What AI won't learn from you

A mid-size Brazilian trucking company (figures in Brazilian reais), 540 people, its own fleet plus contracted drivers. Load scheduling is done by three people, and one of them, with 19 years at the company, handles the hard cases. Asked how she decides, she gives answers that are true and useless: 'it depends on the customer', 'that one doesn't accept delivery after 3pm', 'when it's that route I don't send the new driver'. None of those sentences is written down anywhere.

The company tried putting an agent in charge of proposing the schedule. The agent had access to everything in the system — orders, distances, availability, cost. It got wrong exactly the cases the 19-year veteran solved in thirty seconds. That's not a model failure. It's a problem with a name since 1966.

01Polanyi's paradox #

Michael Polanyi published The Tacit Dimension in 1966, condensing into one sentence the argument that runs through the book: we know more than we can tell. Tacit knowledge is what's demonstrated in doing and can't be transmitted by description — recognizing a face, riding a bike, sensing a customer is about to complain before they do.

David Autor brought the idea into labor economics in 'Why Are There Still So Many Jobs?', published in the Journal of Economic Perspectives in 2015. His argument: tasks that depend on tacit knowledge resist automation not because they're complex, but because we can't specify the rules we follow when doing them. Autor named this Polanyi's paradox.

An honest caveat, which Autor himself anticipates: machine learning partly works around the paradox, because it learns from examples rather than rules. Where there are many labeled examples, the barrier falls. Where there aren't — and in load scheduling at a mid-size trucking company there aren't — it stays exactly where it was.

02Why this decides the agent's outcome #

An agent decides based on what it perceives. What it perceives is what's recorded. At the trucking company, the system recorded the order and the delivery; it didn't record that customer X refuses late-afternoon delivery, that route Y has a stretch a new driver doesn't handle well, or that a particular contracted driver is always late on the first business day of the month.

None of those three pieces of information is a secret or a complexity: they're operational tacit knowledge, kept in three heads. As long as they stay there, no model discovers them — and the agent will keep proposing schedules the experienced person rejects in thirty seconds, which quickly ends the project through loss of credibility.

The operational conclusion is direct and not very exciting for anyone hoping for a shortcut: AI performs in proportion to the context the company manages to make explicit. The expensive part of the project isn't the model; it's the conversation with the person who knows.

03What can be made explicit, and how #

Nonaka described, in his 1991 Harvard Business Review article, the move from converting tacit knowledge into explicit knowledge — and the condition for it: proximity, observation and dialogue, not a form. Translated to the trucking company, it worked like this.

1. Watch the decision happen. Someone sat next to the scheduler for four shifts and wrote down every exception with the reason spoken out loud. Thirty-one exceptions in four days.
2. Ask about the case, not the rule. 'Why can't that one?' gets an answer; 'what are your criteria?' gets nothing.
3. Turn exception into field. Twenty-two of the thirty-one exceptions became recordable attributes: delivery window per customer, route restriction by experience level, on-time history per contracted driver.
4. Leave the remaining nine with the person. These are the ones that depend on reading the situation. The agent's scope was designed to stop right there.
5. Record at the source. The new attributes started being filled in at the moment of the decision, not in a parallel record that goes stale in two months.

Want to see how this looks inside a real operation? Explore the platform.

The fourth item is the most important and the most ignored. Trying to make everything explicit produces false documentation: the person writes what they think they should do, not what they do. Accepting that a part stays tacit is what makes the rest trustworthy — the same care described in what only Paulo knows.

It's worth noting the practical limit here. Making knowledge explicit is work done by busy people, and explicit knowledge ages: the credit rule written in March stops applying when the commercial policy changes in July. That's why the record only holds up when it has an owner, a review date and one single place where it lives — ideally the same place where the process runs, not a file nobody reopens. When the record sits outside the operation, it turns into dead documentation within months, and the company goes back to depending on one person's memory. What survives is what's used every day, corrected through friction, and maintained by whoever does the work, not by whoever writes it down.

04The math on context #

Stated assumptions, at the trucking company's scale — redo it with your own numbers:

Four days of observation changed the agent's performance more than any model swap would have. It's an unglamorous result, and it's the typical one. It's also worth logging that context's provenance: who reported it, when, and with what validity — the practice of documenting the origin of the data feeding a system, argued for by Timnit Gebru and colleagues in 'Datasheets for Datasets' (Communications of the ACM, 2021), applies equally to a trucking company's operational context.

05What to do this week #

At Relevanti, the context an operation makes explicit lives in the same place the agent works, instead of a separate document — see the platform, the breakdown by department under solutions, or bring your most single-person-dependent decision to a conversation.

Polanyi wasn't writing against automation: he was writing about what human knowledge is. Sixty years later, his sentence has become the most practical yardstick for AI projects — the agent learns what you managed to say, and nothing beyond that.