Don’t Retrofit

A faster step four

Every large organization is full of processes built for a world where information was expensive and coordination took people: twelve-step approval chains, departments that exist to move documents between other departments, meetings whose only job is keeping everyone’s picture of reality in sync.

Bolt AI onto that and you get a faster step four of twelve. The other eleven steps stay, and so does the reason they exist.

Economists call this the productivity J-curve: a general-purpose technology pays off only after companies invest in the new processes, skills, and organizational designs it requires.² Incumbents are structurally bad at that investment. Every step in the approval chain has an owner, every department has a budget to defend, and every process encodes someone’s career. A clean-slate company carries none of it. It can ask what the shortest path from a need to a delivered result looks like, given what’s possible today, and build exactly that with three people instead of three hundred.

The incumbent’s org chart is a fossil of yesterday’s costs.

Wherever a process exists mainly to manage expensive information, it is better to build from scratch than to retrofit.

Incumbents aren’t helpless, which is why this is a contest and not a rout. They own the customers, the distribution, the equipment, the permits, the data, and decades of operating knowledge, and some of their cumbersome procedures exist because someone learned an expensive lesson. Those assets are real. They are also the ones AI doesn’t cheapen. What AI cheapens is the coordination layer built on top of them, and that layer is where most incumbent cost and slowness live. The winning entrant learns to tell an obsolete habit from a hard-won safeguard, and keeps the safeguard. The winning incumbent has the nerve to build its own clean-slate replacement before someone else does. Both of them start over.

Building got cheap. Being chosen didn’t.

A clean slate lowers the cost of making things. It does nothing for the harder problem of getting found and getting trusted. Anyone can now ship a product, a report, a prototype, a brand, so the market floods with them, and the scarce thing becomes attention and the trust that turns attention into a decision.

Making things was never the hard part for most businesses. Selling them was, and AI has made that obvious.

That is why distribution and a personal brand matter more than ever. This is not vanity. It is the new moat. A body of public work tells the world what you know and how you think. An online reputation is trust at internet scale, the thing that lets a stranger take a chance on you. Fame, even niche fame inside one industry, opens doors that excellent private work never reaches.

When polished output is free, though, polish stops signaling quality. Anyone can generate a beautiful deck. Nobody can generate a track record: the shipped product, the published test results, the customer who’ll take a reference call, the public correction when you got something wrong. The brand that lasts is a visible record of delivering. Build in public, make precise promises, keep them, and let the record compound.

Your next customer isn’t human

The human customer base has a ceiling, and in much of the world it has already been hit. The UN projects the global population will peak around 10.3 billion in the mid-2080s, and 63 countries and areas, home to more than a quarter of humanity, have already passed their peak.⁴ The number of AI agents acting on people’s behalf has no comparable limit. You can spin up a thousand of them this afternoon.

Increasingly, the first party to evaluate your product will be software, sent by a person or a company to search specs, compare offers, check terms, and complete the purchase. The plumbing is going in now. Stripe and OpenAI released the Agentic Commerce Protocol in September 2025 so merchants could sell directly inside ChatGPT.⁵ In January 2026, Google launched the Universal Commerce Protocol, co-developed with Shopify, Etsy, Wayfair, Target, and Walmart and endorsed by Visa, Mastercard, Stripe, and more than a dozen others, to cover the whole journey from discovery to checkout to post-purchase support.⁶

The largest retailers on earth are rebuilding their storefronts for customers with no eyes.

Build for agents first, and for humans through them. An agent-native business is:

• Legible. Specs, prices, terms, and availability are explicit and machine-readable, not buried in a PDF or behind a “contact sales” button.
• Callable. Clean interfaces an agent can use reliably, with a record of every transaction.
• Accountable. Clear identity and permissions: who sent this agent, what it is allowed to do, and how that authority gets revoked. NIST’s newly proposed design tenets for agent identity center on exactly those questions: distinct agent identities, delegation that can only narrow authority, least privilege, and revocable access.⁷
• Humane at the edges. When something breaks, a real person can step in and fix it.

Agents multiply activity, not money. An agent has no wallet of its own; it spends its principal’s budget, so you are still serving people, through the software they’ve authorized. The businesses that understand this will capture the agent economy. The ones whose sales process still needs a human to sit through a demo will wonder where their leads went.

Here the retrofit trap closes a second time. The incumbent built its entire interface for humans clicking through forms, and adapting that for agents is the steam-engine swap all over again. A company founded today can be agent-native from day one.

The winning company of this decade will be small, sized to the work and not to an org chart (see The Floor Just Moved). It will run on processes designed around today’s capabilities instead of yesterday’s costs. It will be visible from the start, with a public record that earns trust. And it will be legible to the software its customers send.