A company that trades knowledge for a faster answer may discover, too late, that it automated precisely what it still did not understand.

TL;DR

AI can replace tasks, speed up activities, reorganize roles, and in some cases reduce teams. The mistake is starting an implementation by dismissing the people who understand the process. A domain expert is not merely an executor: they know the exceptions, risks, customers, informal rules, and signals that an apparently correct answer may cause a real problem. Instead of throwing that knowledge away, the company can turn the expert into the owner of an AI-assisted workflow: someone who defines criteria, reviews high-risk cases, corrects error patterns, and governs what AI may do. Cutting before learning is not digital transformation. It is outsourcing ignorance to a faster machine.

There is an idea circulating that sounds irresistible to a certain kind of executive: put artificial intelligence everywhere in the company and immediately start firing people at scale.

It must look beautiful in the spreadsheet. Salaries go out, efficiency comes in. A whole team goes out, a monthly subscription comes in. The chart rises, someone calls it digital transformation, and everyone goes home before the inconvenient question appears:

who is left to notice when the AI screws up?

That is the small detail that tends to disappear from the presentation. An employee does not merely perform a visible sequence of tasks. They know the customer who always needs an exception, the supplier who is late whenever it rains, the rule written in the manual, and the other rule that exists because the written one already caused the same problem three times. They know where the process creaks, which number looks valid but makes no sense, and when a technically elegant answer may cause a perfectly real disaster.

That knowledge does not fit entirely inside a job description. Often, the company does not even realize it exists until it fires the person carrying it.

So let us separate a few things. AI can replace tasks. It can accelerate activities. It can reorganize roles. In some situations, it can reduce the number of people needed to run a process. Denying that would merely replace alarmism with naivety, and I am not interested in either.

But concluding that the first intelligent move is to dismiss the process expert is operational stupidity. The company throws away the very person who could turn a generic tool into a real capability.

A task is not a job

When someone says AI “does a person’s work,” they are usually compressing several different things into one convenient sentence.

A task is writing an email, comparing documents, classifying requests, filling in a system, consolidating a spreadsheet, or suggesting a response.

A role combines several tasks, decisions, relationships, and responsibilities.

A job is the organizational form a company chose to group some of those things.

And responsibility is what remains when the automated answer is wrong and someone must explain it to the customer, the board, legal counsel, their own team, or—depending on the automation’s creativity—all four at once.

AI models can perform or assist with a growing range of tasks. That does not mean they have automatically absorbed the context, judgment, and responsibility of an entire role.

The International Labour Organization reached a more sober conclusion when it examined the exposure of almost 30,000 tasks to generative AI: systems capable of producing text, images, code, and other content from instructions and examples. Because most occupations still combine activities requiring human input, job transformation is currently more likely than full automation.

That does not prove no position will disappear. The research does not claim that. It supports a narrower point: treating occupations as indivisible blocks hides what the technology actually reaches—different tasks, to different degrees.

That detail changes the management question completely.

Instead of “how many people can I cut?”, the useful question becomes:

Which tasks can AI take over, which criteria must remain human, and how will we use the time it frees to improve the outcome?

It is a less cinematic question. It is also much harder to place in a post on LinkedIn, a professional social network, next to a rocket emoji. We will survive.

The expert is not merely the person who executes

A mature company should not look at an experienced professional as an expensive pair of hands that AI has finally made disposable.

That professional is a source of domain knowledge.

They recognize exceptions before they become incidents. They distinguish an unhappy customer from a customer at risk. They notice when a data point is formally valid but incompatible with the rest of the case. They know when to follow the procedure and when to interrupt it because reality had the poor manners not to read the flowchart.

Some of that knowledge can be documented and converted into rules, tests, examples, acceptance criteria—the conditions a result must satisfy—and escalation paths—the route a case follows when it needs human or higher-level review. Excellent. That is precisely what should happen during a responsible implementation.

But it does not happen by osmosis—much less because someone bought a platform with “AI agents” written in large letters. An agent is software that combines AI, tools, and a sequence of actions to pursue a goal; it is not an intelligent entity that learns the company through divine revelation.

AI needs to be placed inside a system. That system must know where data comes from, what is allowed, how a decision can be verified, when uncertainty must be exposed, and when the machine must stop and call a person.

