The four content jobs to give AI and the four to never

The most useful boundary for AI in content production is not a policy document. It is a copy of what working journalists actually use it for, because they operate under a correction regime that punishes shortcuts publicly and immediately.

The four jobs they hand over: summarising, transcription, background research, and proofreading. The four they do not: the argument, the anecdote, the quote, and the number.

That split holds up under the evidence and it is easy enough to enforce on a Monday morning.

The four to hand over

Summarising. Feeding in twelve source documents and getting back a structured digest is genuinely faster and genuinely accurate, provided the sources are yours and are attached. The output is a map of material you already possess, not new claims about the world. Verification is cheap because the source sits next to the summary.

Transcription. Interviews, client calls, discovery sessions, sales calls. This is the highest-value use in the whole list and the least discussed. A team that transcribes every client conversation has a permanent library of real language from real buyers, which is the raw material for every piece of copy that will ever sound like it came from someone who has met a customer.

Background research. Getting oriented on an unfamiliar area, finding the names of the standards, identifying who the authorities are, working out what questions to ask. The output is a starting point for verification, never a finishing point. Every claim that survives into the final piece has to be checked against the primary source. Every one.

Proofreading. Consistency, repetition, inconsistent terminology across a long document, structural sag in the middle. Models are good at spotting the paragraph that says the same thing as the one three paragraphs earlier, which is exactly the thing a writer becomes blind to.

The four to keep

The argument. What you think and why. This is the only part of a piece a reader cannot get anywhere else, and a model generating an argument is averaging every argument in its training data, which by definition produces the consensus position. A consensus position is not an argument. It is a summary of other people having had one.

The anecdote. The project where it went wrong. The client who asked the question nobody expected. The month the numbers said one thing and the sales team said another. These are unfakeable, unindexable by anyone else, and the single fastest way to establish that you have actually done the work.

The quote. Never generate a quote. Not for a case study, not for a testimonial, not as a placeholder that somebody will approve later. Placeholder quotes get approved. Then they get published, and you have attributed words to a named human being that they never said.

The number. Every figure comes from a source you can name, with a sample size and a date. A model will produce a plausible statistic on request and it will look exactly like a real one. There is no visual difference between a verified figure and a generated one, which is why the rule has to be absolute rather than case by case.

Why this boundary and not another

Look at what separates the two lists. Everything in the first list operates on material you supplied. Everything in the second list requires knowledge of the world that has to be true.

That is the whole principle. The model is safe when it is rearranging your inputs and unsafe when it is supplying facts.

Once a team understands it that way, they can classify a new task without asking. Am I giving it the material, or asking it for material, is a question anyone can answer in three seconds.

Making it operational

Write it into the brief template, not into a policy nobody reads.

Every content brief gets four fields that must be filled by a human before drafting starts: the argument in one sentence, the anecdote or example, the source for any quote, and the source for any number. If those four fields are empty, the piece is not ready to be written, by a person or otherwise.

That single change does most of the work, because it front-loads the human contribution to the point where it is cheapest to produce and impossible to skip.

Then two checks at the other end. Any number without an inline source comes out. Any quote without a named person who has approved it comes out. No exceptions for deadline, because deadline is the only condition under which the rule ever gets broken.

The honest limits

Transcription accuracy drops on strong accents, crosstalk and technical vocabulary. Budget review time on anything you will quote from, and never quote directly off a raw transcript without listening back to that section.

Background research produces confident, plausible, wrong answers at a rate that has not gone to zero and probably will not. Treat every output as a list of things to check, and expect one in ten to be wrong in a way that would have been embarrassing.

Summarising loses the outlier. The unusual detail in source document nine, the one that would have made the piece interesting, is exactly what a summary drops as noise. Read the sources yourself if the piece matters.

Give it the material and keep the meaning. That line holds in every situation you will meet this year.