Skip to content
Back to the blog
  • Inteligência Artificial
  • Futuro do Trabalho
  • Tecnologia e Sociedade
  • Automação
  • Mercado de Trabalho
  • Ética na Tecnologia

Tacit Knowledge Became a Company Asset

Who Keeps the Value?

In 1911, Frederick Taylor wrote that the knowledge built up in a trade was "the principal asset" of every worker. His proposal was to take that asset away from the worker and hand it to management. In 2026, we do the same thing under another name: we take the way a person works and turn it into a skill that any colleague, or any AI agent, can run. I know this from the inside, because part of my job as head of AI Enablement is exactly that. In this post I go through what Taylor did with the factory worker's knowledge, why the skill repeats that move and goes further, who keeps the value of what gets taught, what this does to junior, mid-level and senior roles, and what we can demand from companies.

What Taylor Did with the Worker's Knowledge

At the beginning of the 20th century, the person who knew how to do the work in a factory was the worker. The foreman knew less. Taylor himself admits it in The Principles of Scientific Management: foremen and superintendents knew, "better than any one else", that their own knowledge fell far short of the combined knowledge of all the men under them.

That knowledge had a name in the book. Taylor called it "rule-of-thumb": practical know-how, passed from person to person, that nobody had ever written down. And he wrote the sentence I borrowed for the title of this post:

"This mass of rule-of-thumb or traditional knowledge may be said to be the principal asset or possession of every tradesman."

His solution was to change who owned that asset. Management would take on the burden of gathering all the traditional knowledge of the workers and of "classifying, tabulating, and reducing this knowledge to rules, laws, and formulae". Each worker would then receive, at least one day in advance, complete written instructions describing the task and the way to do it. Taylor put it in one line: "In the past the man has been first; in the future the system must be first."

Knowledge you can't put into words

Decades later, the philosopher Michael Polanyi gave a name to what Taylor wanted to capture. In The Tacit Dimension, from 1966, he starts from a simple sentence: "we can know more than we can tell". That is tacit knowledge. The mechanic who listens to an engine and knows what is wrong, the support agent who can tell from a customer's tone that they are about to cancel, the developer who looks at a log and already suspects the cache.

That knowledge protected a lot of people for a long time. The economist David Autor called it "Polanyi's Paradox": because our tacit knowledge goes beyond what we can explain, it was hard to write a program that did the same thing. What didn't fit into a rule didn't fit into a machine. Generative AI, and the skill written alongside it, go straight at that protection.

In 1974, Harry Braverman read Taylor with different eyes in Labor and Monopoly Capital. For him, the method had three principles: separate the labor process from the worker's skills, separate the people who conceive the work from the people who carry it out, and use control over knowledge to control every step of the work. Braverman also showed that the office was going through the same process as the factory. The technique was never only about productivity. It was about who runs the work, because whoever holds the knowledge has bargaining power.

Hold on to that idea of separating the people who conceive from the people who execute. It came back, and it came back through a road Taylor never had.

The Skill Is the New Manual, and Now It Reaches Reasoning

If you don't use them yet: a skill is a package of instructions, examples and scripts that teaches an AI agent to do a task the right way. Anthropic, which created the format and later opened it as a standard, describes the skill as a way of "packaging your expertise". I already mentioned in AI Skills by Area that Anthropic compares a skill to the onboarding material for a new hire. The detail that matters here is a different one: that material used to be written to teach a person, and now it is written to teach the agent that does that person's job.

The difference is that Taylor's manual stopped at the body. He timed movements, shovels, ore wheelbarrows. Even with all the changes in management that came after him, what could be extracted was the operational part. With skills, you can standardize a task that depends on reasoning and make sure it is always done the same way. For example: the playbook a senior analyst follows to investigate an incident, the criteria a QA uses to say a card is done, the way someone in sales reads a proposal.

The company that sells the format measured its effect. The January 2026 Anthropic Economic Index shows that the tasks AI covers most are precisely the ones that require more education, and concludes that the net first-order impact is "to deskill jobs". The report also shows the other side: some occupations get upskilled, because AI takes the routine part and leaves the more complex one.

That is why I think the skill is the same move Taylor made, and maybe a worse one. When what gets standardized is reasoning, people lose mental effort, autonomy and room to create. They end up executing what someone else thought through before them.

