Impressions Are the Bill.

LinkedIn told me my last article reached 2,539 people. My own analytics told me six of them read it.

Both numbers are true. For ten days the post sat at forty impressions a day; then, between the 9th and the 13th of July, it climbed — 487, 785, 1,801, 2,539. A fivefold spike, and across the whole run not one new reaction, comment, or measurable click. Whatever LinkedIn's distribution did, the signature is unmistakable: a wide, loosely-matched audience that was never going to act. When I traced where readers actually came from, the number was six.

Measuring outcomes is what I do for a living — and I still spent the better part of two weeks quietly pleased with a curve that measured nothing I wanted. That is how strong the pull of the meter is.

This isn't really a marketing story. It's the same error of attention I've watched serious organisations make with far more money on the line — and this month McKinsey put a number on how widespread it has become.

Tokens are the bill, not the value

In a 2026 McKinsey Quarterly paper, Is that AI agent worth it?, Hämäläinen and colleagues report that 93 percent of respondents to one enterprise survey are over their AI budgets, and a fifth have already constrained AI use because of what it costs to run. The instinct is predictable: cut the token bill — cheaper models, shorter answers, tighter caps.

The authors wave a flag first. As Pay-i's David Tepper puts it: tokens are not value; tokens are the bill (Tepper, as cited in Hämäläinen et al., 2026). The cost of a model call is the easiest thing in the system to measure, so it becomes the thing everyone governs — while the real question, whether the finished work is worth more than it cost to produce, goes unasked, because no one defined the outcome it was meant to produce. Their conclusion is blunt: the unit of governance is the completed business outcome, not the token.

Read that against my post. Impressions were my tokens — the cheap, abundant, flattering number — and I governed them instead of the only thing that mattered: whether the right person read the argument and acted. Impressions are not value. Impressions are the bill.

Why the meter is so easy to trust

The trap is near-universal because meters are engineered to be trusted: precise, real-time, and usually moving the way you hoped. Outcomes are none of those things — lagging, noisy, and often embarrassing.

McKinsey's analysis shows how costly that gap gets at scale. Roughly 60 percent of an agentic task's cost isn't the first answer at all — it's the checking, repairing, and re-verifying afterwards. In programming, the same task can cost thirty times as much depending on the path the system takes; average-cost budgets, the authors note, will not be enough. A meter reading "cost per call" tells you almost nothing about whether the work was worth doing.

So the fix is not a better dashboard. You cannot instrument your way out of never having decided what you were trying to achieve.

This is an architecture decision, not a reporting one

The line that should stop every operator is this: no autonomous system should operate without a defined mandate, budget, and stopping rule. McKinsey calls the underlying discipline allocating intelligence like capital — deciding, before anything runs, where reasoning is worth spending and where it is not.

That is not a monitoring requirement. It is a design-time decision, made in a room where someone commits to what a system is for, what outcome it answers to, how much autonomy it holds, and the condition under which it stops. This is what we mean at Mandantis when we say governance is architecture, not reporting. A dashboard tells you what a system did; an architecture decides what it is allowed to do, and against which outcome, before it does anything. Get those decisions right up front and the meters finally become useful, because you have something to hold them against. Skip them, and every number you collect afterwards is just weather — precise, real, and about nothing you chose.

My stakes were vanity; a CEO signing an AI invoice this quarter is holding millions. But the reflex is identical: govern the instrumented proxy because it is measurable, instead of deciding the outcome first.

Six

I know all of this because I just did it to myself, in public, with a number I should have known better than to trust. The meter said reach. The outcome said six.

Decide what the outcome is first. Then, and only then, go read the meter.

---

If you are staring at an AI invoice you cannot yet tie to an outcome, that is the conversation we have. You can see where your own operating model stands at index.mandantis.com.

Figures on the author's own post are drawn from LinkedIn analytics and Mandantis CDP arrival data, 30 June – 13 July 2026.

References

Hämäläinen, L., Patel, M., Blumberg, S., Catlin, T., & Lala, W. (2026, July). Is that AI agent worth it? Agentic economics and the modern operating model. McKinsey Quarterly. https://www.mckinsey.com/capabilities/quantumblack/our-insights/is-that-ai-agent-worth-it-agentic-economics-and-the-modern-operating-model