-Out is an operating model question rather than a technology one. It forces decisions no pilot ever forces: which decisions agents may take alone, what the guardrails are, who owns the exception queue, what the auditor sees, and what the team does once the work they used to do runs without them. Enterprises that skip those decisions don't get autonomy — they get very expensive assistance.
Why the plateau happens
Copilots are easy to buy and easy to defend. They require no process redesign, no control framework, no answer to the question "what happens when it's wrong at 2 a.m." Autonomy requires all three. So organizations rationally stop at the assistance layer — and then mistake the plateau for the ceiling.
Crossing it takes three things copilots never demand. First, an agentic layer: goal-driven systems that plan, decide, and act across enterprise tools, not just draft text for a human to paste. Second, real operations — monitoring, evaluation, cost management — because an autonomous process that degrades silently is worse than a manual one. Third, a control plane auditors and regulators accept, because in regulated industries the constraint on autonomy is never the model; it is the attestation.
Where to start
Not with the most impressive use case — with the process where the economics of autonomy are sharpest: high volume, rule-dense, deadline-driven, already measured. In most enterprises that is the Office of the CFO, which is why it is where we built deepest first. Reconciliation, invoice processing, and close orchestration have proven they can run touchless at global scale; the pattern then generalizes to risk, operations, and revenue.
The honest test of where you stand: if your team stopped touching a core process for forty-eight hours, would it run? If the answer is no, you don't have an AI capability problem. You have an operating model decision you haven't made yet.