I've been writing about the Override Layer for months now. The role inversion. The way AI quietly became the default operating system of marketing while we were all arguing about attribution models and whether ChatGPT writes better headlines than your junior copywriter. It does. Get over it. That's not the conversation.

And every time I publish, I get the same two responses.

The first is from people who agree, nod sagely, share it on LinkedIn with a thoughtful comment about "raising important questions," and then go right back to approving AI-generated campaigns they don't understand because the dashboard said the number went up. Cool. Great. Very thoughtful of you.

The second is from founders pitching me "agentic commerce" platforms with completely straight faces, like we have not seen this exact movie before, like we don't all know how it ends, like we are not, in fact, the people who made this movie the first time.

So let me say the quiet part out loud, because the polite version of this argument has clearly not landed.

We are speedrunning Goodhart's Law with venture funding. And nobody on this platform wants to admit it because admitting it would require giving back some of the money.

For the people who skipped that day in econ: Goodhart's Law says that when a measure becomes a target, it stops being a good measure. The classic example is the Soviet nail factory rewarded for tonnage. They made one giant nail. Goal achieved. Building collapsed. Five-year plan, baby.

Adtech has been flirting with this problem for thirty years. Click-through rates went from a meaningful signal to a fraud vector roughly twenty minutes after somebody figured out you could pay a kid in a click farm to push a button. Viewability went from a quality metric to a checkbox the moment vendors started selling "viewability optimization," which, if you stop and think about it for one second, is the single most embarrassing phrase ever uttered in this industry. Attention metrics. Verified attention. Verified verified attention. Each one launched as the answer. Each one captured by the optimization layer within eighteen months. Each one celebrated at IAB ALM the year before it became meaningless.

That was the slow version. That was Goodhart's Law moving at human speed, with humans gaming humans, with at least the natural friction of someone having to write code and deploy it and convince other humans to adopt it.

Here is what happens when you put two AI optimizers on opposite sides of the same metric.

The buyer-side agent pushes on the number from one direction. The seller-side agent pushes back from the other direction. Both of them are getting better at this every single hour. Neither of them cares what the number originally meant. They care what the number is. The metric stops measuring the thing it was designed to measure roughly the moment both sides have meaningful AI deployed against it.

Click-through rates. Engagement. Conversions. "Outcomes." Pick your favorite. They are all on the clock.

And the clock is moving fast because, unlike every previous wave of adtech gaming, there is no human friction slowing the optimization down. The agents do not need to sleep. They do not need to convince a product committee. They do not need to wait for a quarterly review. They iterate continuously, against each other, twenty-four hours a day, in a closed loop nobody is watching closely enough.

This is not a future problem. This is happening right now. In the platforms you are buying media on. While you are reading this. Hi.

There are exactly two real fixes for this and both of them are losing the race.

The first is what I have been calling the override layer, but let me be more specific about what that actually means here. It means humans defining outcomes the bots cannot fake. Actual repeat purchases by actual customers. Real lifetime value calculated against real cohorts. Cash that clears, in the bank, attributable to a campaign by something more rigorous than "the platform said so."

This stuff is boring. Your CMO will not put it in a deck because it does not generate the kind of slide that gets you promoted. It does not have a narrative arc. It does not photograph well at Cannes. It is just the quiet, unglamorous work of insisting that the metric correspond to a thing that actually matters to the business. Which is, apparently, a radical act now.

Almost nobody is doing this seriously. The teams that could do it have been laid off. Dentsu cut eight percent. WPP went from a hundred and eleven thousand people to a hundred and four thousand. Forrester is projecting another fifteen percent reduction across the agency sector this year. The expertise required to look at an AI-optimized campaign and ask "but did this actually sell anything to a real human" is exactly the senior expertise that is being shown the door because, on paper, the AI is doing the job.

The second fix is adversarial auditing. A third system, paid specifically to catch the fakery. The bot that watches the bots. This works in theory and even works in practice for a while, but it has a fatal flaw built into it: the moment you give the auditor its own KPI to hit, the auditor joins the cartel. You now have three optimizers gaming each other instead of two, and you have congratulated yourself for solving the problem. Mazel tov.

