Most Companies Are Buying a 6% Improvement

The technology for Ford's production line existed forty years before Ford used it.

Edison had commercialised electricity in New York by 1881, and his dynamo was already efficient enough. Every factory owner in America could have built Ford's factory in 1881. Almost none did. What they did instead was leave the building exactly as it was - four floors, central shafts, belts running off a single drive - and swap the coal engine for an electric one.

That got them about 6% better.

The 3x came from something nobody wanted to do: demolish the factory. Move it out of Manhattan. Build it flat and long on cheap land in New Jersey, and redesign the entire floor around small dynamos placed wherever the work happened, because electricity - unlike a steam shaft - can go anywhere. Ford did that. The productivity gain powered the American century.

Alejandro Maza, who runs product and AI at the Latin American used-car marketplace Kavak, tells this story to his team constantly, because he thinks it is happening again right now, and that most companies are once again buying the 6%.

Here is what the other version looks like.

At Kavak today, 96% of all customer interactions are handled by agents with no human involved. 95% of transactions are fully agent-run. You meet a person when you collect your keys; that is roughly it. Between 100,000 and 200,000 agents are instantiated every single day, each one in its own virtual machine. They wake up, work for three minutes or eight hours or three days, set an alarm for their next task, and go back to sleep.

This is a company that sells used cars and underwrites car loans. Not a software product. Physical inventory, regulated lending, emerging markets, an average ticket most people take three months to decide on.

The results are not marginal. NPS tripled. The agents initially converted 50% better than the human sales team, and now convert 2.1x better. Loan approvals that take two months elsewhere in Mexico take under three minutes. Warranty claims fell 26%.

Maza is clear that none of this came from adoption. It came from three decisions most companies will not make.

The first: they rebuilt the company, not the workflow. "The first instinct is you leave your structure as it is and just give ChatGPT to your team, and then there's no efficiencies, your customers have the same problems, and nothing happens." Kavak rewrote its APIs and its internal systems so that agents - not people - could operate them. The question they asked was not "where can AI help?" It was "how would we build this company in 2035, with 2035-level intelligence?" Then they built that.

The second: an agent per customer, not per task. This is the architectural bet, and it is the opposite of what almost everyone is doing. Most agentic systems are workflows - graphs, functions, handoffs between specialist agents. Kavak had built exactly that: tens of thousands of multi-agent systems, running the business, profitable, growing. Then a more capable model shipped and Maza concluded the harness itself had become the constraint. So they destroyed two years of working architecture and started again.

What replaced it is one long-running agent per customer, with its own machine, its own memory of every interaction going back years, and a single long-term goal: maximise that customer's lifetime value. The company stopped being transactional and became relational. Ten million customers in the database, most with an agent permanently assigned to them. Activating 1% of that base is worth hundreds of millions of dollars.

The third: evals are the brakes, and brakes are what let you go fast. Kavak spends roughly as much engineering time, money and compute on evaluation as on building the agents themselves. Maza's framing is that companies moving slowly on AI are moving slowly because they cannot see what is happening - and the answer to that is not more caution, it is better instrumentation. And the thing they measure is not calls handled or minutes saved. It is: did the customer convert, did they get value, did they come back. "That's where most things break. I see companies measuring number of calls or minutes during the call - superficial KPIs that give you some information but don't really work."

Two details are worth sitting with.

The first is what happened to the org chart. In most deployments, when an agent gets stuck it escalates to a human queue and the trail goes cold - which means the failure never becomes training data. Kavak inverted it. The agent hits a wall and calls an API asking for help, and on the other end of that API is a person. Map it and you get human teams that have an agent, not the other way round. Sometimes the agent is the boss.

The second is the AI CEO. They carved out one Mexican city, Cuernavaca, and put an agent in charge of it. Target for month one was to double profit. It missed - it only delivered 50% more. It reads every number, forecasts, and messages physical staff each morning with their plan, asking for voice notes back. Inventory, financing penetration, satisfaction all improved. "That was the last job AI was supposed to take."

Everyone at Kavak, from the CEO to the mechanics, goes through a six-week internal programme they call the Jedi Academy. They ship a production agent at the end of it. The point is not to turn mechanics into AI engineers. It is that there is no version of this company where you do not know how to work alongside the thing.

And here is the part that should interest anyone building rather than defending: Maza does not think incumbents will do this. Not because they are stupid, but because Schumpeter was right. It is nearly impossible for the CEO of a large public company to stand up and say they are demolishing forty years of accumulated structure to rebuild around a technology that is two years old. They will adopt superficially, book their 6%, and report it as progress.