Who helps build that? The expert who already understands the work.

If the company fires that person first and tries to capture the knowledge later, it is not automating a process. It is trying to interview a ghost.

From executor to capability owner

The most interesting role for that professional is not to keep manually repeating everything the machine can do well. It is not to become a “prompt engineer” either—someone dedicated to designing, testing, and refining the instructions given to AI—simply because someone decided every business problem now needs a new job title in English.

They can become the owner of the AI-assisted workflow.

In practice, that means:

• defining what counts as an acceptable response;
• providing context and real examples;
• reviewing higher-risk cases;
• identifying error patterns;
• turning recurring corrections into rules, tests, or system improvements;
• deciding when AI may act and when it should only recommend;
• monitoring quality indicators, not only speed;
• preserving human responsibility where it remains necessary.

The machine executes part of the work. The professional governs the capability.

That can be a major change. Someone who spent much of the day copying data between systems can investigate discrepancies. Someone writing repetitive responses can focus on cases requiring negotiation, emotional judgment, or deep customer knowledge. Someone consolidating reports can use the time to challenge assumptions and improve the decisions made from them.

That is the gain I actually care about: not doing the same thing with fewer people as a matter of principle, but doing better the things the company never had enough time, clarity, or capacity to do.

Relationships. Prevention. Quality. Learning. Analysis. Innovation. Responsible decisions.

Everything usually announced as a priority and abandoned by Tuesday because operations caught fire again.

The best worker can also teach the machine

A field study involving more than five thousand customer-support professionals found average productivity gains when workers received AI assistance. The gains were larger among less experienced workers. The mechanism matters: the system had learned patterns from successful conversations and offered suggestions that agents could use, adapt, or reject. The full paper was published by the National Bureau of Economic Research (NBER).

This study is not a prophecy for every company. It examined a specific context, tool, and set of metrics. But it illustrates something important: technology can spread the knowledge of strong professionals without the company presenting the system as the independent author of that knowledge.

The study did not examine layoffs or test whether the strongest professionals needed to remain at the company. My synthesis from the mechanism it observed is different: if the system depends on patterns produced by competent people, removing them before establishing a continuous learning cycle may reduce the company’s ability to update what separates an acceptable answer from an excellent one. That is the irony that should worry anyone dreaming of mass layoffs.

For a while, the machine may continue to look intelligent. It reproduces what it learned yesterday. The problem begins when the market, product, customer, regulation, or situation changes.

Without experts close to the workflow, a company can keep receiving fast answers while losing the ability to notice that those answers have grown stale.

AI increases speed. It does not choose a better direction

I have used another metaphor for this: giving AI to an operation without direction is like pointing a Ferrari at a wall and pressing the accelerator.

The technology does not automatically correct the route. It merely brings the impact forward.

A confused process remains confused after automation, now at scale. An unfair rule remains unfair when executed by software, now producing the unfair outcome faster. A bad data point distributed by agents reaches more places before anyone finishes their coffee.

That is why productivity cannot be measured only by completed tasks. We also need to observe rework, exceptions, incidents, satisfaction, reversed decisions, avoided losses, and outcome quality.

The Organisation for Economic Co-operation and Development (OECD) gathers evidence and scenarios pointing in both directions: AI can automate routine tasks, free time for more meaningful work, and raise productivity; it can also reduce headcount, put pressure on workers, and damage job quality when adopted without safeguards, training, and participation.

That is the adult part of the conversation. The gains are real. The risks are real. Direction does not come preinstalled with the model.

Not every team reduction is forbidden

It would be comforting to end by claiming that every company can preserve exactly the same structure by promoting everyone to AI manager. That is not true.

Some processes will need fewer people. Some jobs will change deeply. Certain roles may disappear while others emerge. Companies will also use added productivity to reduce costs—sometimes out of necessity, sometimes by choice.

What I am criticizing is not every headcount reduction. It is using layoffs as the first success metric for an implementation that has not yet demonstrated quality, reliability, or sustainability.

Before deciding that a human capability has become redundant, a company should be able to answer:

1. Which tasks were actually automated?
2. How is quality measured and compared?
3. Who reviews high-risk cases and exceptions?
4. What knowledge still exists only in people’s heads?
5. What happens when data, rules, or conditions change?
6. Who can stop the system?
7. Where will the freed time be reinvested?
8. Is the reduction the result of a learned process, or merely a financial target dressed as innovation?

