Your AI spend shows up in usage dashboards. It disappears somewhere before the P&L.

I'm Vishnu Balasundaram, founder of The Ops Layer. The Routing Layer is my monthly letter on execution inside AI-era companies: each issue takes one problem and works it down to the mechanism, specific enough to act on without hiring anyone. This first one is for founders who signed the AI invoices and cannot yet say what they bought.

Ask a founder what AI costs and the answer comes in seconds. Seats, tokens, licenses, down to the dollar. Ask what it earned and the sentence falls apart halfway through.

That gap is not an accounting problem. It is a measurement asymmetry. Spend is visible because it arrives as an invoice. Usage is visible because every vendor ships a dashboard proving your team logs in. Value has no dashboard, because value only shows up in the shape of the work itself: how fast a piece of work crosses the company, how fast decisions get made, how often exceptions interrupt someone senior, how much gets redone.

So companies measure what is easy and call it progress. Adoption is up. Prompts per employee are up. The board deck has a slide with a rising line on it. And the P&L stays quiet.

McKinsey's 2026 State of AI survey, published this week, puts hard numbers on the silence. Nearly nine in ten organizations now use AI regularly in at least one function, and 44 percent report AI scaling across the enterprise, up from 38 percent last year. Eighty percent of respondents say AI has improved their individual productivity. And yet only 37 percent can attribute any EBIT impact to it at all, a number that has not moved in a year. The high performers, roughly 6 percent of respondents, are the ones who attribute 5 percent or more of EBIT to AI, and they are separated from everyone else most visibly by one behavior: nearly three-quarters of them have fundamentally redesigned workflows around the tools, against one-quarter of everyone else.

Read those numbers together and the diagnosis writes itself. The tools work almost everywhere. The returns show up almost nowhere. The variable that separates the two groups is not the model. It is whether anyone rebuilt the work around the model.

Here is what the leak looks like inside a real company.

A copilot cuts drafting time on a proposal from three hours to twenty minutes. The draft then enters the same approval chain it entered last year, waits in the same queue, gets reviewed by the same overloaded person, and ships on roughly the same day it would have shipped anyway. The task got faster. The workflow did not. The saved 160 minutes evaporated into wait states nobody measures.

Multiply that across a 60-person company and a second problem appears. More output means more decisions, more handoffs, more edge cases that the old process never anticipated. Those exceptions need a human to resolve them, and in a founder-led company they route to exactly one human. You paid for acceleration and bought yourself a longer escalation queue. That is the chaos tax, and it compounds while the usage dashboard glows green.

The fix starts with measuring conversion, the layer between output and margin that nobody instruments. McKinsey's data backs the instinct: high performers are twice as likely as everyone else to have defined processes for measuring the impact of their AI initiatives.

Here is the audit, four traces, one week, no consultants required.

1. Workflow cycle time. Pick the three workflows your AI spend touches most. Measure elapsed time from start to shipped, before adoption and now, and be strict about elapsed: the clock keeps running while work sits in a queue. If drafting time fell 70 percent and cycle time fell zero, your value is leaking in the waits between steps, and no better model will fix that.
2. Decision speed. Pull the last ten decisions that crossed functions. Count the days from the moment the question surfaced to the moment someone decided, and count how many of the ten landed on your desk. If AI made the work faster but decisions still queue for you, the bottleneck did not shrink. It moved.
3. Exception volume. For one week, count every interrupt that needed a human to break an ambiguity: an unclear owner, two conflicting priorities, a handoff with no receiver on the other end. If output rose and this number rose with it, your AI investment is producing escalations at scale.
4. Rework. What share of AI-assisted output gets redone after review? A fast wrong draft costs more than a slow right one, because review capacity, usually yours, is the scarcest resource in the building.

Wherever one of these four numbers refuses to move, that is where the conversion breaks, and that is the exact spot where redesign has to happen: who owns the step, who decides at the step, where the exception routes when the step fails. Structural work. Unglamorous, which is probably why so few companies do it.

I have seen what happens when someone actually does. In enterprise programs I ran, redesigning 90 plus measured workflows produced 50 percent plus cycle-time improvements, and automating 200 plus recurring reports removed over 1,000 manual hours a month. The software was the same before and after. What changed was the path the work traveled.

Your AI spend is already committed. The dashboards already look good. The open question is whether the work around the tools gets rebuilt before the next renewal, because until it does, the invoices are real and the returns are theoretical.

Run the four traces this week. The numbers will tell you where to start.

If the four traces confirm what you suspected, the 7-Day Clarity Sprint is the next step.