Where AI Actually Earns Its Keep in a Marketing Team

Most of the argument about AI in marketing is conducted between two positions that are both wrong. One side has banned it, usually after a bad experience with published content that turned out to be subtly false. The other has routed strategic judgement through it and is now making decisions that sound well-reasoned and are not grounded in anything.

The useful question is not whether to use it. It is which specific tasks move to it, and there is a test that sorts them reliably.

The test

Does the task have a verifiable right answer that a competent person can check in less time than it would have taken to do the work?

If yes, hand it over. The model will be faster, and the verification step catches the failures.

If no — if the output's quality is a matter of judgement, taste, or accountability that someone has to own — then the model can assist with the inputs but cannot produce the output, because there is no cheap verification step and errors do not announce themselves.

That single question resolves most of the disputes I see inside teams. It explains why using it to extract structured data from four hundred reviews is obviously correct, and why using it to decide your positioning is obviously not, even though both feel like "analysis".

Where it pays, concretely

Variation at volume. This is the strongest use case in marketing and it is not "write my ads". It is: here is the single angle I have decided to test, here is the customer language it came from, produce twenty-five headline variants that keep the claim and change the framing. You supply the idea and the judgement. It supplies the permutations, which is genuinely tedious work that humans do worse as the hours accumulate. Given how much of paid performance now rests on creative throughput, compressing this step has a direct line to revenue.

Research compression. Reading two hundred competitor ads, six hundred reviews, or thirty support transcripts and returning a structured list of recurring objections, claimed benefits and price complaints. This is the highest-return use of AI in marketing and almost nobody does it, because it is unglamorous and does not produce a shareable artefact. It produces the raw material for everything else.

Structured extraction. Turning unstructured text into tables. Classify these two thousand search queries by intent. Tag these leads by industry. Pull the price objection out of each of these calls. Verifiable by sampling twenty rows, which is exactly the test above.

Operational QA. Checking that every campaign follows the naming convention. Finding pages missing a meta description. Diffing a tracking implementation against the spec. Boring, rule-based, instantly checkable, and it catches the errors that quietly corrupt your reporting.

First drafts of internal documentation. Process docs, briefs, onboarding material, meeting summaries. Low stakes, high volume, obvious when wrong.

Where it quietly costs you

Positioning and strategy. Ask a model what your positioning should be and you will get a fluent, structured answer assembled from the average of everything written about your category. The average is precisely what positioning has to escape. The output reads like strategy and functions as camouflage, and the cost does not show up for two quarters.

Anything you would publish without reading closely. The failure mode is not gibberish — gibberish is easy to catch. It is a confident, specific, plausible claim that is wrong. A statistic that does not exist, a regulation misstated, a competitor's feature described inaccurately. The fluency of the output is what disarms the reviewer.

Volume content for organic. Publishing forty indistinguishable articles still works occasionally and the trend is firmly against it. Answer engines and ranking systems are both converging on original information as the discriminator, and synthesised summaries of existing material are the one thing they can already produce themselves. You are competing on the dimension where you have no advantage.

Interpreting your own numbers. A model reading your dashboard will produce a tidy narrative for any set of figures, including figures that are wrong. If your platform-reported revenue disagrees with your accounts, AI analysis will simply explain the fiction more articulately. Get the measurement trustworthy before you point anything clever at it.

The customer-language pipeline

If you implement one thing from this piece, make it this, because it converts the strongest capability into the highest-leverage input.

1. Export everything your customers have written or said: reviews, pre-sale questions, support tickets, call transcripts, cancellation reasons, survey free-text.
2. Have the model extract, per item: the problem stated, the hesitation expressed, the alternative considered, and the exact phrasing used.
3. Cluster those into recurring themes with frequency counts.
4. Read the output yourself. This step is not optional and not delegable.
5. Turn the top themes into ad angles, landing page headlines and FAQ entries — in the customers' words, not yours.

Steps one to three took a researcher a fortnight and now take an afternoon. Step four is where the value is created and it is still human. That shape — machine does the volume, human does the judgement, in that order — is the template for every use case worth adopting.

It also feeds directly into page work. The objections that surface at step three are usually the exact reasons a page fails, which makes this the cheapest input available to conversion rate optimisation.

The governance that stops it going wrong

Three rules, and teams that skip them end up back at the ban.

One named owner per workflow. Not "the team uses AI for copy". A specific person owns the output of a specific workflow and answers for it. Diffuse ownership is how unchecked output reaches customers.

A shared prompt library, versioned. When one person works out the prompt that reliably produces good variants, that belongs in a document everyone uses, not in their chat history. Most of the productivity difference between teams is here, not in model choice.

A hard rule on external claims. Nothing with a number, a date, a legal statement or a competitor reference goes out without a human verifying the source. Write it down. The one time it matters, it will have been worth the friction.

Getting a team to this point is mostly a training problem rather than a tooling one. The tools are commodity and roughly equivalent; the difference between a team that gains a day a week and a team that generates confident rubbish is entirely in how the work is divided and checked.

What it does not do

It does not reduce headcount in a marketing team that was already lean. What it does is raise the floor — it removes the excuse for shallow research, thin testing volume and undocumented process, because those were all capacity problems and capacity is now cheap.

Which means the bar moves. When everyone can produce forty competent variants, the advantage goes back to whoever had the better idea to vary, and whoever can tell which of the forty actually won. Both of those remain stubbornly, expensively human, and both are where I would put the time you just saved.