Content marketing in 2026: when are we supposed to do the writing?

I’ve spent my career writing: journalism, copywriting and content marketing. I know what it takes to turn an interesting observation into something another person might find useful enough to read. I also know the less satisfying version, where you meet the brief and the deadline but suspect that a better piece was possible.

I use AI extensively. It helps me work through research, develop arguments, question briefs and connect things I would otherwise struggle to hold together. I’m developing Editorial Intelligence, my own practice around evidence, editorial judgement and applied AI, partly because these tools have made a much wider range of work possible.

That puts me in an uncomfortable position when I ask this: with everything content marketing now involves, when are we supposed to do the writing?

I mean the writing where you sit with an idea long enough to discover that your first explanation was inadequate. Where you go back to an interview because something doesn’t add up, or remove a paragraph you like because it doesn’t help the reader. Where finding the right words is part of discovering what you actually think.

That work still takes attention. AI has changed how I approach it, but the rest of the working day has changed too.

The article has acquired a much bigger surrounding job

My title at Sage is Content Marketing Manager, and the role has always included strategy, delivery and coordination as well as words.

But when I look across the work I’m currently involved in, the range is striking. Alongside writing and editing, there is:

• Interpreting research and developing editorial narratives.
• Finding useful practitioner evidence in events, interviews and conversations.
• Connecting customer stories to questions audiences are asking.
• Exploring generative engine optimisation, or GEO, and what AI discovery means for our content.
• Reviewing video conversations and identifying useful clips.
• Working through YouTube’s role alongside website and social publishing.
• Developing promotional plans and coordinating handoffs.
• Testing ways to preserve evidence so the next project can reuse it.
• Experimenting with AI workflows for customer stories and editorial briefs.

I do most of this within wider teams, and several items are experiments rather than established practice. Outside Sage, Editorial Intelligence adds website development, publishing, distribution, service positioning and maintaining the knowledge behind the work. That’s a separate project I’ve chosen to pursue, not a workload my employer has imposed.

Across both, though, I can see how the range expands. AI makes an adjacent problem easier to investigate. You notice a connection, propose something useful and start exploring it. Sometimes someone asks you to take it further; sometimes you give yourself the responsibility.

Meanwhile, the articles still need writing. The capability has grown, but the week has the same number of hours.

The whole package makes sense to the buyer

The roles and freelance assignments I’ve been looking at raise a related question: how much work sits inside the package being requested?

An article can arrive with audience research, interviews, search optimisation, images, social copy, publishing and performance reporting attached. Each element has a reasonable justification. A buyer wants something that reaches people and achieves a purpose; an excellent document sitting in a folder won’t do that.

Experienced freelancers have often supplied much more than words, and a senior in-house role should involve judgement about where the work fits. The difficulty comes when the package expands while assumptions about time, budget and capacity stay roughly where they were. The writer is still expected to deliver excellent copy, but excellent copy has become one component of a substantially larger assignment.

I explored the economics of that change in The writing job didn’t disappear. It unbundled.. Drafting can be priced down while research, judgement, distribution and coordination are bundled back into the job. Unless those additions are recognised, a broader package can contain a surprising amount of unpaid or unallocated work.

AI makes the package easier to imagine. It doesn’t tell us whether its parts can all be delivered well within the available time.

A faster draft isn’t a faster assignment

AI does save me time. Comparing documents can become much quicker. A transcript can become easier to navigate. A rough structure gives me something concrete to challenge, and adapting approved material for another format may take far less effort than starting again. Sometimes those savings are substantial.

But the time saved on drafting isn’t a reliable measure of the capacity created across the whole assignment. A piece still needs adequate evidence, a defensible argument, accurate attribution and a review proportionate to its claims. Stakeholder decisions and publishing dependencies don’t necessarily accelerate because the first draft did.

There is also a choice hidden inside the saving. If AI frees up an hour, that hour could support a better interview, closer checking, another deliverable or a shorter working day. Those are different uses of the same gain. We can’t assume the hour pays for all of them.

The wider the remit becomes, the more explicit that choice needs to be. Otherwise an improvement in one stage quietly becomes the justification for adding work across several others.

Writing is one of the ways we discover the argument

There is a tendency to treat writing as the final expression of thinking that has already happened. In practice, writing often exposes how incomplete the thinking was.

You try to explain a claim and discover that you don’t quite believe it. You put two examples together and realise they contradict each other. You attempt a straightforward introduction and find you still haven’t decided who the piece is for.

That difficulty can be productive. It sends you back to the reporting, forces a narrower claim or reveals the question you should have asked in the interview. A stronger article sometimes comes from recognising that the material won’t support the article you originally intended.

AI can participate in that process. I use it to challenge arguments, compare structures and surface gaps, and I used it to help develop this article. Sometimes the model gets me to a useful distinction much faster than I would have managed alone. At other times I need to work through the language myself, because that’s how I find the thought, and I want to make that call according to the work.

It can also smooth over the difficulty I need to examine. A paragraph can sound finished before the underlying question has been resolved. A missing explanation becomes a plausible transition, and two contradictory examples become a tidy summary that conceals their disagreement.

