Point of view · June 2026

AI can do the work. It can’t own it.

Last year, roughly a quarter of one Y Combinator batch shipped products almost entirely written by AI. Many of the headlines read it as the end of engineers. I read it the other way.

The number behind those headlines came from Garry Tan, who said that for about a quarter of that batch, roughly 95% of the code was written by AI.

Ninety-five percent of the code being written by AI did not remove ninety-five percent of the work. It moved it, to the part that was always hard. Someone still has to decide the architecture, read the security exposure, catch the edge case the model was confident about and wrong about, and put their name on the result.

The typing got cheap. The judgment behind it did not.

I should put my own bias on the table. I sell judgment for a living, so of course I am going to tell you judgment is the scarce thing. Do not take it from me. Take it from what happens to a company structurally when the work and the ownership of the work come apart, which is exactly what just happened.

The thing that got scarce

Execution used to be the constraint. It isn’t anymore. For most of my career the question was never what good looked like; it was whether you had the hours, the headcount, and the budget to produce it. A founder can now ship in a weekend what used to take a team a quarter.

The investors are saying it too. Andreessen Horowitz put it plainly in its Big Ideas for 2026: the moat is not the model. As frontier capability commoditizes, the durable advantage moves up the stack, into proprietary workflows, domain data, and the operational logic a company already runs on. The usual next move is to credit taste, the person who can tell good output from bad. That is half of it. Discerning which call is right and owning that the call was yours are two different things, and only the second one ever has a name on it.

When execution stops being the bottleneck, the new scarce resource is a person willing to own the call, not just generate the work that informs it. Osman “Ozzie” Osman, a co-founder of Monarch, put the cleanest frame on it I have read: accountability is not punishment, it is confidence.

“It’s the confidence you have in work that has someone you trust’s name on it.”

Osman “Ozzie” Osman · Co-founder, Monarch

AI can generate the work, but it cannot be accountable for it, because it has no reputation to lose and no judgment of its own to answer for. That is not a knock on the technology. It is simply what accountability has always meant.

Which is why organizations hesitate, correctly, to act on analysis no one stands behind. You can trust a person who is accountable for being right; a model gives you no one to trust that way. The signature is what makes an output trustworthy. Not the polish.

The first producer with no stake

“Everyone puts their hand up and says, we opine on the decision, we chip in on the decision. But we don’t own it.”

An operator who has spent years inside heavy industry, on how decisions actually get made.

AI is the first producer with nothing on the line. For most of corporate history, work got delegated but the stake never left a person: a junior analyst built the deck, a partner signed it, and both had a name, a reputation, and a career on the line. When Apple named a Directly Responsible Individual, or Amazon a single-threaded owner, it was formalizing who carried that stake.

It can produce the analysis, the code, the memo, the plan at a level that used to require a senior person, and it has no reputation to lose and no consequence to absorb. The work still gets produced. What used to stand behind it is gone.

That is what makes “who signs it” load-bearing. It is not a new framework; the Directly Responsible Individual has existed for fifteen years, back when the producer was always someone. What changed is that the most productive worker in the building is now the one entity that cannot be held to account, and the org chart has not caught up.

The gap was there before AI: a company measures its numbers to the decimal but owns the decisions underneath them almost nowhere. Revenue, SLAs, the quarterly number all have a steward. The quality of the calls that produce them has none. The biggest sit with the CEO, the next tier with the C-suite, and from there they cascade and diffuse until the thousands a model now drafts by the hour scatter, many to no one in particular and plenty impossible to track. No function exists to ask whether, across all of them, the company decides well.

Now drop AI into that. If no one owned the decision when a human was slowly typing it out, the gap does not close when a model produces it in seconds. It widens. This is not hypothetical. When Air Canada’s chatbot gave a customer a refund policy that did not exist, the tribunal rejected the airline’s argument that the bot was a separate entity and held the company liable. A court sanctioned the lawyers who filed a brief built on AI-invented cases, not the software that wrote them. Deloitte, not its model, refunded a government for a report full of references that were never real. The liability landed where it always lands, on whoever put their name to the work.