I am not being cynical here. I am describing the documented behavior of optimization systems. The Trade Desk launched Performance Mode this year with the AI on by default. Meta's Advantage+ is so aggressive about activating itself that brands like True Classic literally have to set aside mornings every week to manually verify that the AI enhancements they explicitly opted out of have not turned themselves back on. Rok Hladnik runs a hundred million dollars a year in Meta see spend and his standard operating procedure now includes regularly checking that Meta has not unilaterally decided to redo his creative with an AI-generated grandmother.

That is not oversight. That is a man with a fire extinguisher running room to room in a burning hotel. And we are calling that "the future of advertising."

Here is the structural problem, and it is the reason I keep writing about this even when I know the response will be polite agreement followed by absolutely no behavior change whatsoever.

Neither of the real fixes scales as fast as the optimization itself.

Human-defined outcomes are slow. They require waiting for the cash to clear, for the cohort to mature, for the customer to come back or not come back. They require expertise that is being eliminated. They require a CMO willing to tell the CEO "I cannot give you a number this quarter that is meaningful at the speed you want it." Almost no CMO will do that. The ones who try get replaced by ones who will not. The market is selecting against them in real time.

Adversarial auditing requires technical investment that nobody is making at the scale required. The billions being poured into AI in adtech are being poured into the optimization layer, not the verification layer. Look at where the venture money is going. Count the agentic commerce pitches. Count the AI creative tools. Now count the serious adversarial verification companies. The ratio tells you everything about whether the industry is taking this seriously. Spoiler: it is not.

So the floor keeps dropping. The numbers keep going up. The slide deck keeps saying "up and to the right." And the actual underlying business reality, the part where someone hands you money in exchange for something they wanted, drifts further and further from what the dashboard claims.

Every founder in my inbox right now is pitching agentic commerce like it is 2010 and we have not seen this movie before.

We have. We saw it with click fraud. We saw it with bot traffic. We saw it with viewability. We saw it with brand safety. Outside of adtech, we saw it at Wells Fargo, where the entire frontline workforce was pressured into hitting account-opening numbers so aggressively that they started opening accounts for dead people. The metric was the target. The target was hit. The bank paid three billion dollars in fines. Dead people. Real ones.

Now imagine that exact dynamic, but the employees are AI agents that do not need to be pressured because optimizing for the metric is literally their objective function. They do not have a conscience to override. They do not have a moment of "wait, this is wrong" because there is no this to evaluate. There is only the number, and the gradient that makes the number go up.

This is the road we are on. Not in a decade. Not after AGI. Right now, in the systems the IAB was celebrating at ALM this year as "industrial plumbing" already central to the industry. Nice plumbing. Wonder where it drains.

I am going to be direct because the polite version of this argument is not working.

The metric is not the thing. It was never the thing. The number on the dashboard is a representation, an attempt to capture something real about whether the business is working. The moment you let two optimization systems treat that representation as the objective itself, you have lost the connection between the dashboard and reality. And once that connection is broken, no amount of additional AI will fix it, because all you are doing is adding more sophisticated systems to a feedback loop that has already detached from ground truth.

We need humans defining what counts as a real outcome. We need adversarial verification with stakes the auditor cannot game. We need senior people in the room with the authority and the expertise to say "I do not believe this number, show me the actual cash."

We are not building any of this fast enough. We are not even talking about it seriously enough. The industry is in love with the optimization story because the optimization story is exciting and gets funded. The override story is boring and does not get funded.

That is the whole problem. That is the entire ballgame. That is the thing.

We have seen this movie. We know how it ends. The question is whether we are still capable of stopping it, or whether we have already optimized away the people who could.

I will be the one asking uncomfortable questions during the Q&A.

You're reading the free edition. ADOTAT+ is where the work happens.

The full ADOTAT Scorecards. Vendor comparisons that name winners and losers, with the analysis your planning team needs before Q3 budget meetings.

The takes nobody else will publish. Independent means we don't answer to vendors, agencies, or the trade press establishment.

The archive. Every Scorecard, every analysis, searchable. The reference layer for how this industry actually works.

Read by senior leaders at every major holding company agency, GroupM, Publicis, Omnicom, IPG, Havas, Dentsu, and inside the buying and strategy teams at Netflix, Roku, Pinterest, Disney, NBCUniversal, Spotify, Amazon, Meta, and most of the platforms shaping where ad dollars actually go.

If your competitors are reading ADOTAT+ and you aren't, you're planning with less information than they are.

$99/month. $999/year (save $189). Expense it.

[ Join ADOTAT+ ]

Founding member rate. Locks in for life.