What AI Actually Does Across the Funnel

Marketing took to generative AI faster than most functions, and for good reasons: the work is text-heavy, the volume demand is real, and the cost of an individually mediocre output is usually low.

It is also the function least able to say whether it worked, and that is worth confronting at the start.

The reason is structural rather than a failure of rigour. Marketing outcomes are influenced by seasonality, competitor behaviour, pricing, product changes, macroeconomic conditions and the rest of the marketing mix, all moving at once. Attributing a change in results to a change in how the copy was produced is genuinely hard, and the existing lesson on the attribution problem covers why.

So the honest position is that most claimed AI marketing gains are unmeasured, and the confident ones are usually measuring output volume rather than outcome.

That has a specific consequence for this cursus. The useful question is rarely does AI improve marketing, which cannot be answered generally, but which specific tasks does it do adequately, at what cost, with what risk. Some of those answers are clear and favourable. Others are not, and the difference matters more than a general verdict.

The measurement discipline is the subject of the third lesson, and it is the part that separates a function that improved from one that produced more.

Mapping capability to stage, with an honest assessment at each.

Research and strategy: synthesising customer interviews, summarising competitor material, clustering feedback. Genuinely useful, and the output is an input to human judgement rather than a decision.

Creation: copy, variants, adaptation across formats and channels, translation. The strongest fit, because volume is the constraint and quality per item is moderate.

Targeting and delivery: who sees what. This is where the regulatory edges live, and where the platform's own optimisation does most of the work regardless of what you do.

Personalisation: adapting content to segment or individual. Real, and it is where the diminishing returns are steepest and the creepiness threshold lives.

Analysis: summarising performance data, drafting reports, surfacing patterns. Useful for description, unreliable for causal claims, which is the distinction that matters most.

And customer interaction: chat, response drafting, routing. Covered in the customer support material, and the same grounding requirements apply.

The pattern: strongest where volume is the constraint, weakest where the output is a causal claim.

The clearest wins share a property: the bottleneck was producing enough variations, and quality per variation was never the limiting factor.

Ad variant generation. Testing requires many versions, and producing thirty headline variants was previously a day of work that produced fifteen because people tire. A model produces the thirty, a human selects, and the test is better powered because it has more arms. This is a real capability increase and it is measurable through the test itself.

Channel adaptation. One piece of content becomes a social post, an email, a landing page section and a short-form script. The substance was already decided; the work was reformatting, and reformatting is what these systems do best.

Localisation. First-pass translation and cultural adaptation across markets, reviewed by a native speaker. The economics change substantially when the first pass is free.

Metadata and structured content. Product descriptions across a large catalogue, alt text, meta descriptions, structured data. Tedious, high-volume, and individually low-stakes.

And the long tail. Content for segments, products or queries that were never worth a person's time. This is the genuine expansion rather than a speed-up: work that was not being done at all.

The common thread is that a mediocre output is worth more than no output, which is exactly when volume generation pays.

The same capability produces the function's characteristic failure mode, and it is worth naming because it is currently widespread.

When producing content becomes nearly free, the constraint moves to distribution and attention, neither of which became cheaper. So the equilibrium is more content competing for the same attention, which is a worse position for everyone including you.

Three specific consequences.

Undifferentiated output. Models trained on the same corpus, prompted similarly, produce similar copy. A market where everyone generates produces convergent messaging, which erodes exactly the differentiation marketing exists to create.

Content that exists rather than earns. Publishing volume for its own sake was already a questionable strategy and it is now cheap enough to do at scale. Search engines have responded to mass-produced low-value content repeatedly, and the existing lessons on how Google ranks the web and on generative engine optimisation cover the mechanics.

And the review bottleneck. If a team generates ten times the content and reviews it at the old rate, standards fall silently, which is the pattern the managing-AI-teams cursus describes as the default outcome.

The practical discipline: decide what you are producing more of, and why anyone would want it. A content calendar that doubled because generation got cheaper is a supply-side decision dressed as a strategy.

The most common quality complaint about AI marketing copy is that it sounds generic, and the diagnosis matters because it determines the fix.

