The Signature the Models Can't Give

This builds on The Math Under the AI Bet, but you do not need to have read it.

This weekend, on a walk along the Rhine, I listened to Dan Shipper on Lenny’s Podcast say something that landed with more weight than he probably intended. Automation is a lie. He meant it specifically, not as a slogan: every agent still needs a human watching it. Every is an AI-native company. Its team uses Claude Code, Codex, and the newest agent surfaces for almost everything a non-engineer used to need an engineer for. And by Dan’s account it has roughly doubled in headcount over the past year.¹

That is the data point worth sitting with. The most AI-forward small company you could name is hiring faster, not slower, and the person who runs it says automation is a lie.

A month ago I argued the same shape from the opposite end of the economy: the AI architecture that survives the math is one where a human signs at the top.² Dan has the platform and the track record here; he called Claude Code a year before most of us saw it. What I have is a narrower vantage, regulated enterprise, pharma, agchem, supply chain, and from inside it one piece of his prediction reads differently. He is right about the work. There is just one thing my corner of it forces that his does not have to.

What Dan actually predicted

It is worth stating his case fairly before adding to it, because he is right about more than he is wrong about.

His through-line was that work over the next year reorganises around agents, and that the reorganisation makes humans more central, not less. He expects companies to converge on a single capable super-agent rather than one agent per person, with Shopify and Ramp as early shapes of that pattern.

He expects the work surface itself to move into tools like Codex and the new co-working environments. He expects SaaS to survive and grow, because agents amplify the demand for software rather than eating it. And he expects product managers and full-stack designers to do well, because taste and judgement become the scarce inputs once baseline competence is cheap.

And underneath all of it, the line he kept returning to: agents need humans. Not as a courtesy. As a load-bearing fact about how the work actually runs.

Where it lines up with the math

Most of this maps cleanly onto the argument I made in April.

Augmentation, not replacement, is the same conclusion reached from two directions. He watches it in his hiring numbers; I derived it from unit economics and the regulatory ceiling. The architecture is multi-layered in both accounts: cheap, narrow models doing the bulk of the work, frontier models used where they genuinely earn their place, a human holding the top. SaaS survives in both accounts for the same reason: the agent needs something to act on, and that something is software.

Even the “models make yesterday’s competence cheap” insight, which Dan puts more sharply than I did, is a precise restatement of the capability-commoditisation point. When a skill becomes free, its scarcity moves up a layer, to the judgement about when and whether to use it.

There is academic weight behind the same observation now: a doctoral dissertation by Zhe Zhu at the University of Vaasa finds that generative AI’s impact depends more on how it is situated inside decision-making practices than on raw model capability.³ A practitioner running a thirty-person company and a Finnish researcher, working from completely different evidence, landed on the same point. When that happens, it is probably real.

This is where my world parts from his. Not because I see more, but because I work somewhere that adds a constraint his does not.

Where the argument has to go further

The disagreement comes down to one question: why the human is in the loop at all.

Dan treats “agents need humans” as a function of current model limitations: the human is in the loop because the model is not yet good enough to be left alone. That framing has a built-in expiry date: as the models improve, the human’s role thins. In his world, that is probably true. His customers are doing creative, language-heavy work where the cost of being wrong is a bad draft, a missed nuance, a redo.

In the world I work in, the human is not in the loop because the model is weak, but because the law puts them there: EU AI Act Article 14, Sarbanes-Oxley 404, MiFID II, the GxP qualified-person signature, Basel III model validation, none of which soften when the next model ships. The signature is a legal object, not a capability gap.

There is also a floor beneath his that he rarely has to visit. Most of the decisions I build for are not language problems at all. They are pattern recognition on structured data: supplier risk, claim adjudication, fraud, demand forecasting. That work runs on cheap, narrow machine-learning models an auditor can reproduce, with a frontier model used sparingly on top. In Dan’s world the work is language most of the way down, so that layer stays out of sight. In mine it carries the load.

That is the productive disagreement. One constraint dissolves as the technology improves. The other does not move at all.

Here is what I mean concretely

A pharmaceutical batch release. A loan approval. A market trade flagged and acted on. In each case, a model can read everything, weigh everything, draft the decision, and lay out the evidence. And in each case a named human still has to put their signature on the outcome, not because the model could not produce it, but because when a regulator asks who decided, “the agent did” is not an answer that survives the room.