Recognising that is slow work, especially when the draft is fluent enough to survive a quick read. This is part of what I mean when I say the AI draft is not the article: publishable work depends on the decisions made around and within that draft.

If writing becomes a small finishing task squeezed between channel planning and the next meeting, we risk losing some of the scrutiny that makes the finished work worth reading. There’s no contradiction in using AI extensively and wanting time to write carefully. The contradiction comes from expecting AI’s speed and careful writing’s quality without allocating time for the second.

The responsibility stays with the person publishing it

Even where AI reduces effort, responsibility remains. I still need to know whether a source supports the sentence, whether the customer really said what the draft implies and whether a claimed benefit belongs to the product, the process or several things working together.

That doesn’t require manually checking every harmless transformation with equal intensity. It requires knowing where an error would matter and looking hardest there. A title variation and a customer outcome claim deserve different levels of scrutiny.

As the number of outputs and channels grows, there are more places to exercise that judgement. A source can be represented accurately in an article and exaggerated in its social introduction. A useful video extract can lose the qualification that made the original answer defensible.

Reuse is still valuable. But some of the time saved in producing variations has to remain available for checking whether the variations preserve the meaning. Otherwise the apparent efficiency depends on someone accepting responsibility for work they haven’t had time to assess.

I’m helping expand my own job

It would be convenient to place all the responsibility on employers and clients. That wouldn’t be an honest account of my experience.

I enjoy the wider work. I want to understand how evidence travels across channels, how AI changes discovery and whether a better workflow can stop somebody having to reconstruct the same context next month. AI lets me explore ideas that previously would have stayed in a notebook.

Some of that work deserves more of my time than another article would. A better brief, a clearer handoff or a decision to reuse existing material can be more valuable than producing another page. In The real AI productivity gain is scope, not speed, I described how AI makes it feasible to understand more of the problem around an assignment.

But understanding more of the problem creates a temptation to take responsibility for more of it. I can see how an interesting investigation becomes a proposed project before I’ve decided what it should displace.

That is a boundary I need to manage myself. Enthusiasm doesn’t create capacity, and a clever prototype doesn’t automatically deserve maintenance. Some ideas should remain possibilities until there is a clear reason to invest in them.

Care still needs a purpose and a deadline

There is a strong counterargument here. Writers can overwork pieces, confuse personal satisfaction with reader value and spend hours improving sentences that were already good enough. A commercial team can’t reserve unlimited time for craft simply because the writer cares deeply about it.

A brief service update doesn’t need the development process of an investigative feature. Some material should be produced quickly, checked appropriately and published. An experienced writer should be able to recognise when further work will have little effect.

The same test applies to my systems work. Spending an afternoon improving a workflow that nobody reuses is no better than endlessly polishing an introduction. I’ve argued that the work that improves the work needs protecting, but it also needs to show that it improves something.

So the answer can’t be to give every article more time. It is to make a deliberate judgement about the attention each assignment warrants. That judgement should consider uncertainty, evidence, audience and consequence, as well as length and format.

A short piece making a sensitive claim may need more scrutiny than a long guide built from settled material. Counting words won’t reveal that difference.

Make the trade-offs part of the brief

For an in-house team, a broader remit should trigger a conversation about existing output expectations. If someone takes on video planning, GEO research or workflow development, which deliverables change, move or receive support? The additional work needs a place in the plan.

For a freelance assignment, the package needs to be visible before the fee and timetable are settled. Research, interviews, drafting, revisions, derivatives, publishing and reporting are all scoping questions. AI may change the effort involved, but a client and freelancer still need to agree what is being supplied.

For an individual practitioner, including me, there is a more personal question: am I solving a recurring problem, or opening another interesting project? An idea can be worth keeping without becoming this week’s commitment.

Across all three, I’d make a few decisions explicit:

• What must this piece achieve? Give it a clear reader need and a reason to exist.
• Where does it need sustained attention? Identify the uncertain claim, missing evidence or difficult explanation before treating the draft as nearly finished.
• What does the package include? Make the work around the article visible.
• How will we use any time saved? Decide whether the gain supports quality, additional scope, development or available capacity.
• What will stop or move? Give new responsibilities an actual trade-off.

Those decisions won’t remove deadlines or make every week comfortable. They do give the person doing the work something more useful than an instruction to be efficient.

Anthropic’s evidence: the work moves, and attention becomes the bottleneck

A few days after I first published this piece, Anthropic released How Employment is Changing at Anthropic, dated 16 October 2026. It is unusually useful evidence because it examines what employees actually do inside a company that has adopted AI deeply, rather than asking them to predict what might happen.

The headline is not that humans have stopped working. By August, the runtime of employee-directed AI agents was six times Anthropic’s estimated human working hours. That is agent runtime, not six hours of equivalent human productivity for every hour worked. The researchers estimate human hours from headcount and an assumed 40-hour week, so the comparison deserves care.

More revealing to me is the new constraint. Anthropic estimates that even non-technical employees would spend about half their working week reading everything their agents produced. Employees respond by creating dashboards, monitoring pipelines and other systems to manage the output. The share of interactive Claude sessions devoted to this kind of ‘meta-work’ rose from 10% to 22%.

