Where AI Actually Fits in a Content Programme You Already Run

Most advice about an ai content strategy starts from a blank page: here is how AI transforms content, here are forty-seven use cases, here is a diagram with robots in it. That advice is useless if you already have a programme that works. You have a calendar, a brief template, a freelancer or two, a monthly report someone actually reads. You don’t need transformation. You need to know which three or four steps in your existing workflow are genuinely constrained, and whether a model helps at those steps specifically.

My answer, after running this in a four-person team and watching a dozen others try it: four insertion points. Research synthesis, structural drafting, repurposing, reporting. Everything else in the pipeline either doesn’t benefit or actively gets worse.

Map what you actually do first

Write out your real workflow, not the one in the deck. A typical in-house programme publishing six pieces a month looks roughly like this:

1. Quarterly planning and topic selection
2. Research: keyword data, customer calls, competitor coverage, SME interviews
3. Brief writing
4. Drafting
5. Editing and fact-checking
6. Design, upload, internal links, publish
7. Distribution and repurposing
8. Measurement and reporting

Now mark where time goes versus where value comes from. In every programme I’ve audited, steps 2, 3, 7 and 8 absorb a disproportionate share of hours while contributing almost nothing that a reader would recognise as your voice. Step 4 is where the value sits. Step 5 is where trust gets earned or destroyed.

That asymmetry is the whole argument. Put AI where the hours are and the voice isn’t.

Insertion point one: research synthesis

This is the strongest case and the one people skip because it isn’t glamorous.

Here’s the worked example. Last quarter we had nine recorded sales calls (transcribed with Otter, roughly 6,500 words each), eleven support tickets tagged “onboarding”, and a Semrush export of 340 keywords in our cluster. That’s about 60,000 words of raw input. Reading it properly is a day and a half of someone’s week, which is why nobody ever did it.

We loaded the transcripts and tickets into a NotebookLM notebook and asked a deliberately narrow question: which specific objections appear in three or more calls, and what exact words does the buyer use? Not “summarise these calls.” The narrow question is the trick. We got back seven recurring objections with quoted phrasing, and two of them (“we already have someone doing this internally” and “how long before I can show my boss something”) had no corresponding content anywhere on the site. Those became two briefs that week.

Time cost: about 50 minutes including sanity-checking the quotes against the source transcripts, which NotebookLM makes straightforward because it cites back to the passage. Against a day and a half that never happened, that’s not a marginal gain.

Claude handles the same job well if your inputs fit in context; a 200k-token window swallows roughly 150,000 words, which is more raw research than most quarterly cycles produce. For competitor coverage specifically, run a Screaming Frog crawl of a rival’s blog, export the URL and title list, and ask for the gaps against your own export. Two minutes of setup, and you get a genuine content gap list rather than a keyword tool’s guess at one.

What this step is not: asking a model what your customers care about. It has no idea. It’s synthesis over evidence you supply, and that distinction is the difference between a useful insight and a plausible-sounding hallucination about your market. If you want the fuller planning picture, including how this feeds topic clusters and quarterly sequencing, our guide to AI-assisted content strategy and planning covers the planning layer in more depth.

Insertion point two: structural drafting, not prose drafting

The distinction matters more than anything else in this post.

Give a model a brief and ask for a 1,500-word article, and you get the thing everyone complains about: competent, airless, indistinguishable. Give it the same brief and ask for an argument structure, and you get something genuinely useful, because structure is a logic problem rather than a voice problem.

What we ask for now: the claim the piece makes, the three or four supporting sections, the objection each section pre-empts, and the specific evidence each section needs that we do not yet have. That last item is the valuable one. It produces a shopping list. “Section 3 asserts that migration takes under a fortnight but you have no data supporting this” is exactly the note a good editor gives, and it arrives before anyone has written a word.

Our brief template used to take about two hours to fill properly. It’s now 35 to 40 minutes: 10 minutes prompting against the research synthesis, 25 minutes of a human cutting, reordering and adding the things the model couldn’t know. Across six pieces a month, that’s roughly nine hours back.

Then the human writes the draft. Actually writes it. This is the step where a one-to-five-person team still earns its keep, and handing it over is how programmes end up with 40 posts a year that nobody links to.

Insertion point three: repurposing

Repurposing is the step everyone knows they should do and nobody does, because it’s the least interesting work in the programme and it always loses to the next deadline.

It’s also nearly perfect for a model. The source material exists, it’s already been fact-checked, the voice is already established in the text you’re feeding in, and the output is short enough that editing it is fast. Take a published 1,800-word pillar, paste the whole thing in, and ask for six LinkedIn posts that each make one argument from the piece and do not summarise the whole thing. Specify the constraint: no “here are 5 things”, no emoji bullets, no opening with a one-line hook and a line break. You’ll still rewrite half of them. Half is fine when the alternative is zero.

A realistic figure from our own tracking: one pillar produces six LinkedIn posts, one newsletter section and a four-slide carousel in about 45 minutes of combined generation and editing. The same output took most of a day when done by hand, which meant it happened for maybe one piece in five.

Descript is worth a mention if you have any video or webinar recordings sitting unused. Its transcript-based editing plus an AI pass for clip selection turns a 45-minute recorded session into four or five short clips in under an hour.

Insertion point four: reporting

Monthly reporting eats a half-day and produces a document whose main function is to prove the programme exists.

Connect GA4 and Search Console to Looker Studio once, then export the monthly figures as CSV and hand them to a model with your actual questions attached: which pages gained or lost more than 20% of clicks month on month, which queries moved from position 11-20 into the top 10, which published pieces have had fewer than 50 sessions after eight weeks. Ask for the anomalies, not the totals. The totals are already on the dashboard and nobody reads them.

We went from roughly three hours to 45 minutes, and the reports got better, because the time now goes into the “so what” paragraph rather than into copying numbers between tabs. One caveat worth stating plainly: models are unreliable at arithmetic on pasted tables. Verify any figure you’re going to say out loud in a meeting. Use the model for pattern-spotting and narrative, and let Looker Studio own the numbers.

Leave these alone

Editing and fact-checking. A model will confidently smooth a factual error into more fluent prose, and it cannot tell you that the pricing changed in March.

Final voice pass, obviously. Anything involving a customer’s name or an unpublished figure, unless you’re on an enterprise agreement with the data controls to match. Topic selection as a first move, because a model’s suggested topics are a fossil of what already ranks, and your programme’s edge comes from what your customers said last Tuesday.

Start with one

Pick the single step where your team currently loses the most hours and cares the least about the output. For almost everyone reading this, that’s the monthly report or the repurposing backlog. Run it for four weeks, track the actual minutes, and keep a note of how much you rewrote.

If you’re rewriting more than half, the prompt is too vague or the step was the wrong one. If you’re rewriting under a fifth, move to the next insertion point and check back in a quarter, because the model will drift and your standards will slip in the same direction without you noticing.