Which leaves the rebuild to whoever is willing to do it from scratch.

Source: Kavak's Playbook for Rebuilding a Company Around AI - a16z

Your Org Chart Is Now the Bottleneck

For thirty years, every slow-shipping company had the same alibi: we don’t have enough engineers. Hiring was the constraint. Capacity was the constraint. Implementation was the constraint.

That alibi just expired.

David Heinemeier Hansson — the man who wrote Ruby on Rails, who spent two decades arguing that beautiful hand-chiselled code was the whole point — has not written a single line of the code shipped in his new operating system by hand. Not one. In three months, he merged over a thousand pull requests into it. Three days after release, the plugin marketplace had 330 extensions. He rewrote a Python graphics library into Rust, a language he openly hates, in forty-five minutes, without reading the output. He calls his current state delirium, and he means it as a compliment.

So here is the question that should keep every founder awake: if one person can now do that, why is your team of forty shipping the same four features a quarter it shipped last year?

The answer is not your engineers.

The human in the middle

The most useful thing Hansson says in five hours has nothing to do with code. It is this: to get the 10x, you have to talk to the agent yourself, and you cannot put another human in the middle of that conversation. The bandwidth collapses. A product manager relaying intent to a designer relaying it to an engineer who then prompts the model is three lossy translations of a thing you could have said directly in nine seconds. When you have a PM, two designers, a VP above them and a CTO above them, and everyone needs to be in the shaping process to justify their existence — that, in his words, is where all the productivity goes to die.

The approval chain was tolerable when implementation took six weeks. It is the entire cost when implementation takes six minutes.

Read that again as an org-design statement, because that is what it is.

The scarce resource changed identity

The second thing worth stealing: most companies are not bottlenecked on building. They are bottlenecked on knowing what to build. Hansson’s example is Microsoft — decades of effectively unlimited engineering capacity, and the output speaks for itself. Writing a lot of code has never produced great software. What produces great software is someone with taste deciding what should exist, in what order, for whom. Make a lot of bad ideas real, very fast, and you have simply industrialised being wrong.

Which means the scarce resource in your company just changed identity. It is not engineering hours. It is judgement. Hansson puts it flatly: software is product management. What should it do? Who for? What ships in version one? Those skills were never evenly distributed across your engineers, and now they are the only ones that price.

Where it breaks. He is honest about this, and the failure is instructive. At 37signals they let the designers loose on an existing product with agents. Every individual pull request looked defensible. Taken together, they destroyed the architecture, and humans had to clean it up by hand. Greenfield is where the miracle happens. A large system with years of accumulated decisions still punishes people who cannot see structure. If you are running a real codebase, that distinction is worth more to you than any productivity statistic.

Stop specifying harder

Then the part most managers will get wrong. The instinct, when you hand work to something powerful, is to specify harder — longer briefs, tighter constraints, more rules. That instinct is now actively destructive. Anthropic shrank the system prompt for its newest model by roughly 80% because the model was being damaged by over-prescriptive humans. Hansson’s analogy is the pointy-haired boss who walks in, understands nothing, and starts telling you how to code. You sulk. You do worse work. Why would an agent be different?

The discipline is the opposite of what you were trained on: say the vaguest thing that produces something real, then react to the real thing. Nobody knows what they want until they hold it. That was the whole insight of the agile movement, and agents have finally made it cheap enough to actually practise.

The strategic read

Now the part that matters for anyone building a business rather than a side project.

There is no accumulated advantage here. Hansson’s claim is that if you had spent the last year backpacking, untouched by any of this, you could reach the frontier in two weeks. Nothing compounds. No one is building a lead in tooling. So the moat is not who adopted agents first — it is who is structurally capable of using them, and that is a question about your management layers, not your stack. Incumbents are supertankers tuned for a world that stopped existing. That is the innovator’s dilemma arriving on a schedule measured in months instead of decades.

And the opportunity is smaller and more specific than the AI discourse suggests. The old joke about Microsoft Office was that everyone only uses 5% of it — the catch being that we all use a different 5%. Hansson’s answer to that: build your own 5%. He replaced his writing app in twenty minutes and had abandoned the original within two days. Every internal tool your team tolerates because the good one is too expensive or the vendor won’t build your workflow is now a weekend.

None of this is a prediction. It has already happened; the distribution is just uneven. The companies that will look stupid in eighteen months are not the ones that were late to the models. They are the ones that bought the models, plugged them into an approval chain built for 2015, and then concluded the technology was overhyped.