That is a recognisable pattern in my own work, even though I operate at a much smaller scale. Searching a long event transcript with AI is quick. Deciding which answer deserves to become a video clip, checking the speaker’s exact meaning, establishing its relevance to GEO and getting the right team to use it are separate tasks. A system can surface possibilities faster than I can responsibly act on all of them.

I’ve also been building a seven-stage customer-story workflow with source hierarchies, verification rules, metric statuses and human approval gates. The aim isn’t simply to generate a case study. It’s to make decisions and evidence traceable from brief to approved story and downstream content. That involves designing the process, testing it against actual projects and deciding when not to change it. Those are real tasks, even if a draft appears more quickly.

The same applies to preserving transcripts and editorial research. Material that would once have disappeared into a folder can now be searched and reused for an ebook, article or future campaign. That is valuable. But someone still has to distinguish a reusable insight from a tempting quote that doesn’t support the new claim.

Anthropic also documents a product-content team that automated much of its writing through a context graph and moved writers into creating prompts and evaluations. That’s a particularly sharp example of a job changing shape. It doesn’t prove that all writers will make that transition, or that they should. It does suggest that language expertise may increasingly be applied to how systems behave, as well as what audiences read.

What OpenAI’s Spotlight comments add

There was an interesting contrast at Semrush Spotlight in London on 13 October. As Adweek reported, Leila Woodington, OpenAI’s head of B2B marketing for EMEA, described field marketing as a particularly strong channel for enterprise sales. Her point was about the distance between recognising ChatGPT and persuading a business to deploy it across thousands of employees and multiple decision-makers.

I don’t want to force that into a competition between OpenAI and Anthropic. They were answering different questions. OpenAI’s reported comments concerned the challenge of adoption. Anthropic’s research examines what happens to work organisation after unusually intensive adoption.

Put together, they suggest three different stages that marketing and business leaders sometimes collapse into one: buying an AI tool, getting people to use it meaningfully, and redesigning responsibilities once it works. A successful demonstration doesn’t resolve who checks the outputs, owns the source material, approves claims, maintains the workflow or stops lower-value work.

This is where the conversation becomes more interesting than ‘AI saves time’. A team can be convinced to adopt AI and still have no clear answer about what the resulting capacity is for. If the answer is simply ‘more of everything’, the new bottleneck becomes the human who must choose, coordinate and sign off.

The uncomfortable skills question

Anthropic reports that junior technical positions fell from 27% to 9% of technical hiring, while hiring managers increasingly emphasised autonomy and judgement. That is a change in the composition of hiring at one fast-growing AI company, not evidence that junior jobs everywhere have declined by the same amount. The report also describes internal redeployment and substantial company growth.

Still, it raises a question that reaches beyond engineering. If organisations value judgement more highly but automate the simpler assignments through which people used to develop it, where does the next generation learn?

My own editorial judgement came from years of reporting, rewriting, getting things wrong, receiving edits and learning when a sentence was technically true but misleading. I wouldn’t want to romanticise every repetitive assignment from that period. Some were inefficient. But not all of the learning was separate from the doing.

An editorial team using AI should therefore ask not only which tasks can be automated, but which experiences newer colleagues need in order to develop source scrutiny, interviewing, argument and accountability. Otherwise we could become more efficient at producing work while weakening the pipeline of people able to assess it.

Productivity is not a capacity plan

One detail in Anthropic’s research caught my eye: median meeting time rose by around 20% over the period studied. That doesn’t establish that AI caused the increase; the business was expanding quickly. But it is a useful warning against assuming that automating implementation removes the coordination around it.

I can recognise the tension in my own week. AI has made it possible to explore customer intelligence, transcript reuse, GEO, video strategy and editorial workflow design alongside conventional delivery. Some of those projects have created tangible value. Others remain experiments. They also generate decisions, dependencies and handoffs that don’t vanish when the research stage becomes faster.

The report strengthens my original argument, but it also makes me refine it. The problem isn’t just that writers have too many adjacent responsibilities. As AI makes more work possible, deciding what deserves human attention becomes a central part of the job. That can be genuinely valuable strategic work. It can also become a way to overload a person with an ever-growing queue of things to inspect.

I don’t think the answer is to resist AI or insist on writing every first draft manually. I want to keep using it, and to keep developing these systems. But productivity gains need a second decision: what will we do with the time, what new responsibilities are justified, and what will we deliberately leave undone?

I want the wider work and the writing to be viable

I’m excited by how much more I can do now. Editorial Intelligence would be much harder for me to develop without AI, and some of the most interesting parts of my working life sit in the connections between evidence, strategy, writing and systems.

I also know what careful writing asks of a person. It needs enough attention to notice the convenient claim, enough curiosity to return to the source and enough room to change direction when the first argument fails.

AI can help create that room. It can also help us fill it with additional obligations before the writing begins. Which outcome we get depends partly on how we define the work, value the package and decide what matters most.

If we want care in the finished writing, we have to make room for it in the working week.