The move already has a name on GitHub

In April 2026, MIT Technology Review told the story of a Chinese project that went viral on GitHub, Colleague Skill. It started as a parody: it imports a colleague's chat history and files and generates, from that, a manual for an agent to do their job. The article reports that bosses were asking employees to document their own work so it could be automated. The response came fast. A product manager published an "anti-distillation" skill, designed to sabotage that kind of extraction.

Taylor also had a word for resistance: soldiering, deliberately holding back the pace so management would never find out how much a day's work could really produce. A hundred and fifteen years later, soldiering became a repository.

It depends on where the skill lands

None of this means every skill is bad. It depends a lot on the place. Some companies use skills the right way: they automate the operational part, the boring and repetitive side of the day, and give people time back to think and build new things. Other companies only need a delivery factory, where quality and innovation matter less than volume. In those, the skill becomes a deskilling tool: people know less and less and just execute the next task.

Even people on the technical side have spotted the risk. The Thoughtworks Technology Radar from November 2026 put an antipattern called "agent skill cognitive surrender" in its Caution ring: teams that adopt huge libraries of ready-made skills without understanding what they do. Cognitive surrender is almost a technical name for what Braverman described on the factory floor.

If the skill concentrates knowledge, the next question is unavoidable: whose knowledge is it once it becomes a file?

Who Keeps the Value of What Was Taught

That question has no single answer, because not every skill carries the same amount of a person. I have two very different examples.

The first is by far the most used skill at the company where I work. It operates our project management tool: it creates cards, comments, updates statuses. I didn't build the first version. Someone created it on their own initiative, without anyone asking, and I have maintained it ever since. At first it was basically a guide to using the tool's API. The writing and communication standards came in later. And even those standards already existed in the team before they became a skill. So how much of it belongs to one person? Not much. The skill generates a lot of value, but what it encodes is a collective way of working.

The second example is the opposite. We have skills created by professionals who do outstanding work. They documented their own process of thinking and executing, built a workflow and turned it into a skill. That is individual. A few years ago, nobody would have had access to that method without mentoring or training from that person. The method was theirs.

And here is the uncomfortable part: in most cases, the company doesn't need a specialist. It needs the result. The person still knows how to do everything from scratch, without AI, but the rarest thing they had is now available to everyone, including whoever will take their seat.

Research already asked the question

This shift has been measured. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed more than 5,000 support agents who started using an AI assistant, in a study published in the Quarterly Journal of Economics in 2025. Productivity went up, and newcomers gained the most. The authors' explanation is almost the theme of this post: AI seems capable of "capturing and disseminating the behaviors of the most productive agents".

The best agents supplied a good share of the examples used to train the system, and they were the ones who gained the least from it. The economists themselves write that their work raises questions about whether and how workers should be compensated for the data they provide. They also point to a risk: with AI suggesting everything, the best people may stop coming up with new solutions. Without people inventing, the next skill has nowhere to come from.

"You're already paid for that"

The most common answer in this debate is that the person already gets a salary, so whatever they create at work belongs to the company. With all due respect, that's a load of nonsense. Your salary pays you to do your job. If your job is checking invoices, turning your method into a skill was never part of the deal.

There is an exception, and I am it. I work on AI adoption, so writing and maintaining skills is part of my role, and part of my salary pays for that. Other people have no such obligation. When someone in finance, QA or sales packages their own method, they are delivering more than the role asks for.

Brazilian law, by the way, doesn't protect methods. Article 8 of the Brazilian Copyright Act says that ideas, systems and methods "as such" are not subject to protection. The text of the skill can belong to the company. The way of thinking inside it still belongs to whoever did the thinking. The bare minimum is that people can take their own method with them when they leave, as long as it carries no company secrets. The company keeps the copy. The person keeps what was always theirs.

Talking about who keeps the value without talking about who does the extracting would be dishonest of me. So let's move to my side of the table.

I Hold the Stopwatch Too

In Taylor's factory, the person who wrote down the worker's knowledge was the time-study man. He watched, measured and turned what he saw into written instructions. When I look at my own work honestly, the resemblance is there.

