AI in content strategy, and what you have to decide

Generative tools made first drafts cheap. They did not make review cheap, and they did not change what happens when something inaccurate goes out under your name. That gap is where most of the practical difficulty sits, and it is a content strategy problem rather than a technology one.

The useful questions are narrow. Where is a machine draft acceptable, who checks it, what do you tell people about it, and what is the work never allowed near it. Answer those explicitly and most of the rest follows.

What actually changed

The cost of producing a plausible draft fell close to zero. Three things follow from that, and they are worth separating.

Volume expectations rise before capacity does. Once someone senior sees a usable draft produced in a minute, the request rate changes. The constraint moves from writing to reviewing, and review is done by the same people who were already the bottleneck.

Errors became harder to spot. A weak human draft usually reads as weak. A machine draft reads fluently whether or not it is correct, so the surface signals reviewers have relied on for years no longer indicate much. Reviewers have to check claims they would previously have skimmed.

Provenance became a live question. Where a claim came from, who verified it, and whether a source exists are now things you need a record of, because the tool will produce a citation-shaped string whether or not the citation is real.

The decisions worth making explicitly

Most organisations arrive at these by accident, one exception at a time. Deciding them in advance is cheaper.

Where machine drafting is acceptable

Usually fine for internal summaries, first structure, alternative headlines, and reformatting something already approved. Usually not for anything asserting a number, a legal or regulatory position, a customer commitment, or a quotation from a named person.

Who is accountable for the output

Not who ran the tool. A named person who has read the final text and can answer for it. This is the single decision that prevents most of the failure patterns below, and it is the one most often left implicit.

What is checked, and to what depth

A stated minimum: every number traced to a source that was opened, every named person or organisation confirmed to exist and to have said what is attributed to them, every link followed.

What is disclosed, to whom

Covered below, because it is the question stakeholders raise first and the one with the least agreement.

What never goes near a third-party tool

Unpublished financials, personal data, customer records, anything under an NDA, and anything that would be a problem if it appeared in a competitor's output.

These decisions belong with your existing editorial rules rather than in a separate AI policy. Content governance, and who is allowed to change what covers the structure they sit in.

Disclosure, and what people actually want to know

Blanket labelling of anything a tool touched is easy to write and close to meaningless, because on that definition a spellchecker qualifies. It also tends to be read as a warning, which is not usually what was intended.

The questions readers and stakeholders raise are more specific. Did a person check this. Is the author a real person who stands behind it. Were the sources verified. Is anything here generated to look like something it is not, such as a testimonial, a case study or a photograph of a customer.

Disclosure that answers those questions is worth writing. A named author who is accountable for the piece, a note on how sources were verified, and a firm position that you do not fabricate people or evidence. Where a regulator, client contract or platform rule requires a specific label, that requirement decides the wording and is not a matter of editorial preference.

On what it is reasonable to disclose more generally, transparent communication, and what it is reasonable to disclose covers the boundary in more detail.

Review capacity is the real constraint

Planning that assumes drafting was the bottleneck will overproduce. If drafting is 4 times faster and review capacity is unchanged, you have built a queue, and queues get cleared by lowering the standard.

Options that work: reduce the number of pieces and keep the standard; expand who can review, which requires training reviewers to check claims rather than read for style; or narrow where machine drafting is used to the low-risk categories where a lighter check is genuinely enough.

The option that does not work is asking the same reviewers to handle more volume at the same depth. That is the point at which a fabricated statistic gets published, and the reputational cost of one of those exceeds the saving from a year of faster drafts.

The failure patterns

Four recur often enough to plan for.

Fabricated sources. A confidently formatted citation, a plausible author, a DOI-shaped string, and no such paper. This survives review whenever the reviewer checks that a source is cited rather than that it exists.

Laundered claims. A number appears in a draft, gets published, is picked up elsewhere, and is then found by a search that appears to confirm it. The original had no source. Tracing a figure to a primary document is the only reliable defence, and it frequently ends with the figure being dropped.

Voice drift. Output converges on a general register. Across enough pieces the material stops sounding like the organisation, which matters most for the pages doing the persuading.

Silent scope creep. A tool introduced for internal summaries ends up drafting customer communications, because nobody wrote down where the boundary was.

Measuring whether it is helping

Faster drafting is not a result. It is an input, and measuring it tells you almost nothing about whether the content is working.

What carries information:

• Time from brief to published, which captures the review queue rather than only the drafting step
• Corrections and retractions published, tracked before and after adoption
• How often a reviewer sends a draft back, and for what reason
• Whether the pages produced this way perform comparably to those that were not

The last one requires the same discipline as any other content measurement, and data-driven content strategy, and what data cannot decide covers where those numbers stop being able to answer the question.

Be careful with vendor and survey figures on adoption and attitudes in this area. Many are commissioned by companies selling the tools or selling the fear of them, the questions are often not published, and the numbers move quickly enough that a figure two years old describes a different situation. Where you cite one, name the source, the date and the sample, or leave it out.

Where this leaves the strategy

The underlying work did not change. You still need to know who the content is for, what decision it helps them make, and how you will tell whether it worked. Your guide to content strategy frameworks covers that groundwork, and none of it is altered by how the draft gets written.

What changed is that producing more has become easy while being right has not. If you write down where machine drafting is allowed, who answers for the output, and what gets checked, most of the rest is ordinary editorial practice.