Work

AI Will Not Replace People. It Redraws the Job.

Most arguments about AI and employment use the wrong unit. Professions are not automated; tasks are, and a job is a bundle of tasks that can be rebundled without the title changing at all. What follows is the mechanics of that rebundling, on both sides of the employment contract.

Replacement is the wrong unit of measurement

A profession is a bundle of tasks held together by a job title. An accountant reconciles ledgers, explains a number to a director, decides what to do when the paperwork does not match reality, and signs off on something that carries a penalty if it is wrong. Automation does not arrive at the level of that bundle. It arrives at the level of the individual task.

So the honest question is never whether AI replaces the accountant. It is which of those four things it takes, what that does to the value of the rest, and how many people the new bundle requires. The distinction cuts both ways: a role can survive as a title while losing most of what made it well paid, and a role can become more valuable because the tedious half is gone.

What AI takes first

The tasks that move earliest share a shape, regardless of industry or seniority.

• They repeat. The same operation on different inputs, many times over. Repetition is what makes the investment pay back and what made the task describable at all.
• They have a knowable correct answer. Extracting a date from a contract, classifying a support ticket, turning a specification into routine code. The output can be checked.
• They involve volume nobody enjoys. Reading four hundred documents to find the six that matter, or producing the first draft that will be rewritten anyway.
• They are bounded. Everything required is contained in the task, with no need to walk down the corridor and ask what the client actually meant.

Notice what is absent from that list: difficulty. Drafting a competent contract clause is hard, and an LLM does it in seconds. Handling a client who is angry for reasons unrelated to the contract is not hard, and no model does it at all.

What does not move

Responsibility does not transfer. When a decision carries financial, legal or reputational consequences, someone has to own it. A system can produce the recommendation; it cannot be the party that answers for it. That is not a limitation of the technology, it is what accountability means.

Judgement under incomplete information transfers badly. Models are strong when the inputs are present and weak when the decisive input is what nobody wrote down: that the supplier is quietly in trouble, that the deadline is political, that the brief is not what the person actually wants. Much of senior work is the handling of what is missing from the file.

Negotiation and work with people stay because their content is not information transfer. Getting two departments to agree, holding a client through a bad quarter, telling someone the answer is no: here the other party being human is the entire point. And defining the problem stays. Systems answer questions; deciding which question is worth answering sits upstream of any tool.

What happens to a role when half of it is automated

The usual outcome is not deletion but displacement upward. When the production part of a job is handled by a machine, the human part moves toward specifying what should be produced, checking what came back, and carrying the consequence. A translator becomes an editor and a guarantor of meaning. An analyst spends less time assembling the report and more deciding what it should measure and whether the numbers hold.

Two consequences follow. The new bundle needs fewer people for the same volume, because one person supervising output covers more ground than one producing it. And it demands more experience per person while removing the junior tasks that used to build that experience. Organisations that ignore the second point discover it a few years later, with nobody ready to promote.

Roles that did not exist before

New work appears around automated systems, and none of it is exotic.

• People who describe processes precisely enough to automate them. Most processes exist only as habit. Turning habit into an explicit sequence with defined inputs, edge cases and exceptions is analytical work, and it is the bottleneck in nearly every automation project.
• People who own the quality of what the system produces. Someone has to define what a good answer looks like, build the checks, and notice when accuracy drifts.
• People who set the boundaries. What the system may decide alone, what requires a human signature, which data it may touch, what it must refuse. This sits between legal, risk and engineering, and is becoming a role rather than a paragraph in a policy.

None of these are jobs about models. They are jobs about process, evidence and limits, which is why the people who understood the business first tend to fill them.

What it means for a specialist

A skill stops being scarce when a machine supplies it cheaply. That is the whole mechanism, and it applies to skills that were reliably valuable for years: producing clean standard output, remembering how something is done, writing the routine version of a document.

What becomes scarce is the inverse. Stating a problem precisely enough that a system or a team can act on it. Telling a good answer from a plausible one in your own field, which takes domain knowledge rather than tool knowledge. Putting your name on a decision. Being effective with people under no obligation to agree with you. The practical move is not to collect tools as a hedge, since tools change. It is to move toward the part of the work where you decide and answer, and let automation take the part where you type.

What it means for a company

The reflex is to treat automation as a headcount question, and organisations that lead with cuts tend to lose twice. The first loss is capacity: removing people at the moment output per person rises means banking the saving and forgetting the growth. The same output from a smaller team is one option; more output from the same team is usually worth more, and that option disappears with the team.

The second loss is knowledge. The people who can describe how the work actually happens are the people doing it. Cut them before the processes are documented, and the programme loses its only reliable source of requirements.

What tends to produce more: choose a few processes with real volume and a tolerable cost of error, keep a named person accountable for each automated decision, redeploy the freed hours into work that was starved, and measure the result on a business number rather than an adoption metric. Redeployment is the harder management task, which is why it gets skipped, and where the return sits.

Where jobs genuinely disappear

Honesty about the mechanism requires honesty about its edge. Where a role consists almost entirely of automatable tasks, bounded, high in volume, with a known correct answer and little judgement attached, that role does contract. Usually not to zero, but to fewer people supervising instead of producing. Entry-level document handling, first-line support of routine questions, basic content and code at volume: these bundles are thinning, and telling the people inside them otherwise is not kindness.

What serious companies do about it is unglamorous. They say which roles are changing and on what horizon rather than letting people infer it. They move people toward the supervision and exception work the same automation creates, retrain ahead of the change rather than after it, and rebuild an entry path so people still acquire judgement without the junior tasks that used to build it.

The shape of the change

AI does not remove people from work that carries consequences. It removes the mechanical parts of that work and moves the human contribution toward specification, verification and responsibility, raising the bar per person and lowering the number of people needed for a fixed volume of output. Neither a catastrophe nor a non-event: a redistribution that rewards specialists who move toward judgement and companies that spend freed capacity rather than pocket it.

AKVANT Technologies is an AI consulting and engineering firm. We work with companies to find where AI pays off — and build only there.