The useful question about AI and writing is not "can it do this job." It can do most jobs passably, which is exactly what makes the decision difficult. The useful question is what happens if it gets the job wrong, and how much it costs you to find out.
That reframing turns a vague anxiety into a working rule. Sort your writing tasks by how expensive a wrong answer is and how quickly you would notice one. Everything cheap and visible is a candidate for delegation. Everything expensive or slow to detect is not, regardless of how well the model performs on a sample.
The question is reversibility
Reversible work can be thrown away without consequence. A headline that does not work costs you nothing but the minute you spent reading it. A rejected paragraph costs a minute of editing. Irreversible work commits you: a number you published, a promise you made to a customer, a claim that is now indexed and quoted.
The second axis is detectability — how long it takes to notice the mistake. A draft with the wrong tone is obvious on first read. A subtly wrong technical explanation might survive review, get published, and only surface when a knowledgeable reader writes in three months later, by which point it has been copied elsewhere.
Plot tasks on those two axes and the delegation boundary draws itself. High reversibility, fast detection: delegate freely. Low reversibility, slow detection: keep it. The awkward quadrant is low reversibility but fast detection — commitments you can catch by reading carefully. Those are delegable only if you actually read carefully, which is a bigger condition than it sounds.
Start with reversible work
The safest first category is volume work with no external commitment: alternative headlines, section headings, transition sentences, list titles, subject lines for internal review, and first versions of paragraphs you were going to rewrite anyway.
There is a second category worth adding early: work that is easier to judge than to produce. Generating twenty possible openings and picking one is faster than writing one, even though picking still requires taste. Recognising a good option is a much cheaper mental operation than producing it, which is why this category has the highest return of anything on the list.
Keep the outputs short at first. Long outputs hide their own problems, and the effort of checking a two-thousand-word draft is close to the effort of writing it. A paragraph or a heading is checkable at a glance.
Four things never to delegate
Four categories should stay with a person, not because the model cannot produce something plausible but because plausible is the failure mode.
- Facts you have not verified. Any number, date, name, version, limit or requirement. A generated figure is not an estimate; it is a sentence shaped like a fact.
- Commitments. Prices, timelines, guarantees, refund terms and anything else a reader could hold you to. These are commercial promises, and they need a human who can be held responsible.
- Positions you will have to defend. What you think about a contested question, and why. A model can summarise a debate; it cannot hold a view, and readers can tell the difference between an argued position and a survey of opinions.
- Anything attributed to a person. Quotes, paraphrased advice, testimonials. Generated quotes are the single most damaging thing to publish, and they are also the easiest mistake to make, because the format is familiar.
Note the pattern: these are the places where the sentence's value depends on it being true and belonging to you. Style is delegable. Accountability is not.
Let AI do the structure
Structure is the best-kept delegation opportunity, and almost nobody uses it. Ask a model to propose four different ways of organising the same material — a chronological version, a comparison version, a version organised by objection — and you will usually get one arrangement you had not considered. Then you choose, which is the part that needs judgement, and you have spent two minutes instead of an hour.
This works because structure is a search problem with a small number of good answers, and searching is what models do well. It also avoids the most common failure of delegated writing, which is that the model supplies the structure by default and every article ends up shaped the same way.
Do the thinking yourself
The work you should never hand over is deciding what the piece claims, which example proves it, and who specifically it is for. Those three answers are the article. Everything downstream — arrangement, phrasing, transitions, titles — is execution, and execution is where delegation pays.
The practical test before you start: can you state the claim, the evidence and the reader in three sentences without a tool? If yes, delegation will speed you up. If no, delegation will produce a fluent piece about the topic that argues nothing, and you will spend longer fixing it than writing it would have taken.
A three-column triage
List your recurring writing tasks and sort them into three columns:
- Delegate. Reversible, fast to check. Generate options and choose.
- Draft then verify. The output can be generated, but every factual statement must be traced to a source you can open. Useful for explanatory and instructional writing where you already know the material.
- Write yourself. Commitments, positions, attributions and anything you would be embarrassed to correct later.
Most teams find the middle column is where the real productivity lives, and that they had been treating it as either column one or column three. "Draft then verify" needs a defined verification step, otherwise it silently degrades into delegation with no check, which is the worst of both.
The first-draft rule
A workable default for the middle column: the first draft may be generated, the first judgement may not be. Someone reads the whole thing and decides what is true, what is missing and what the piece actually says. Reordering or polishing before that pass is wasted effort, because the judgement usually changes what needs to be there.
This is also why "AI wrote it" is rarely the real problem when published content goes wrong. The tool is not accountable for a published claim, and readers do not blame a tool. The failure is always the skipped judgement, and that step takes minutes when it is planned and hours when it is retrofitted.
When the boundary moves
Models improve, and some things in the "write yourself" column will become safe to delegate later. Revisit the boundary twice a year, but change it deliberately rather than by drift. The way to move it responsibly is to take one task from the middle column, add the verification step that makes it safe, run it for a month, and check whether the checks actually caught anything.
What should not move is the underlying rule. Any sentence whose value comes from being true, or from being yours, stays with a person. That is not a limitation of current tools; it is a description of what publishing something under your name means.