Antonio Bustamante and the two-hour problem

Antonio Bustamante had finished onboarding a food distributor to Silo when the customer asked for a little time. There was some data entry to do. Two hours of it. Out came a stack of invoices, and the customer began typing them into the new system. Bustamante had helped build an attractive piece of cloud software. The customer still supplied the fingers.

He tells the story with the embarrassment left in. The buttons were round and blue; the interface was clean. In California's Central Valley, however, a distributor's working day remained stubbornly attached to paper. A better screen had given an old chore a more agreeable destination.

That scene sits near the center of his thinking at bem, the San Francisco company he leads with co-founder and chief technology officer Upal Saha. Their work concerns the awkward stretch between information arriving and software being able to use it. Bustamante has spent enough time on both sides of that stretch to regard another beautiful interface with a certain suspicion.

The internet arrived before the business plan

His route to California began in Spain. As a child in a town outside Barcelona in the 1990s, he tinkered with computers in a household that got internet access early. Websites, blogs and applications followed. The appeal was immediate: here was something a young person could make, put into the world and watch other people use.

There were lessons in distribution as well as construction. His teenage experiments included a blog business and advertising revenue. Legal threats also entered the picture. The internet was happy to provide an education, though it occasionally sent the syllabus on legal stationery.

His formal studies took him to the Technical University of Madrid for computer science and to Uppsala University for human-computer interaction. The combination is pertinent. One discipline asks how a system works; the other asks what happens when a person encounters it. Much of his subsequent career has unfolded in the uncomfortable space between those questions.

His early San Francisco jobs included Hattery and fuseproject, where his work crossed engineering and user experience. He then joined Kite's founding team as an engineer and designer. An AI programming assistant in the years before today's language-model boom, Kite put him close to a question he still pursues: how much useful work can software take off someone's hands?

Produce has a deadline

At Silo, which he helped found after Kite, the users were growers and food distributors. Their software had to deal with orders, inventory, payments and the movement of perishable goods. A transaction in that setting belongs to a physical chain of events. The truck leaves. The produce ages. An unanswered message can have consequences that a tidy dashboard cannot wish away.

Bustamante led engineering, product and design there. That breadth put the promise of software and the experience of using it within the same field of view. The invoice story is revealing because he makes himself part of its failure. The customer was using the product as intended. The product had left an expensive part of the job untouched.

He and Saha encountered another version of the same problem inside their own engineering work. Before useful product features could operate, somebody had to build the layer that accepted and transformed inconsistent incoming information. Then somebody had to adjust it. Then maintain it. The supposedly preliminary work had developed a talent for becoming permanent.

They carried that experience into bem. In June 2024, the pair announced a $3.7 million seed round led by Uncork Capital, with Kevin Mahaffey, Roar Ventures and Garry Tan among the participants. The initial proposition was deliberately usable by other builders: send information through an API and receive it in the format your application needs.

An argument with the assistant

By December 2024, Bustamante was writing against the industry's enthusiasm for AI agents. He argued that making one employee more efficient left larger questions about departments, handoffs and business processes unresolved. His preferred unit of ambition was the system around the worker. That is a larger promise, and a more demanding one to demonstrate.

His criticism also has a commercial logic. A reusable platform can be embedded in many products and workflows. A bespoke intervention must be rebuilt for the next customer. Bustamante's bet is that companies will assemble more of their own software from components, using infrastructure that understands their inputs and rules.

There is an entertaining tension in the vocabulary. The company is called bem, short for Brilliant Enterprise Magic. Its chief executive spends considerable time arguing that a buyer should demand explanations. Apparently, even enterprise magic ought to keep its receipts.

That insistence gives his writing its recurring subject. In an April 2025 essay, he described the hidden work organizations do to compensate for software's assumptions: people translating formats, repairing failed connections and dealing with exceptions. He wants that labor treated as a design problem. It should be visible when a product team decides what to build.

The useful part of being wrong

For Bustamante, a system's mistakes are part of its design brief. In his writing on accuracy, the relevant question is how well an output fits a particular customer's documents, requirements and decisions. A generic score cannot settle all of that. The people doing the work possess context the model needs.

He has described an approach in which corrected outputs become examples for evaluation and improvement. Operators can review results, preserve what the right answer should have been and track how a workflow changes. The aim is to make their expertise accumulate. Otherwise, a person can spend the afternoon correcting software that arrives the following morning equally uninformed.

In another essay, he took on the habit of endlessly patching prompts. A team finds an unusual input, adds an instruction and discovers that a previously successful case now behaves differently. His argument favors a recorded feedback process, with corrections tied to the work that produced the error. For a founder who has watched integrations consume engineering time, the attraction is plain.

By August 2025, he was describing production AI as a glass box: define the desired output, run the function, observe its behavior and feed corrections back. The language makes control part of the product. A customer needs to inspect what happened and decide where human attention is still required.

“AI is not a replacement for human trust.”

Antonio Bustamante

Keep the quirks. Lose the ego.

His advice to other founders is less polished than a product announcement and more revealing. He recommends emotionally mature co-founders who understand engineering. He makes room for quirks, curiosity and concentrated work. Two hours of uninterrupted attention appear on his list of things worth defending.

He also advises founders to plan their experiments, question their company's reason for existing and stop to talk when a team loses clarity. These are practical instructions from somebody who expects uncertainty to persist. A plan helps people move; a conversation can prevent them from moving confidently in incompatible directions.

There is room for retreat, too. He has publicly counseled admitting defeat and recalculating. In conversation, he describes a career with repeated failures and the need to keep building without protecting an oversized ego. The distributor with the invoices fits that outlook: a customer can puncture a founder's self-congratulation without intending to be unkind.

He shares the floor as well as taking it. Through bem's Hard Software conversations, he has spoken with builders including Temporal co-founder Maxim Fateev and MotherDuck chief executive Jordan Tigani. The interests are recognizable: infrastructure, the experience of building products and the distance between a technical idea and a working business.

What the meter should count

The product announcements he published in June 2026 show where that work is going. Data Join brings information from multiple documents into a single record, retaining a trail to the material behind each field. Agentic File System Access makes a document collection navigable by agents, with a knowledge graph connecting entities across files. His criticism of the agent metaphor leaves room for tools that agents can use. The infrastructure still has to account for where an answer came from. A shipment, after all, can have several documents telling different parts of its story.

His more recent work extends the same questions into ownership. In August 2026, he argued that a company's corrections should build an asset it keeps: the examples and models that encode its own operational judgment. The review queue becomes more than a place to clear errors. It becomes a record of what the organization knows.

In October, his attention turned to pricing. If an agent creates three short-lived helpers to complete a document task, how many software seats has the customer used? He raised the problem in an interview about AI economics. The familiar monthly license starts to look peculiar when a supposed occupant exists for ninety seconds.

He has a stake in that debate: bem charges by function call. But the question is consistent with his career. Count the work, examine the handoffs and ask whether the commercial arrangement reflects what happened. An abstraction earns its keep when the person paying can connect it to something useful.

The ambition still comes back to the distributor. Somebody has information. A system needs it. A human being spends valuable time carrying it across. Bustamante is building around that crossing, drawing on the years when he helped make the destination more attractive. The unfinished business is the stack of invoices waiting beside the screen.