Your engineers are not the constraint. Go look at your calendar.

Source: DHH on the Lex Fridman Podcast #501

Every Second Is a Bite

A video is a dish.

Not a message, not a piece of content, not an asset. A dish. Someone sits down, picks it up, and eats it — and every second of it is one bite.

That framing changes everything about how I make them.

In a restaurant, the diner controls the bites. He decides what to pick up first, what to skip, whether the sabzi comes before the dal. In a video, I control the bites. Second one is my bite. Second two is my bite. Second nine — mine. He has no fork. He opens his mouth and I decide what goes in, in what order, at what speed. There is no other format on earth where the maker has this much control over the exact experience of the consumer.

That control is the whole point. And it is a responsibility, not a privilege.

The plate must be amazing — every bite

If I control every bite, then every bland bite is my fault. Nobody else’s. Not the algorithm’s, not the audience’s, not the platform’s.

A restaurant can survive one dull dish on the menu because the diner just doesn’t order it again. A video cannot survive one dull second, because a dull second is a bite of nothing — and a person who gets a bite of nothing puts the fork down. That is the swipe. It is not rejection of me; it is a normal human reaction to a mouthful with no taste.

So the standard is simple and brutal: there are no filler seconds. No “let me explain what this video is about.” No throat-clearing. No bite that exists to set up a later bite. Every second has to taste like something on its own, and every second has to make the next one worth taking.

Retention is the main variable

Restaurants live on one thing: did he finish the plate?

Not “did he walk in.” Not “did he like the decor.” Did he finish. A finished plate is the only proof that the food was good, and a finished plate is the only thing that produces the next customer — because the man who finished it goes and tells someone.

In our terms, retention is the finished plate. Watch time is the finished plate. The share button is him bringing his friends and family to the restaurant. Nobody in the history of food has recommended a place where they left half the meal. Nobody will share a video they didn’t finish. Shares are downstream of retention, always, without exception. So I don’t chase shares. I chase the finished plate, and the shares arrive on their own.

Swipe-away is the shopfront, not the food

There are two ways to lose a person. One: he walks past the restaurant. Two: he sits down, takes a bite, and leaves.

The first is the swipe-away rate — the first frame, the thumbnail, the hook. That is the shopfront and the smell from the kitchen door. It matters, but it corrects itself over time. You post enough, you learn what makes people stop at your door, the smell gets better, the board gets better. It is a volume problem and volume fixes it.

The second is the one that decides everything. If he sat down and left, the food was wrong. No shopfront fixes that. No hook fixes that. And every person who leaves mid-plate is a person who will never bring anyone.

So my attention goes to the plate, not the door. The door improves as a side effect of a great plate — the smell of a kitchen that never serves a bad bite eventually reaches the street.

The goal is not satisfaction. It is addiction.

Jo mere restaurant pe aayega, wo nashedi ban ke jayega.Whoever comes to my restaurant leaves addicted.

A satisfied customer is a customer who might come back. An addicted one has no choice. That is the bar. Not “he liked it.” Not “he found it useful.” He finished the plate, sat there for a second, and felt the pull for the next one — before he had even decided to want it.

That means I am designing the second visit while serving the first. The last bite has to leave a taste that only the next dish can resolve. The craving is the product. The video is just how it gets delivered.

What this means at the cutting table

When I sit down to make one, the dish rule gives me every decision:

Taste every second. If a second doesn’t taste like something on its own, it goes. Not shortened — gone.

The first bite decides the meal. Whatever the strongest flavour in the video is, it goes in the first bite. Nothing is saved for later; there is no later if the first bite is weak.

No explanation bites. Nobody eats a menu. Show the dish, don’t describe it.

Season for the person eating, not for the cook. What I find interesting about my own footage is irrelevant. What tastes good to him is the only measure.

Read the plates coming back. The retention curve is the plate returning to the kitchen. Where the curve drops is exactly the bite that was bland. That is not data; that is the chef’s feedback loop, and I look at it for every single dish.

I am the chef

The feed is a street with ten thousand restaurants, and the person walking down it has no loyalty, no patience, and no reason to stop at mine. I cannot control the street. I cannot control who walks it or how fast.

I can control every bite on my plate. That is the one thing that is completely, totally mine — and it is enough, because the plate is the only thing on that street that has ever turned a stranger into a regular.

---

Every second is a bite. Make every bite amazing. The rest follows.