The Unhedgeable Signal

Every layoff press release is a firm admitting the work was already worthless to it. The dangerous part is what that admission does to the price of everything left. A computer scientist, an economist, and a biologist proved the same thing without ever meeting. AI just made it matter. Everyone is treating AI as a discount. It is a repricing. The discount is on the cheap work, and the repricing falls on everything cheap work used to vouch for.

The headlines read the same way every week. The Big Four trim their analyst pyramids. The agencies cut the deck-builders. A fund swaps a research team for a model. The story is always cost: the work got cheaper, so the workers got fewer. Read that as a confession, not a strategy. When a firm automates a task, it is telling you something it decided long ago: that task was already cheap. The intelligence in the slide, the diligence pack, the sensitivity table. They had quietly written it off. AI just let them act on a verdict they reached years earlier. They are right that the work was cheap. They are about to learn what "cheap" costs.

The Conventional Wisdom

The received view in 2026 goes like this. AI collapses the cost of cognitive work. Whatever can be specified can be generated. Margins compress in any business whose product was expertise. So the smart move is to ride the cost curve down: automate hard, shrink the labour line, pass some savings to the client, keep the rest. It sounds like efficiency. It is the most expensive mistake on the board.

Value in a market under deception pressure was never the work. It was the cost of the work that could not be faked. Collapse that cost and you have not made yourself efficient. You have made yourself indistinguishable. There is a formal floor under that claim. Three fields, none of which cite each other, none of which were trying to describe consulting, already proved it. They describe the same machine.

Three Proofs of One Machine

Start in a computer lab in 1967. Gene Amdahl was arguing about parallel processing, and he proved a hard limit. The speedup from parallelising a task is capped by the fraction that must stay serial. Automate ninety percent and the last ten percent, the part that cannot be split, becomes the entire ceiling. That is Amdahl's Law. The half people miss runs the other way. As the parallel work falls toward zero cost, the value does not vanish with it. It migrates. It piles up entirely in the serial fraction. Whatever cannot be parallelised becomes the whole price. So ask the question of advisory work: what is the serial fraction? Not the analysis. The judgment call at the end. The commitment nobody can make on your behalf.

Now walk into a Chicago economics seminar in 1921. Frank Knight split two things people had been blurring together. Risk is calculable. You can assign it probabilities, so you can insure it, so you can hedge it. Uncertainty is not calculable. No probabilities attach, so it cannot be priced and cannot be laid off onto anyone else. From that split Knight drew a conclusion that still unsettles people. Risk earns no long-run profit, because anything hedgeable gets competed down to its cost. Profit exists only as the payment for bearing true uncertainty: the cost you cannot offload. Look at what that does to Amdahl. The serial fraction, the judgment call AI cannot touch, is Knight's uncertainty. It is unpriceable for the same reason it is unautomatable. It is unhedgeable.

Now leave the building and go watch a peacock. In 1975 the biologist Amotz Zahavi proposed the Handicap Principle, and in 1990 Alan Grafen turned the intuition into a formal proof. A signal is honest in equilibrium if and only if its cost is prohibitive for a faker to bear. Make the signal cheap and mimics flood in until it means nothing. The cost is not a defect in the signal. The cost is the signal. The peacock's tail is credible precisely because a sick bird could not afford to grow one.

The only credible signal is an un-fakeable cost. The only un-offloadable cost is the bearing of uncertainty. And bearing uncertainty is exactly the serial fraction AI cannot reach.

Call it the Unhedgeable Signal. In any market under deception pressure, the sole durable source of value is a cost that cannot be hedged, pooled, insured, or faked. AI is a machine for making costs hedgeable. It can do one thing to that signal: raise its price, by destroying everything cheap around it.

The Trap, Stated Plainly

Here is what the cost-cutters walked into. AI did not make their analysts cheaper. It made analysis cheap. And by Zahavi, a cheap signal is no signal. The deck, the model, the diligence pack were never just deliverables. They were costly signals of seriousness. A client could not easily tell good judgment from confident nonsense in advance, so they read effort as a proxy: the firm that spent three weeks and a senior partner's nights on your problem was betting its reputation on the answer. The cost vouched for the conviction.

Automate that cost and the proxy stops pointing at anything. You can generate the same pack in an afternoon. So can your competitor. So can the client. The signal has gone to zero, which by definition means it no longer signals. You did not cut cost. You cut your own credibility and filed it under efficiency.

This is the formal underpinning of what I have called Conviction Capital. The argument there was that when information goes to near-zero cost, real options theory inverts: the scarce factor stops being the wisdom to wait and becomes the willingness to commit under irreducible ambiguity. The Unhedgeable Signal is the proof underneath that thesis. The conviction premium is the uncertainty premium is the handicap. Same equilibrium, three proofs, a century apart. Dixit and Pindyck priced the right to defer. Knight, Amdahl, and Zahavi price the right that matters now: the cost of committing when you cannot hedge.

The Machine Can Hold Capital

A sharp objection deserves a straight answer. Agentic AI can hold capital. An autonomous agent can take a position, post collateral, carry genuine skin in the game. If the agent bears real, un-fakeable cost, then the handicap is no longer uniquely human. Zahavi does not care whether the costly tail belongs to a bird or a bot. That is correct as far as it goes. The handicap is not human by birthright. It is cost that cannot be offloaded. If an agent bears that cost, it sends that signal.

But follow the cost up the chain. Some principal chose to deploy that agent. Chose the strategy, funded the collateral, accepted the downside, put their name on the outcome. The agent did not decide to be unhedged. A human decided to trust it unhedged, on the record. The un-offloadable commitment does not disappear when you insert a machine between yourself and the risk. It regresses to whoever pulled the trigger.

The handicap moves. It never vanishes. You can automate the bet. You cannot automate the act of standing behind it.

What Survives

So price the analysis at zero. It is going there with or without your permission. Charge for the conviction. Charge for the part of the work that costs you something you cannot get back if you are wrong: your name on the result. That part was never the deck. The deck was always the cheap, hedgeable, parallelisable wrapper around the one expensive thing inside it.

You cannot automate the part that costs you something. It was never the work. It was the willingness to be wrong on the record.