Part of what I do is help people extract the way they get things done and turn it into a skill. Another part is measuring: whether the skills are being used, and whether they are being used the right way. I explain how we organize and publish these skills in Skills Hub.

The difference I can defend is in what I measure. I want to know whether a skill is useful, whether it saves work and whether it needs improving. I don't want to monitor who uses it or build a ranking of people.

The difference I can't guarantee is what happens next. I train people and I measure, but I'm not the one who decides what to do with it. The people above me decide. The same number that shows a skill saved ten hours a week can become free time for the team to think, or it can become an argument for not backfilling a position. I already argued which of those two paths I think is right in AI Governance in Practice. But arguing is not deciding, and anyone reading this post deserves to know that.

That math weighs heaviest on people who are just starting out. And that is where a question comes in that I think few people are asking the right way.

Junior, Mid-Level and Senior Were Built for a Different Job

The conversation about AI and careers usually stops at hiring juniors. And the data is worrying. I already wrote in AI and the Future of Professions about the Stanford study on workers aged 22 to 25. The August 2026 revision pushed the gap to 19% below expected, almost all of it from less hiring rather than from layoffs. The authors warn that these are early indicators, not proof of causation. A Harvard study using data on 62 million workers found a similar pattern: at companies that adopted generative AI, junior employment fell and senior employment kept rising.

Not even the industry agrees on what to do. Matt Garman, CEO of AWS, said that replacing juniors with AI is "the dumbest thing I've ever heard", because ten years from now nobody will have learned anything. Tobi Lütke, CEO of Shopify, told teams to prove AI can't do the job before asking for more headcount.

But I think the bigger question isn't whether companies will hire juniors. It's what "junior" will mean.

Categories of autonomy

Junior, mid-level and senior were created to rank people by how far they can go on their own in operational work:

  • Junior: helps the company, but still needs help to do the day-to-day work.
  • Mid-level: does it alone and no longer needs support.
  • Senior: understands enough to guide other people.

This logic is old. The Dreyfus brothers' model of skill acquisition, from 1980, goes from the novice, who relies on "context-free rules" and step-by-step instructions, to the expert, who acts on intuition and often can't explain what they do. Notice that the context-free rule, the lowest rung on the ladder, is exactly what a skill delivers ready-made. And the expert who can't explain what they do is Polanyi's tacit knowledge all over again.

The career handbook at 37signals, a company known for having little hierarchy, makes this even clearer. Its level matrix has an axis called Trust, which measures how much a person's work needs review: at level 1, everything is "thoroughly reviewed"; at level 4, work is "shipped without review". The document says nothing about AI. But that is exactly the axis AI scrambles, because today a large part of what needs review was written by an agent, not by a person.

The operational work that ranked people is disappearing

Models get better, harnesses (the environment built around the agent, with context, memory and checks) get more advanced, and better and better tools for security and quality keep showing up. The operational part that separated one level from the next is changing very fast.

At a retreat with senior engineers from large companies, Thoughtworks recorded the emergence of a "middle loop", supervisory work that sits between writing code and shipping it: directing, evaluating and fixing what agents produce. These are skills that, according to the report, are rarely "recognized in career ladders". The same text says many people were hired to translate pre-digested tickets into code, and that "that work is disappearing". The biggest risk, in their view, isn't even the junior. It's the mid-level. Martin Fowler, who attended the retreat, summed it up: "It's the mid-level developers who face the greatest challenges."

It makes sense. If mid-level was defined by "does it alone", and the agent also does it alone, the definition loses its footing. GitLab admitted something similar in an official document: its current interview process primarily measures skills that "AI has commoditised", and they are redesigning hiring to assess judgment and validation discipline. Addy Osmani described the effect on people starting out in one sentence: "Taste used to be a byproduct of the reps. Agents took the reps."

Nobody is proposing to throw out the ruler

The curious part is that, as far as the research for this post could see, nobody with real weight is proposing to abolish junior, mid-level and senior because of AI. What exists is people asking to redefine the levels. In the abstract for his talk at QCon London this year, Alasdair Allan wrote that progression frameworks, built for a world of gradual skill accumulation, "no longer match reality". Ossama Chaib called for defining what each level means in a world with agents and readjusting the pay bands. Deloitte was already arguing in 2022 that jobs are not the best way to organize work, but it kept levels in its own proposal, and AI was only the third reason on its list.

