Leading Indicators That an Autonomy Program Is Working

Waiting for the P&L to confirm an autonomy program is like steering a ship by its wake. The signals that actually tell you whether it's working show up months earlier, in places the finance report never looks.

Three months into an autonomy program, the operations leader is in front of the board, and the board wants to know one thing: is it working. The honest answer, if you only look where the board looks, is that it is too early to say. The cost line has barely moved. Headcount is unchanged, because nobody was let go and nobody is going to be until the leadership is sure. The margin looks like last quarter's margin plus noise. By the only measure the board reflexively trusts, the program is indistinguishable from having done nothing at all, and there is a real temptation, sitting in that room, to conclude that the thing has not landed and the budget was a mistake. That conclusion would be wrong, and the reason it would be wrong is that the P&L is the last place an autonomy program shows up, not the first. Financial statements are a trailing record of decisions that were already made and work that was already absorbed. They tell you what the operation was, one accounting period in arrears. If you wait for them to move before you believe, you have chosen the slowest possible instrument to read the fastest-moving change in the business.

The problem is not that the financial signal is false. It is that it is late, and lateness has a cost of its own. A program that is genuinely working but has not yet touched the margin will look, to an impatient board, exactly like a program that is failing, and the two get killed the same way. This is not a hypothetical failure mode. Gartner has predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs and unclear business value — and "unclear business value" is very often not the absence of value but the absence of a way to see it before the financials confirm it. The organizations that pull the plug are frequently the ones measuring the wrong thing on the wrong clock. What they needed was a set of leading indicators — operational signals that move early, move first, and predict the financial result with enough lead time to hold your nerve.

The signals move upstream of the money

The good news is that an autonomy program leaves a trail long before it reaches the income statement, and the trail is legible if you know which instruments to watch. The first and cleanest of them is resolution rate: the share of incoming work — tickets, cases, requests, exceptions — that gets carried all the way to done without a human touching it. In the first weeks this number is low and it should be, because the Autonomous AI Workers are handling only the most unambiguous cases while the operation learns to trust them on the rest. What matters is not the level but the slope. A resolution rate that climbs week over week, ten percent to twenty to thirty-five, is telling you something the P&L will not say for two more quarters: that the routine work is genuinely leaving human hands and not coming back. When that curve bends upward and stays bent, the cost line is already committed to following it. You are simply seeing the cause before the effect posts.

The second signal is subtler and, to my mind, the most reliable of all, because it is the hardest to fake. Watch what happens to the escalations — the cases the system does not resolve on its own and hands to a person. In an operation that is merely deflecting work, escalations arrive raw: a ticket bounced to a human with nothing attached, the equivalent of a shrug, so that the person has to start the investigation from zero and the escalation has saved no one anything. In an operation where autonomy is actually working, the escalations start arriving pre-investigated. The case that lands on a specialist's desk comes with the context already assembled, the relevant history pulled from the systems that hold it, the likely cause identified, and a recommended action attached — so the human is no longer investigating but deciding, which is a fundamentally cheaper act. When your team stops saying "these escalations are noise" and starts saying "these escalations are already half-done when they reach me," you are watching the program work in real time, months before that shift shows up as a lower cost-to-serve. The quality of the handoff is a leading indicator; the cost of the handoff is the lagging one.

The third signal is the one the people doing the work feel first, and it is time-in-queue: how long a unit of work waits between arriving and being acted on. This is the number that collapses earliest and most dramatically, because it does not depend on the autonomous layer resolving anything at all — it depends only on the work being read, understood, and routed the moment it arrives rather than sitting in a backlog until a human gets to it. A queue that used to hold work for six hours starts holding it for six minutes, not because anyone is working faster but because nothing is waiting for a person to notice it. Time-in-queue is a leading indicator precisely because it moves before resolution rate and long before cost: it responds to the presence of a system that is always reading the inbound stream, even in the earliest phase when that system is still handing most of the actual decisions back to people. When the queue drains and stays drained, the backlog that used to hide unquantified losses — the renewal nobody chased, the exception that aged into a write-off — stops accumulating, and that is money the P&L has not yet learned about.

How to read them before the financials confirm

Having the signals is not the same as reading them, and reading them badly is its own way to kill a good program. The discipline that matters is to watch slopes, not levels, in the early going. A resolution rate of thirty percent means nothing in isolation; a resolution rate that went from ten to thirty in six weeks means almost everything, because the trajectory is the forecast. The instinct to benchmark against some absolute target — "we need eighty percent autonomous resolution before I call this a success" — is exactly backwards in the first quarter, when the only question worth asking is whether the curves are bending in the right direction and holding. A program on a healthy slope at a low level will beat a stalled program at a high level every time, and the slope is visible weeks earlier.

The second discipline is to insist on instrumentation as part of the deployment itself, not as an afterthought bolted on when the board starts asking questions. This is one of the quiet advantages of building autonomy on a platform designed to be observed — where the AI Missions running across the operation emit a continuous record of what they resolved, what they escalated, and how long each unit of work waited, so that resolution rate and time-in-queue are not numbers someone has to go reconstruct from logs but live readings off the operation. A system with Human-in-the-Loop wired into its decision points produces the escalation-quality signal as a natural byproduct, because every handoff is a structured artifact with its investigation attached rather than an email into the void. The measurement is not a separate project. It is the exhaust of a program built to be watched, and an autonomy effort that cannot show you these curves on demand is not one that is failing so much as one you have no way to know is succeeding — which, in front of an impatient board, amounts to the same fatal thing.

The last discipline is to connect the leading indicators to the lagging one out loud, before the lagging one moves, so that everyone in the room understands the sequence they are watching. When resolution rate climbs, cost-to-serve will fall — later. When escalations arrive pre-investigated, the cost per exception drops — later. When time-in-queue collapses, the silent losses in the backlog stop compounding and the revenue-leakage lines firm up — later. Saying this explicitly is what buys a program the runway to reach the quarter where the financials finally agree. It reframes the P&L not as the verdict but as the confirmation, and it turns the three-month board meeting from an anxious referendum into a status check against a forecast the operational data already made. This is the posture behind the broader shift toward an autonomous enterprise: the recognition that an operation running on Autonomous AI Workers throws off a richer, faster, and more honest signal than the one accounting was built to capture, and that learning to read that signal is not a reporting nicety but the difference between a program that survives long enough to pay off and one that gets canceled a quarter before it would have.

So the mental model to carry out of the board room is this. The P&L is the wake, and steering by the wake means you learn where you were, never where you are going. The leading indicators — resolution rate bending upward, escalations arriving already investigated, time-in-queue collapsing toward nothing — are the bow of the ship, and they are pointed at the outcome long before the water behind you records the turn. An autonomy program that is working announces itself in those three curves weeks or months before the money moves, and the operators who learn to trust the bow over the wake will keep the programs their competitors killed for looking, too early and in the wrong place, and finding nothing yet to see.