Dan’s super-agent works beautifully precisely because his industry’s accountability surface is thin. Nobody audits a marketing brief the way a regulator audits a batch record. Move the same super-agent into pharma, banking, or insurance and it has to file under someone’s name. That single requirement changes how the system is designed, who is allowed to build it, and what the budget looks like over five years.

The shape of the work converges anyway

What is striking is that even across that divide, the daily texture of the work ends up looking the same.

Dan describes the lived experience inside Every as more reviewing, more gardening, more managing of agents than doing the task by hand. The architecture I have described demands exactly that, for a different reason. When the system drafts and scores and the human accepts and signs, the human’s day fills with reading, checking, and judging rather than producing. Dan arrives there because the agents are productive enough to need supervising. I arrive there because the signature requires the supervision to be real. Different cause, identical working day. The work that survives is the work of accountability.

There is a quieter complication in Dan’s strongest career advice, which is to ride whatever model is newest, try it on everything you do, and chase the moments where it surprises you. The advice is correct. The friction in following it is realer than it sounds. Recently a CTO of an insurance-tech firm in Zurich told me, with some heat, that he was all-in on Cursor. Pressed on why, he defended the choice more than he explained it. And yet for full module development he quietly switches to Claude Code. The stated loyalty and the working practice do not line up, which is the tell. The harness you name as yours is an identity; the one you reach for when the work gets heavy is a judgement, and the two are not always the same. Most practitioners ride the model they started with and rationalise it afterwards. The people who genuinely keep switching, the ones Dan’s advice describes, are rarer than the advice makes them sound.

I do not buy the absolute version of his “CLIs are over” either. I live mostly on the command line, and last month I still paid fifty dollars for a writing app when I already own three editors that open the same files. The point was never CLI against GUI. It is the surface where the next action is cheapest: typing for prose, dragging cards for triage, the terminal for nearly everything else. The surfaces multiply. They do not retire each other on a schedule.

What could move the ceiling

All of this rests on one assumption: that the regulated signature stays required. That is the load-bearing claim, and the one I am least sure of over a long horizon. Regulation is not fixed. The FDA’s posture today is a poor guide to where it lands in three years, and an agency that now insists on a named human could, with enough evidence behind it, come to accept a validated autonomous decision in some narrow class of cases. If that happens, the ceiling moves, and Dan’s reading, that the human stays only until the model is good enough, turns out to be the more durable one. I do not think it moves soon, or far. But I have been wrong about ceilings before, and “the law will not allow it” is a weaker claim than “the law does not allow it today.”

What both of us are building for

Strip the two accounts down and the same structure sits underneath each.

A system that carries the load. A human who carries the consequence. Dan came at it through productivity, from a company where the consequence is a manager’s judgement. I came at it through audit, from industries where the consequence is a signature with legal exposure. Neither of us designed the architecture. The work did, in two places at once, because the constraint that shapes it is older and more stubborn than the technology that revealed it.

Dan said one more thing near the end of the hour that I have not stopped turning over: the edge of AI sits wherever the technology meets a real human doing real work, which is rarely San Francisco. From Basel, building for industries where that real human still has to sign, it lands as exactly right.

The system carries the load. The signature carries the law. Dan saw where the work is heading. I only know the room where someone still has to sign.

Sources

1. Dan Shipper, in conversation with Lenny Rachitsky, Lenny’s Podcast, May 2026. Episode on YouTube. All quoted lines (“automation is a lie”, “agents need humans”, “models make yesterday’s human competence cheap”), the headcount figure (roughly 15 to 30 over the year), and the super-agent, SaaS, and CLI-to-GUI predictions are taken from that episode, verified against the transcript.

2. The Math Under the AI Bet, Pramod Prasanth, 23 April 2026. /essays/the-math-under-the-ai-bet

3. Zhe Zhu, Generative Artificial Intelligence in Organizations: Strategic Decisions and Human Adaptations, doctoral dissertation, University of Vaasa, 15 May 2026 (ISBN 978-952-395-272-0). Open-access PDF. Cited from the dissertation text, not the press summaries.