What a Remote AI-Heavy Posting Has to Say About Review

What does a remote AI-heavy posting have to say?

Two things, in two short paragraphs. The first names the review arrangement: who the work goes to, and on what terms. The second names the delegation boundary: what belongs to the assistant, and what never leaves the person. An office answers both without anybody writing them down, because both are visible from a desk. A remote posting has to write them down, and an AI-heavy one has the least to fall back on if it does not.

Most remote hiring advice is still about tools, time zones and async rituals. Those matter and they are solved. What changed is quieter: colocated work taught the job through two channels nobody ever documented. People watched somebody more experienced do the thing, and people got corrected on visible, obviously rough output. Remote removes the first. An assistant that returns something polished on the first pass removes the second, because a draft that looks finished does not invite the correction a bad draft used to attract.

The posting is also competing against other postings that use the same words. Indeed Hiring Lab put AI-related job postings at 6.3% of US postings in August 2026, past their prior peak of 3.3% in 2022 ¹. That counts postings mentioning AI, which measures employer language rather than adoption, and it is enough to make the point: a candidate reading three AI-heavy remote postings in one sitting has nothing to tell them apart unless one of them says something specific.

The phrase itself carries almost no information. In nationally representative US surveys in late 2024, 23% of employed respondents had used generative AI for work at least once in the previous week, and 9% used it every work day ². Self-reported, and a low bar. Between the person who opened it once last week and the person who uses it every working day sits a range of practice wide enough that "AI-heavy" tells an applicant nothing about their week.

Name the reviewer, the cadence, and the standard

Three specifics, roughly forty words. Who reads this person's work, how often it happens, and what it is checked against. "A senior analyst reviews every client-facing model before it ships, twice a week, against the source system rather than against how it reads" is a complete answer, and it is the sentence most postings replace with a line about a collaborative culture.

The third specific is the one teams get wrong, because the obvious standard is the useless one. In a Boston Consulting Group field experiment, on one task deliberately chosen to sit outside AI capability, consultants using GPT-4 were 19 percentage points less likely to reach the correct answer, 84.5% of the control group against 60% and 70% in the two AI conditions, and the group given a prompt-engineering overview did worse than the group given none ³. One task, one sample, a 2023 model. What generalizes is that nobody could tell which side of the line the task was on, because the wrong output read exactly as well as the right output.

So the standard has to sit outside the document. Reviewing generated work against how it reads is reviewing it against the thing the model optimizes. Reviewing it against the source system, a recomputation, a named person who knows, or a test that fails is reviewing it against something that can disagree.

Write the cadence in units the candidate will actually feel: per deliverable, weekly, or at a named checkpoint. "Regular feedback" is read as none. And name the role rather than the person, so the sentence survives the reviewer changing jobs. If the honest answer is that nobody reviews the work today, that is worth knowing before the posting goes up, because it is a description of a job with no correction loop, and it will show up in month two as a new hire who demoed brilliantly and then fell apart.

Write down what must not be delegated

One paragraph naming what the assistant is expected to do and which calls stay with the person. It is the difference between a posting that attracts people who work well unsupervised and one that attracts people who assume nobody is checking, and it costs nothing when it is true. It is also the only part of the AI clause a candidate can act on.

Write it as two short lists rather than a philosophy:

• Handed over, expected. First drafts, summaries, the initial data pull, routine correspondence, boilerplate code, the meeting notes. Say this out loud, because a candidate who thinks generation is discouraged will hide it and a candidate who thinks everything is fair game will hand over the wrong thing.
• Stays with the person. The claim that goes to a client, the number in the board pack, the decision between two options, anything about a colleague or a customer, and every case where being wrong is not reversible in an afternoon.

The boundary is more useful in a posting than a tool list, and it dates far more slowly. Naming products tells a candidate which subscription you hold; naming behaviors rather than tools tells them what the job is. If part of the offer is access, say that too, since advertising the tool access is a real draw and costs a sentence.

One test before publishing. Read the paragraph as a candidate who is good at this and ask whether it describes anything they would have to change. If the answer is no, it is too vague to be worth the space. A boundary that does not exclude anything is not a boundary, and the version that excludes the right people usually names one specific thing the person is not allowed to hand over.

Why does silence about review cost both ends of the pool?

Because the two strongest kinds of applicant read the silence in opposite ways, and both walk. A strong early-career candidate reads a remote AI-heavy posting with no review arrangement as a job with nowhere to learn, since both of the channels that used to teach the work are gone by default. A strong senior reads the same silence as a team with no support and no bar, which is a job where their judgment will be uncheckable and therefore unrewarded.

Who stays is the problem. Candidates who are indifferent to review are often indifferent because they have never needed it, which is a poor match for a role that is now mostly judgment exercised out of sight. That is how a posting written to sound flexible ends up filtering for exactly the working style it can least afford at a distance.

The honest version also protects the first ninety days, which is where this bill actually arrives. A hire who has never been told what the standard is will set their own, and an assistant will help them meet it convincingly. Structuring the first ninety days of an AI-heavy hire is far easier when the posting already named a reviewer and a cadence, because onboarding then confirms an expectation rather than introducing one.

There is a second-order version worth writing into the posting if it is true: whether asking a person is expected before asking the assistant. Teams have started noticing that new hires ask AI before they ask anyone, and remotely there is nobody sitting nearby to make the human option the easy one. A single sentence saying who to ask, and that asking is expected rather than tolerated, does more for a remote AI-heavy role than any paragraph about culture.

None of this lengthens the posting much. Two paragraphs, maybe eighty words, in exchange for a document that describes the job as it will actually be experienced from a kitchen table three time zones away.

See the benchmarks >