What does the future of work look like?

Watch a commercial credit reviewer work through a file and you are watching two jobs happen at once. The first job is reading. She pulls three years of financials, an appraisal, a tax return, a page of covenants, and a folder of correspondence, and she reconciles them against each other and against the credit policy: does the debt service coverage clear the threshold, does the collateral value hold, does the borrower's story on page 4 survive the numbers on page 190. This is most of the hours. The second job takes about ninety seconds. Having read everything, she decides. She assigns the risk rating, she puts her name on it, and if the loan sours in two years it is her judgment the workout committee pulls up. Same person, same desk, same afternoon, two completely different kinds of work.

When people ask whether AI is going to take that job, they are asking about the wrong noun. AI is coming for the first job and cannot touch the second, and most of the future of work is contained in the gap between them. The reading, the reconciling, the checking against policy: machines are now genuinely excellent at all of it. The rating, the ownership, the signature a committee can subpoena: machines structurally cannot produce it, and not because they are not clever enough yet. Almost every job that feels threatened right now is, on inspection, two jobs stapled together under one title, and the staples are about to come out.

Orientation and direction

Call the first kind of work orientation. Orientation is figuring out where you are. It is reading a long document and knowing what it says, recognizing a pattern across a thousand cases, checking a claim against a rule, reconciling two ledgers that should agree, triaging a queue by urgency, summarizing, drafting the memo from the precedent memos. Orientation has a tell: the answer already exists, latent, inside the material and the patterns. The reviewer is not inventing the debt service ratio, she is extracting it. The work is real, often hard, usually the thing that took years to get good at. But it is recovery of something already there, and recovery is exactly what large models do well.

Call the second kind action-directional, or just direction. Direction is choosing where to go when the material does not decide it for you. It is committing to a course under genuine ambiguity, owning the consequence when it lands, standing in front of a regulator or a customer as the accountable party, deciding what matters when the pattern has not formed yet. Direction has its own tell: there is no answer in the material to recover, because the answer is a choice, and the choice belongs to someone who can be held to it. The rating is direction. The batch release is direction. The relationship and the signature are direction. You cannot extract them from the file, because they were never in the file.

The cut runs through every operational job in the economy, and once you see it you cannot unsee it. In a bank, the analyst who reads the transaction monitoring alert is doing orientation; the officer who decides the alert is a real suspicion worth filing on is doing direction. On a plant floor, the reviewer who reads three hundred pages of a batch record against the specification is doing orientation; the quality lead who releases the batch, and owns the recall if it was wrong, is doing direction. In freight, the clerk who assembles the paperwork behind a cargo claim is doing orientation; the person who decides to fight the claim or waive it is doing direction. In each pair the first job is most of the headcount and almost none of the accountability, and the second job is almost none of the headcount and all of it.

The argument has a lineage

None of this is a new observation, only a newly urgent one. In 2014 the MIT economist David Autor wrote a paper with the odd title Polanyi's Paradox and the Shape of Employment Growth, borrowing a line from the philosopher Michael Polanyi, who observed in 1966 that we can know more than we can tell. Polanyi's point was that much of human competence is tacit: we recognize a face, ride a bicycle, sense that a loan smells wrong, without being able to write down the rule we are following. Autor's point was that you cannot automate what you cannot specify, so the tasks that resisted software were exactly the ones whose rules lived in a person's head and could not be dictated to a machine. For thirty years that was the safe ground. Then the models learned to do tacit things without anyone writing the rules down, and the safe ground moved.

So the old dividing line, routine versus non-routine, is not where the border sits anymore. Reading a dense contract and catching the clause that contradicts the term sheet used to be the definition of expert, non-routine work, and a model does it now. What survived is not the non-routine but the non-delegable, which is a different thing. Here the useful rhyme is Hans Moravec, the roboticist who noticed in the 1980s that AI found the hard things easy and the easy things hard: a computer could play master-level chess long before a robot could reliably pick up the chess piece. Knowledge work is living through its own version of that inversion. The part everyone treated as the skilled core, the reading and the analysis, turns out to be the automatable part. The part nobody thought to write on the org chart, being the one who is answerable, turns out to be the irreducible one.

Automation did not come for the hard part of the job. It came for the part that felt like the hard part.

The third thinker worth having in the room is Daron Acemoglu, who shared the 2024 economics Nobel and who spends his book Power and Progress, written with Simon Johnson, insisting on one point: the direction of technology is a choice, not a weather system. Acemoglu draws a hard line between automation, which takes a task a person was doing and hands it to a machine, and augmentation, which creates new work that makes people more valuable. His warning is about what he calls so-so automation, technology just good enough to replace a worker but not good enough to raise what the whole operation can do, which trims the wage bill and improves nothing else. The orientation-versus-direction cut is where that choice actually gets made. Point the machines at orientation and let people keep direction, and you get augmentation: the reviewer does the work of ten and spends her day on the judgment. Point them at both and pretend a system can be accountable, and you get so-so automation with a compliance incident attached.

Why the line is structural, not a stage we pass through

The reflex is to assume this is a capability gap that the next model generation closes. It is not, and the reason is worth being precise about. Direction is accountability, and accountability is a relationship between a decision and a person who can be made to answer for it. A bank examiner does not want a well-calibrated probability that a transaction was laundering. He wants a named human being who decided, on the record, and who can be questioned, sanctioned, or fired. A system cannot occupy that role, not because it reasons poorly but because there is no one home to hold responsible. You cannot fine a model. You cannot strike a model off a register. You cannot put a model in front of a review committee and watch it care about its answer. Responsibility requires a self with something at stake, and that is not an item on a roadmap.

