The Lights-Out Enterprise
Every operating advantage of the last thirty years came from the same move: take work people did by hand, encode it, and run it at machine cost. ERP did it to record-keeping. Outsourcing did it to labor arbitrage. The move now available is larger than either, because for the first time the thing being encoded is not the transaction but the operator. Functions that required hundreds of people to run can now run themselves — exception-based, always on, at a fraction of the cost, with better evidence than the human process ever produced.
We call that end state Lights-Out, and this paper states our thesis plainly: Lights-Out is not a technology purchase. It is an operating model decision, and it is engineered — by designing the right Target Operating Model, encoding the right Data Model, and putting governed agents on the runbook. Enterprises that treat it as a procurement exercise will buy copilots and plateau. Enterprises that treat it as an operating model decision will change the cost structure of their functions.
1. The economics have inverted
For every dollar the enterprise spends on software, roughly six are spent on services and headcount doing the work around it. Software vendors have spent three decades fighting over the one; the six sat untouchable because work required people. Agentic systems break that constraint. The consequence is a repricing of the entire services economy: work that was sold by the hour is becoming work sold by the outcome — per invoice processed, per account reconciled, per control tested. Buyers should insist on it. If your AI vendor prices by the seat, you are buying a tool and keeping all the work. If they price by the outcome, they are taking the work — and carrying the risk.
If you sell the tool, you are racing the model. If you sell the work, every model improvement makes the work faster, cheaper, and harder to compete with.
2. Work divides into intelligence and judgment
The line that matters inside any function is not automated versus manual; it is intelligence versus judgment. Intelligence is rule-dense, verifiable, high-volume: matching, coding, testing, filing, reconciling. Judgment is what to accept, what to risk, what to sign. Agents should take the intelligence — all of it, at machine speed, under controls. People should keep the judgment, concentrated at a single surface: the exception queue.
The boundary is not fixed. Every human decision on an exception is data about how judgment is exercised, and disciplined organizations feed it back into agent policy. Today’s judgment becomes tomorrow’s intelligence. The practical instrument for managing that boundary is an autonomy dial — observe, assist, recommend, act with approval, lights-out — set per process, per entity, per materiality threshold, and moved only on evidence.
3. Lights-Out is engineered in four moves
No function drifts into autonomy. The enterprises that get there make four deliberate moves. First, design the Target Operating Model — processes, roles, thresholds, controls — as a runnable specification rather than a binder. Most transformation programs fail here: they document an operating model no system can execute. Second, encode the Data Model: entities, hierarchies, calendars, policies, control matrices — the semantic layer agents reason over. Agents pointed at ungoverned data produce ungoverned outcomes. Third, put agents on the runbook, wired into the systems of record, with maker-checker enforced in the runtime rather than in policy documents. Fourth, operate exception-based and ratchet: prove every action on an Evidence Ledger, and promote autonomy level by level as the evidence accumulates.
Lights-Out isn’t a model upgrade. It is the right Target Operating Model and the right Data Model, run by governed agents.
4. Evidence is the unlock, not the afterthought
The binding constraint on enterprise autonomy was never model capability; it is attestation. In regulated functions, the question that stops every autonomy program is not “can the agent do the work?” but “will the auditor accept it?” This is why governance must be structural. When every agent action lands on an immutable ledger — inputs, rationale, tools, approvals, outputs — assurance changes character: you test populations instead of samples, controls monitor continuously instead of annually, and the audit becomes an export instead of a project. Done properly, the autonomous process is easier to audit than the human one it replaced. That is the argument that wins the audit committee, and it cannot be retrofitted.
5. Sequence through the outsourced wedge
Where to start is an economics question, not an ambition question. Start with work the enterprise already sends outside — payables processing, reconciliations, shared-services intake, audit preparation. Three things are true of that work: the budget already exists, it is already bought as an outcome, and replacing a vendor is an easier decision than restructuring an organization. Land there, accumulate evidence, then expand toward the judgment-heavy, insourced work no outsourcer could ever take. The outsourced task is the wedge; the insourced work is the prize.
6. The organization moves up
The most common objection — “what do the people do?” — has an empirical answer: they move from execution to governance. The controller stops chasing reconciling items and starts owning the exception queue, the control thresholds, and the autonomy promotions. Judgment, formerly diluted across a thousand routine touches, concentrates where it compounds. Functions running this way report a paradox worth stating: fewer touches, more control. The touches were never the control; the evidence is.
The test
One question tells you where you stand: if your team stopped touching a core process for forty-eight hours, would it run? If not, you do not have an AI adoption problem. You have an operating model decision you have not yet made. The stack exists. The evidence pattern is proven — 90%-plus touchless reconciliation across 37 countries, 75% lower cost per invoice, Day-1 finance through a $9B separation without interruption. What remains is the decision.