Hire a Person or Hand the Work to AI: How to Decide

Sort the work into producing and deciding

List the tasks in the req and mark each one produce or decide. Produce makes an artifact: a draft, a query, a summary, a first pass. Decide commits the company to something: what ships, what goes to a client, what number lands in the board pack. Producing can move to a model tomorrow. Deciding stays with a person who can be asked why.

The mark tracks accountability, not difficulty. Plenty of deciding is easy and plenty of producing is hard, and neither fact changes who has to answer for the outcome. The test is simple: if this went wrong in front of a customer, would the answer be a person's name or a vendor's?

The sort usually comes out lopsided, with far more producing in a role than deciding, and the deciding part is the one that cannot move. That is why the salary-against-subscription comparison feels so convincing and works so badly. It compares the whole role against the tool, when only part of the role was ever in scope.

Run the sort on the actual task list. Job descriptions are written to attract people, so they tend to describe the deciding part and skip the rest. Ask the person currently absorbing this work to list what they did last week. That list is the real req.

What the replacement evidence actually supports

Skills get replaced, not jobs, and only a thin slice of skills at that. Indeed's GenAI Skill Transformation Index rated almost 2,900 work skills against more than 53.5 million US postings and found 19 of them, 0.7%, very likely to be fully replaced by generative AI; 40% of skills fell into hybrid transformation, which the analysis defines as work where human oversight remains necessary ¹.

Those ratings come from large language models scoring skills, with nobody watching work get done, so they describe what current models are judged capable of. That sets an upper bound on transformation and records nothing about what any employer has changed. A skill is also not a job: no role in the analysis consists only of the 19 fully replaceable skills.

The demand side points the same way. The Burning Glass Institute found skills exposed to automation were 16% more likely than baseline skills to see demand decline in postings, while skills exposed to augmentation were 7% more likely to see demand increase, and that the occupations seeing the most automation are simultaneously seeing the most augmentation ². Those are relative likelihoods against a baseline group. Neither number is the size of a change.

And adoption itself is still partial. The Census Bureau's Business Trends and Outlook Survey put AI use among US businesses at 19.8% as of 3 May 2026, rising to 37% among firms with at least 250 employees ³. The unit there is the firm and the question asks about the previous two weeks, so it counts activity at a floor and says nothing about depth. Read together, the three say the same thing: roles are changing shape, and very few are disappearing.

Measure the review hours the automation creates

Take one week of the work in question, automate the produce half on paper, and count the hours somebody would spend checking the result. Two numbers come out: hours of review created, and the name of the person who has to find them. If that person cannot find the hours without dropping something else, the automation has not removed the work. It has moved it and hidden the cost.

A worked version, on a marketing req that is mostly production. The week contains twenty first-draft briefs, six campaign summaries and four client updates. Hand the drafting to a model and thirty pieces still arrive, each needing somebody to read it against the brief and check the factual claims in it. At fifteen minutes each that is seven and a half hours a week, roughly a fifth of a person, landing on a reviewer who was already at capacity.

The number resolves the decision three ways:

• Under the slack that genuinely exists. Hold the req, write down who owns the review, and set a date to re-check the volume.
• Above the slack, below a full role. Reshape the req instead of cancelling it. What is needed is a reviewer with the judgment to send work back, which is a different hire from the producer originally scoped.
• Well above a full role. The automation increased headcount need instead of reducing it, which happens more often than the planning deck suggests. Open the req.

Doing this across every open role at once is the same exercise at portfolio scale: how many people you actually need next year if AI takes on part of the workload.

When is the answer genuinely no hire?

When the produce half is large, the decide half is small, and a named person has the review hours without dropping anything. That combination is real and it does occur. It usually looks like a role that was mostly volume in the first place: a queue that grew faster than the thinking behind it, or a task somebody picked up because there was nobody else free.

Write three things into that decision before it goes into the plan. Name the reviewer, in the document, by name and not by team. State the volume ceiling above which the decision is revisited, because the reason the hours fit today is that the volume is what it is today. And set the re-check date, ninety days out, with the volume number attached so the conversation starts from evidence instead of from whoever is most tired.

The failure mode to avoid is a cut made on the assumption that AI absorbs the work, followed by a quiet rebuild of the same capacity under different titles a year later. That path costs the severance, the rehiring, and the institutional memory in between, and the second hiring round is harder because the requirement changed while nobody was writing it down (what to hire for after a cut the AI did not absorb).

The entry-level version of this question deserves its own arithmetic, because the cost of not hiring shows up three years later rather than this quarter: does hiring entry-level still pay off.

See the benchmarks >