There is a second reason the line holds. Orientation works because the past is a good guide: the pattern in the data and the precedent in the file are enough to recover the answer. Direction is needed precisely when they are not, when the case is genuinely new, when two policies collide, when the right call is the one that has never been made before because the situation has never occurred before. That is the moment the model has nothing to recover, because there is nothing yet to recover, and the whole value of a person is that they will decide anyway and own the decision. A machine can tell you what usually happens. It cannot want a particular outcome and take responsibility for making it happen, and most consequential work, at the top, is exactly that.

The historical rhyme

If this sounds like it should end in mass unemployment, history says look closer at which layer actually gets automated. The spreadsheet is the cleanest case. When VisiCalc shipped in 1979, and Lotus and Excel followed, it did to manual calculation what these models do to manual reading: it made the tedious middle instant. It did not empty the profession. By one widely cited tally drawn from US employment data, the decades after the spreadsheet saw the number of bookkeeping and accounting clerks fall by several hundred thousand, while the number of accountants and auditors grew by more. The tabulation layer, the human calculator, shrank. The judgment layer, the person a client pays to model the what-if and sign the return, grew, because once crunching numbers was cheap, people wanted far more of it done. The spreadsheet did not delete accountants. It deleted the part of accounting that was orientation and grew the part that was direction.

The economist James Bessen documented the same shape in the most quoted example of all, the bank teller. The intuition was that the ATM would end the teller, and for two decades the opposite happened: as machines took over dispensing cash, the number of tellers in the United States rose. A branch now needed fewer tellers, which made a branch cheaper to run, which led banks to open more branches, and the teller's job drifted from counting bills toward the relationship work the machine could not do. The honest coda is that mobile banking later did what the ATM never did, and the teller count finally fell, which is the real lesson rather than a comforting one. Automating some of a job's orientation can grow the job. Automating essentially all of it does not. The tellers who thrived were the ones whose work moved up into direction, and the comfort in the story is conditional on exactly that.

The early evidence from this wave rhymes too. Anthropic's Economic Index, which studies how people actually use its models across millions of anonymized conversations, has found that the dominant pattern is not the machine swallowing a whole task but the machine and the person splitting it, and that in its recent data this augmentation pattern edged ahead of full automation rather than the reverse. It also found the usage clustering into a narrow band of tasks, the reading, drafting, coding, and analysis that are orientation almost by definition, rather than spreading evenly across whole occupations. That is what a task-level recomposition looks like from the outside while it is happening: not jobs vanishing, but the orientation tasks inside them peeling off to the machine, and the human weight shifting toward the end that requires a decision.

What disappears and what concentrates

So here is the concrete forecast, job by job rather than in the aggregate. What disappears is the reading pile. The hours a person spends recovering information that was already sitting in the documents, the triage, the reconciliation, the first draft from precedent, the checking of the thing against the rule, all of it collapses toward instant. What concentrates is judgment, with less and less padding around it. The reviewer's afternoon stops being ten hours of reading wrapped around ninety seconds of deciding and becomes closer to the reverse: a queue of decisions, each arriving already read, already reconciled, already drafted, waiting for the part only she can do. The job does not disappear. It gets denser. Every hour becomes the hour that used to be ninety seconds.

The job does not disappear. It gets denser. Every hour becomes the hour that used to be ninety seconds.

The honest discomforts

That density is not free, and pretending otherwise is how these essays lose the reader who actually runs an operation. The uncomfortable part is the ladder. The way you become a credit reviewer who can rate a loan in ninety seconds is by spending years doing the reading, badly at first, then well, until the pattern lives in you the way Polanyi described. The reading pile was the apprenticeship. It was how orientation slowly hardened into the tacit judgment that direction runs on. If the machine does the reading, the bottom rungs of that ladder are gone, and nobody has a clean answer for how the next generation climbs to the judgment without the years of orientation that used to build it. The optimistic case, which is Autor's more recent argument, is that AI can extend expert judgment to a wider set of workers, letting people with less training reach further instead of hoarding the high-stakes calls among the few who survived the old apprenticeship. That is a real possibility and not a guarantee, and the design of the seam is what decides which way it breaks.

The new work is at the seam

The recomposition also creates work that did not exist before, and it clusters in one place: the seam between the orientation the machines do and the direction the people keep. Somebody has to manage the AI employees the way you manage any team that does the reading, setting their scope and reviewing their output. Somebody has to design the approval rules, deciding which calls can be waved through on a clean record and which always stop at a person, which is genuine policy work with real consequences when it is wrong. Somebody has to audit the machine work, reading the trail of what was pulled and cited and concluded, and catching the misses before they ship. This is not a footnote to the future of work. In an enterprise where the orientation layer is staffed by machines and the direction layer by people, the seam between them, the approval, the evidence, the audit trail, becomes the most important surface in the building. It is where accountability is either preserved or quietly lost, and it will be designed either deliberately or by accident.

This is the division Winsen One is built for, and the whole product is an argument about where the line goes. Its AI employees take the orientation work as a job rather than a suggestion: they own the queue, read every document index-first with a page citation under each finding, reconcile, draft, and flag. The people keep the direction and the signature. And the approval envelope, where every output is submitted to a named person with its evidence attached and nothing leaves without a decision, is the seam, built on purpose as the most important surface rather than left to accident. The future of work is not people or machines. It is orientation staffed by machines, direction concentrated in people, and a well-designed seam holding the two apart where they have to stay apart. We decided to build the seam.