Without answers, the layoff plan is not strategy. It is a bet made with someone else’s operational knowledge.

Adapting people also takes work

There is another common mistake: declaring that employees are now “AI managers” and considering the transformation complete.

It is not enough to hand out accounts, run a two-hour lecture, and ask everyone to become innovative on Monday.

People need training connected to their real work. They need to understand limits, security, privacy, validation criteria, and responsibility. They need suitable tools and enough safety to report failures without punishment because the automation dashboard became less attractive.

They also need to participate in redesigning the process. The people doing the work know bottlenecks that rarely appear in a board meeting. Ignoring them and hiring an external team to “put agents everywhere” is a fine way to build an impeccable solution to the wrong problem.

Adaptation is not teaching every employee half a dozen magic commands. It is turning experience into operational governance.

The company gains more when it preserves the ability to think

I am not defending repetitive work merely to preserve the appearance of busyness. Quite the opposite. Much of what I build exists to remove useless friction, coordinate tasks, consolidate context, and free people from carrying entire processes in their heads.

But freeing someone from a task does not make that person useless.

It may give them time for what never fit into the workday: talking to a customer properly, investigating the cause of an error, improving a product, documenting knowledge, training someone, noticing an opportunity, or questioning a decision the data appears to support but reality contradicts.

A company that uses AI only to reduce staff will probably capture part of the gain. A company that uses AI to expand the capacity of people who understand the business may discover gains that were not even visible before implementation.

That requires a change in posture: the employee stops being treated as a cost attached to tasks and is recognized as a carrier of judgment, context, and responsibility.

That is not romanticism. It is knowledge architecture.

Learn before you cut

If AI is going to take over part of someone’s work, that person should be at the center of the implementation.

They help map the workflow, expose exceptions, build examples, define metrics, and identify what may or may not be delegated. Then they monitor the system, correct failures, and turn lessons into reusable improvements.

Only then does the company begin to understand its new capability. It may decide to expand the service, raise quality, absorb more demand, shorten lead times, create new offers, or—yes—operate a process with a smaller structure.

The difference is that the decision comes after learning, not before it.

AI does not have to replace the expert. It can turn the expert into the manager of a capability they previously executed almost entirely by hand.

Firing that person before capturing their judgment is not digital transformation.

It is outsourcing ignorance to a faster machine—and hoping nobody notices before the next quarterly report.

Start with the real work

At i-9.ai, I do not begin with a promise to replace a team. I begin with the process, the bottleneck, and the people who know where operations fail.

If you want to discover which tasks can be automated without throwing away the knowledge that sustains your company, get in touch. The first conversation can begin with what consumes time today—and what must not be lost.

Continue reading

• Your company does not need to discover where to put AI: a practical map separating process, automation, assistive AI, agents, and human decisions.
• Generic agents can be your worst first contact with AI: why a ready-made solution does not know your operation’s exceptions, risks, or culture.
• The best answer is not the one that pleases me most: how to design systems that disagree, show limits, and return verifiable evidence.
• Written but not read: AI does not sign for you: the responsibility that remains human when AI participates in execution.

Learn more

• Tacit knowledge: practical knowledge that is difficult to convert fully into rules and manuals.
• Human-in-the-loop: designs in which people continue participating in system decisions, reviews, or corrections.

References and limits of use

• ILO, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” (2025): supports the task-based analysis and the conclusion that occupational transformation is more likely than full automation in the scenario studied. It does not predict the fate of a specific company or job.
• Brynjolfsson, Li, and Raymond, “Generative AI at Work” (National Bureau of Economic Research, Working Paper 31161): a study of 5,179 customer-support agents that found heterogeneous productivity gains with AI assistance. Its specific context does not justify transferring the percentages to other sectors.
• OECD, “Beyond automation: Decoding the impact of Generative AI on regional labour markets” (2024): gathers evidence about task acceleration, work reorganization, new skill requirements, and the risk of headcount reductions. It does not prescribe the organizational strategy argued for in this article.

These recommendations are an authorial synthesis about organizational architecture and responsible AI implementation. They do not replace employment, legal, financial, security, or impact analysis specific to each company.