AI Has No Skin in the Game

For twenty years the prediction was that ad agencies would disappear. Buyers and sellers would transact directly, the tools would keep getting better, and the middle would go away. The tools got better. The middle is still here.

That fact is worth sitting with, because the industry is now running the same prediction again at a much larger scale. The new version goes like this: build an agent for every participant and every task in the transaction, and the system finally becomes efficient and transparent. A planning agent, a buying agent, a yield agent, a measurement agent. Automate the participants and the friction drops out.

It won’t play out that way, and the reason is hiding in the word we chose.

“Agent” was a legal term long before it was a software term. An agent, in the original sense, is a party who acts on behalf of a principal, owes duties to that principal, and can be held liable and dismissed. Software agents borrowed the word and none of the obligations.

Every agent is built by someone, for someone. The buyer’s agent optimizes for the buyer. The seller’s agent optimizes for the seller. Every agent in the middle optimizes for whoever deployed it, with that stakeholder’s bias and targets built in. So when the transaction fills with agents, the conflicts of interest don’t disappear. They execute faster. Negotiation doesn’t end either; it moves to agent against agent, at machine speed. And someone still has to decide whose interest wins when the goals directly collide. Machines don’t resolve a conflict of interest. They run one.

Which surfaces the deeper problem. Say your agent makes a bad call. A consequential one, with real money attached. What do you do? You can’t fire it. It’s software you deployed, doing what you built it to do. Blaming your own agent is blaming yourself.

Every consequential transaction needs a party who carries the risk, owns the outcome, and answers for it when it goes wrong. That, I’ve come to believe, is why the agency prediction kept failing. The intermediaries that survived do two jobs at once. They create real value, and they carry responsibility for it. Look at which ones actually got disintermediated over the last thirty years and which didn’t. Travel agents created informational value, and the internet took that job. Real estate agents, lawyers, and yes, ad agencies deliver judgment with consequences attached: they can be held to account, made to fix it, and replaced if they don’t. Automation keeps copying the first job. It cannot take the second, because accountability is the one thing a tool cannot carry.

An AI tool has no skin. Only a principal does.

None of this argues against building or deploying agents. The leverage is real, and the teams building orchestration layers are solving a genuinely hard problem: getting fleets of agents to work in concert. But orchestration coordinates the agents. It doesn’t decide whose interest wins, and it doesn’t own the outcome. The better the agents get, the more the whole system depends on the person standing above them.

So where does this leave the people? Here is the honest version, in the order it deserves.

AI will displace most of the people in most of the functions it touches. That much is true, and pretending otherwise insults everyone living through it. But the judgment does not leave the system. It moves up a layer. Somebody has to decide what all these agents optimize for. Somebody has to adjudicate between them when their goals collide. Somebody has to own the result when a loop of machines produces an outcome nobody chose. That is decisioning at a higher altitude than the work it replaced: judgment, discernment, full ownership of outcomes. Command center work. The people who can do it, who can stand at the top of the stack and answer for what the machines below them do, are worth more now than they have ever been.

The bar rises. That is the event. The spread between someone who can own an outcome and someone who can perform tasks has widened enormously, because the tasks are what the machines took.

And here is the hopeful part, which I believe as firmly as the hard part. The same tools that raised the bar also cut the cost of reaching it. Someone early in their career can now build genuine working depth in three or four disciplines in a fraction of the time it took my generation to build one. They can connect those depths and become formidable years earlier than the old path allowed. What they cannot compress is judgment. Judgment is still made of reps: the deal that went sideways, the hire that didn’t work, the decision owned in front of other people and answered for afterward. The tools accelerate skill. Wisdom still runs at the speed of experience.

So if you’re deploying agents this year, and everyone is, the question is not which ones. The question is who in your operation has enough skin in the game to answer for what they do.

The X is the HXV publication on the crossing point of technology, commercial reality, and human consequence. New essays roughly monthly. Subscribe below.