An output is not an answer

The frameworks are a free download now. The judgment to say which one is right is not. There is a confident counter-version of all this, and it deserves an answer because it is half right: the issue trees, the prioritization matrices, the maturity assessments the big firms charged a fortune for, a stack of prompts will generate in an afternoon. True. The leap from there, that the analysis is therefore the same and all that is left to sell is change management, is where it fails.

Ask a model for the analysis and it gives you one. Ask it for the opposite and it gives you that too, with equal confidence, and abandons either one the moment you lean the other way. It is not that you cannot get the unwanted answer out of a model. You can; red-teaming and steelman prompts do it every day. It is that the model does not care which answer is right and pays nothing if it is wrong. The contrarian view is available. The stake behind it is not, and that stake, being on the hook for the call, is the whole of what a human signer adds.

I watched the cost of the easy answer years ago, at a Fortune 150 retailer whose store labor was stuck in a doom loop. Their scheduling chased last week’s sales, so a slow Monday cut the coming weekend, and the weekend they cut was usually the one that mattered. Leadership wanted a tidy fix: train the managers to be better. But no manager could win the job they had been handed. The comfortable answer was right there, well formatted and easy to sell. The right one meant rebuilding the metrics and the tools underneath them, harder to do and harder to hear. A model would have happily drafted either. What it could not do was own which one was true, and carry the cost of being wrong.

The issue tree was always just scaffolding. What turns an output into an answer is a person willing to say: this one. This is the answer. My name is on it, and the reputation is the collateral. That holds as much for the analysis your own team generates in an afternoon as for anything a firm ever sold.

You can’t bully a signature.

What this asks of a leader

The job now is to put production and ownership back together on purpose. The risk if you do not is the failure mode Ozzie names: producing volume in inverse proportion to expertise, confident work in exactly the areas you are least equipped to judge. The output gives no tell. A sound call and a hollow one look identical, and you can only separate them by knowing who stood behind the work. Three moves make that real.

1. Name the signer for every consequential output, the work where being wrong costs real money, safety, or trust. Not who generated it, who will stand behind it. You cannot do this for every email and do not need to; the discipline is deciding what counts.
2. Hold the signer to the soundness of the call, not just the number. Companies already own outcomes: revenue, SLAs, the quarterly number. What almost no one owns is the quality of the decisions underneath them. That is the layer AI just flooded, and you get the decision quality you measure.
3. Make sure the ownership is real. Not by re-doing the work, but by knowing where the model is most likely wrong and checking there. If the signer could not have caught the error, the signature is decorative.

Here is the part I will not pretend is solved. The outputs are good, often good enough that vetting them at the speed they arrive is genuinely hard, and not only for a junior person; the experts I trust feel it too. I have gone looking for the people who say they have this solved, and so far every solution I have pressed on has thinned out under real load. The pattern I keep seeing is the limits getting downplayed and the benefits getting sold. Maybe someone has genuinely cracked vetting at scale. I would not claim it is most of them yet, and I would love to be wrong.

What I can offer is where to look first, not a cure. Models smooth over the contradiction, drop the dissent that did not fit, and sound most certain exactly where they are weakest. So the signer’s attention goes to the load-bearing number, the claim that would flip the decision, and the source that is a little too convenient.

As AI eats the execution layer, the decision layer expands with it. The expensive, genuinely senior work is now owning the call: framing the real question, knowing which dropped data point matters, and being the name on the result. Whether or not anyone gives that a title, every company now has to build the discipline on purpose, because the worker that used to deliver the owner for free is no longer the one doing the work.

The companies I would bet on are the ones that re-weld production and ownership deliberately and stay honest that the vetting is hard, over the ones quietly trusting work that no one really stands behind.

Who signs it.

A position firmly stated and loosely held. If you think this is user error, that you have the vetting solved and I am simply doing it wrong, I genuinely want to hear it. I would rather learn it than be right about it.

Eric Tribe· erictribe.com