In Brazil, that change is even harder. Under article 461 of the CLT, Brazil's main labor code, a formal career structure or a job and salary plan is what lets a company pay people in the same role differently without running into equal pay claims, and that plan can come from internal rules or from collective bargaining. Changing the categories means changing the plan.

I don't have the authority to say how this will turn out. But I believe it will change, because the work is changing and the way we classify work will have to keep up. Operationally, an AI with access to all of a company's documentation knows more than any single person there. That's a fact. Whether it does things the best way is another conversation.

The question I want to leave is about the expert. Will the expert be the person who carries out a task brilliantly? Or the person who reinvents how to use the tools the company already has, the ones AI already masters, and turns that into a process others can repeat? The Dreyfus brothers had a sixth stage, above the expert, for those who "transform the style of their skill domain". If that is the answer, the expert of the future is precisely the person who writes the skills. And then the question of who keeps the value stops being a side issue and becomes the central question of a career.

What We Can Demand

If the problem is structural, the way out can't depend on the goodwill of each manager. I see three paths, from the simplest to the hardest.

1. Authorship, at the very least

Authorship and recognition are the basics. Whoever created a skill has to be able to say on LinkedIn and in their portfolio that it is theirs, and the company has an obligation to confirm that the achievement belongs to the person, not to the company. That includes the right to take the method with them, as I said above, when there are no company secrets involved. In The Developer's Manifesto I wrote about signing what we make. A skill is one of the things that deserve a signature.

2. Rewards proportional to use

A skill with heavy adoption deserves a bonus, as something extra. And someone who keeps delivering new methods all the time is delivering far more than the role asks for, so the fair thing is a raise.

Measuring this is hard, and I'm not going to pretend it isn't. How many hours did a skill save? How much rework did it prevent? How much money did it bring in? Does the reward come as time, a percentage, a raise or a bonus? These are open questions. But there are models to copy:

  • In Germany, the Employee Inventions Act guarantees "adequate" compensation when a company uses an employee's invention, and calculates the amount based on economic use, the person's role and the company's share in the result. It even applies to technical improvements that can't be patented, as soon as the company starts using them. It is almost the yardstick I described above: the individual weight varies, and whoever has that as part of their job gets less.
  • In Brazil, the Industrial Property Act, in article 89, already allows a company to give an employee inventor a share of the gains from a patent, through negotiation or internal rules, without that counting as salary. The design exists. It just hasn't reached skills.
  • At one company, Omnisend started giving quarterly raises of 2% to 4% to people who use AI with business impact, including processes that other colleagues and teams go on to adopt. The executive himself admitted to Fortune that measuring that impact is "honestly really difficult". The difficulty didn't stop the company from trying.

3. Collective agreements

The strongest path is the one that doesn't depend on a single company. In 2023, Hollywood screenwriters went on strike and came out with a contract stating that AI-generated text doesn't count as literary material, that AI-generated material can't be used to reduce a writer's pay or credit, that the company can't force anyone to use AI, and that the union reserves the right to challenge the use of writers' work to train AI. Authorship, compensation and collective bargaining in the same document.

For people working in tech in Brazil, this still seems far off. But the knowledge that becomes a skill is the same kind as the knowledge that becomes a screenplay: it is the way someone thinks about the work. If screenwriters managed to put that in a contract, we can put it in collective agreements and in job and salary plans too.

The Asset Still Belongs to Whoever Knows

Taylor was right about one thing: the knowledge of a trade is the principal asset of the people who work. The difference is that, in 1911, taking that asset required a stopwatch, a foreman and years of observation. Today all it takes is a text file and an afternoon. And what goes into the file now isn't just the movement of the hand. It's the way of thinking.

That doesn't make the skill an enemy. Used to clear boring work out of the way, it gives time back and opens room for new things. Used as an assembly line, it teaches people to know less and less. The technology is the same in both cases. What changes is who decides and who keeps the result.

The skill can stay with the company. The knowledge has to keep belonging to whoever knows.

References

Matheus Mierzwa

AI Enablement Lead & Frontend Tech Lead

I write about what I build and lead in practice.