Two boardrooms are making opposite decisions about AI, and both are getting it wrong. In the first, someone has done the arithmetic, concluded that a language model writes a 1,200-word article for pennies, and shrunk the content team to a couple of people who are good at prompting. In the second, AI has been banned outright over quality, privacy or the suspicion that the whole thing is a bubble.

Both rooms think they are being prudent. Neither is asking the question that decides whether their content works: what is left that a competitor with the same tools cannot replicate by lunchtime?

The budget squeeze is real, and it is pushing bad decisions

Marketing has been shrinking for years. Gartner’s tracking puts the average marketing budget at 11% of company revenue in 2020, 9.1% by 2023 and 7.7% today. That is a structural contraction, not a rounding error, and generative AI arrived exactly when finance teams needed a story about doing the same work with less of it.

The story does not survive contact with the rest of the data. A separate Gartner study found 59% of CMOs say they lack the budget to execute their 2025 strategy. If AI were genuinely making teams cheaper and more capable, that number would be falling. Instead, capacity was cut on the promise of a tool.

The question no spreadsheet can answer

There is a difference between being data-led and data-informed, and it decides whether your content strategy survives the next two years. Data-led thinking takes the numbers as the complete picture: AI is cheaper per article, therefore replace the writers. Clean logic, fast decision, no further inquiry. Data-informed thinking treats the same figure as the opening of an investigation. What is this measurement failing to capture? What happens to our output’s effectiveness when every competitor runs the identical play with the identical tools?

That second question has an uncomfortable answer. If you can produce your content with a prompt and minimal human involvement, so can everyone else in your category. Within a year an entire industry is publishing near-identical pieces on near-identical topics, synthesised from the same crawled sources, aimed at the same buyers.

There is no reason a model would write a better article for you than for the competitor who asks it the same question an hour later. Which means AI-generated content is not an advantage. It is the new floor, the baseline your work has to clearly beat before anyone notices it exists.

Content is infrastructure, not a line item

Marketing writing gets treated as the easy kind. In practice a single asset has to do several jobs at once.

• Be genuinely informative, and survive scrutiny from someone who knows the subject.
• Do a specific job in the funnel: attract, nurture or convert, without drifting from the brand’s messaging.
• Demonstrate real expertise rather than assert it, and stay distinctive enough to be remembered as yours rather than the category’s.
• Be structured for machines, with the right entities, topics and relationships, without losing a human reader who can leave at any moment.
• Contain something quotable: a line, a statistic, a framing worth repeating elsewhere.

A model handles the first item well. Expecting it to spin all six plates at the standard of an experienced creator is where the disappointment starts. However detailed the prompt, something goes missing, because you are asking a system to assemble something distinctive out of material that is already public.

And there is the irony at the centre of the debate. With AI-mediated search shaping how buyers find anything, your content’s new job is to be the source a model cites. Asking AI to generate content that AI will want to cite is a dog chasing its tail: if the model supplied the information, it already has more authoritative places to get it from. Citations go to material carrying something the model could not otherwise reach, which is what we cover in what content needs to look like to earn AI citations.

The internet is not the sum of human knowledge

Models cannot use information that was never crawlable, and the gap between “on the web” and “known by people” is enormous. A model can tell you when the local collectables market runs, but not which dealer has the hard-to-find comic you have been hunting for years, because that is only discoverable by turning up and digging through boxes. It can summarise the First World War in detail and struggle badly on the Iranian famine of the same period, which survives mostly in Persian-language histories and family memory. A grandmother’s account of living on one almond a day is not in any index.

Your business has its own version. The real reason a deal was lost. The workaround the implementation team invented in year three. The pattern a senior specialist has seen forty times and never written down. None of it is online, all of it is raw material nobody can copy, and none of it can be prompted into existence.

There is a second problem: training data skews heavily English, baking in cultural bias that is easy to miss when the output reads smoothly. If you publish across markets, that distortion is a strategy risk, as we examine in why AI visibility strategies break outside English.

What the law firms figured out first

The useful precedent is not in marketing. A survey by the Australian Financial Review found most law firms had adopted AI tools, and that 70% of them increased lawyer hiring, because someone qualified has to review and sign off on what the machine produces.

That is not a failed AI strategy. Headcount reduction was never the point: the tools handle research and drafting, and the humans get more time for the judgement work that earns the fee. The same logic applies to content teams. AI is a force multiplier, best pointed at the drudgery around the real work:

• Summarising dense research and transcribing interviews with internal experts.
• Building outlines and structural options.
• Drafting derivative assets such as social posts, and checking finished work against the brand style guide.
• Producing a throwaway rough draft purely to get past the blank page.

Every one saves hours. None of them is the thinking. If results have slipped since you leaned on these tools, the cause is usually the workflow rather than the model, a point we unpack in why AI content alone will not fix your rankings.

The cuts you cannot easily reverse

Britain has an expensive lesson in cutting infrastructure on narrow metrics. In the 1960s the government appointed Dr Richard Beeching, a physicist from ICI with no transport background, to make the loss-making railways profitable. He assessed each route in isolation on passenger numbers and operating costs, then closed whatever failed. Between 1963 and 1970 that logic shut 2,363 stations and more than 5,000 miles of track, roughly 30% of the network, with 67,700 jobs lost.

The Edinburgh to Carlisle line went in 1969. On paper it was unprofitable; in practice it carried access to jobs, education and services across the Scottish Borders, and closing it pushed those costs onto roads rather than eliminating them. Thirty miles reopened in 2015 as the Borders Railway at an estimated £300 million for seven stations. Beeching solved the problem he was given and undermined the purpose of the thing he managed.

Content teams are vulnerable to the same error, because the value that disappears when you cut them (institutional context, editorial judgement, brand voice, the relationships that get subject matter experts to talk) is exactly the value your reporting never itemised.

Rebuilding is harder than dismantling. Experienced creators do not wait around to be re-valued; they move on, and each cohort that leaves takes the apprenticeship route for the next one with it. Add the months of underperforming output in between and reversal costs multiples of the saving.

Frequently asked questions

Should we use AI to write our articles at all?

Use it for structure, summarising, transcription and rough drafts, then treat the output as raw material a knowledgeable person shapes, verifies and adds to. Publishing it largely untouched puts you at the category baseline, where differentiation is impossible.

Does AI-generated content get penalised in search?

Production method is not the deciding factor; usefulness and originality are. The practical risk is invisibility rather than penalty, because content that recycles what is already indexed gives nothing a reason to surface it.

How do we prove content value when budgets are being cut?

Stop defending it on cost per asset. Report the outcomes narrow metrics miss: assisted pipeline, branded search demand, citations and mentions in AI answers, and sales cycle influence. If your reporting only counts publishing volume, you are handing finance the case for cutting.

What content can AI not replace?

Anything grounded in proprietary experience: customer interviews, internal expert knowledge, original data, first-hand results, and opinion attached to a named person with a track record. That is what human hours are worth spending on.

Before you swing the axe

Your spreadsheet is right that AI-generated content is cheaper. It is silent on whether that content will accomplish anything, and the silence is being read as approval. Content is the infrastructure of your brand’s discoverability, which matters more in an AI-mediated search environment, not less, including the awkward possibility that a thin strategy leaves models recommending your competitors instead of you.

The real question was never human versus machine. It is whether you are equipping skilled people with better tools, or dismantling the system that makes your brand findable and credible.