A base model produces the average of its training corpus, which for marketing copy means the conventions of a great deal of undistinguished marketing copy. Prompting can move it, and it moves back, because you are working against the distribution.

This is precisely the case the fine-tuning cursus identifies as fine-tuning's genuine strength: a behaviour and register problem, not a knowledge problem. Examples convey a house voice far more reliably than instructions describe it, and a few hundred pieces of your own best copy is a realistic dataset.

The cheaper intermediate step, and the one to try first: few-shot prompting with several strong examples of your own writing, plus an explicit style guide describing what you avoid. This gets a substantial share of the benefit at no training cost, and it is what the prompting-first principle recommends.

What neither fixes. A brand voice that was never actually defined. Many organisations discover, on attempting this, that their voice was whatever the two people who wrote everything happened to do, and it was never articulated. That is a finding worth having, and it has to be resolved by humans before any amount of prompting or training encodes it.

And note the trade: a strongly encoded voice applied at volume makes the sameness problem worse if the voice itself is conventional.

Creation is largely unregulated. Targeting is not, and marketers are often the last to know which rules apply to them.

The case that surprises people most: targeted job advertising is explicitly inside Annex III of the AI Act as a high-risk employment use. A marketing team running recruitment campaigns is operating in the high-risk tier, and this is frequently news to them because the recruitment budget sits in marketing while the compliance analysis sits in HR.

Beyond that, several regimes bear on targeting.

Advertising in regulated categories, credit, insurance, housing, employment, is subject to non-discrimination rules in most jurisdictions, and platform delivery optimisation has produced documented cases of skewed delivery even with neutral targeting parameters, because the optimiser infers who engages.

Profiling for advertising engages data protection rules on consent and legitimate interest, and consent requirements for tracking are enforced actively.

And the Article 50 transparency duties, applying from 2 August 2026, require disclosure where synthetic content is generated or manipulated, which reaches AI-generated imagery and video in campaigns.

The practical instruction: if a campaign targets in a regulated category, involve someone who knows the rules before it runs rather than after. The marketing function's usual speed is a liability precisely here.

Marketing copy makes claims, and claims are regulated in a way that most marketers know well and most generative deployments ignore.

The risk is specific. A model generating product copy will produce plausible specifications, benefits and comparisons, and it has no access to whether they are true. A generated line stating a product is the fastest in its class, or reduces costs by forty percent, or is certified to a standard, is a claim your organisation is now making, with no basis and no record of where it came from.

Consumer protection law across jurisdictions prohibits misleading claims, comparative advertising is separately regulated, and in regulated sectors, financial services, health, food, the constraints are considerably tighter and specific claims may require substantiation held on file.

So the design rule for any generative marketing pipeline: factual claims must come from a source, not from the model. In practice this means supplying the specifications, approved claims and substantiation as retrieved context, and instructing the model to use only what it is given, which is the grounding pattern from the hallucinations cursus applied to a commercial rather than an informational risk.

And the review step has to check claims specifically rather than reading for quality. A reviewer assessing whether copy reads well will pass a fluent false claim, because fluency is exactly what the model is good at.

The organisations that get caught here are not being careless. They are reviewing for tone.

An adoption order for a marketing function, ordered by value per unit of risk.

First, the volume work where quality per item is moderate and the review is cheap. Ad variant generation for testing, channel adaptation, localisation first passes, metadata and catalogue content. These are the clearest wins and none of them makes a claim or targets anyone.

Second, research synthesis. Clustering customer feedback, summarising interviews, digesting competitor material. The output informs human judgement rather than reaching a customer, so the risk is low and the time saved is real.

Third, brand voice work, starting with few-shot prompting and moving to fine-tuning only if consistency at volume demands it and you have the examples.

Fourth, with grounding and a claims-specific review, anything customer-facing that makes factual assertions.

And with actual legal involvement rather than a self-assessment: anything that targets in a regulated category, and any recruitment advertising, which is high-risk under the AI Act whatever the marketing team assumed.

The sequencing puts the measurable, low-risk capability first, which also builds the credibility to argue for the harder cases. And it defers the two things that produce most of the incidents in this function: unsubstantiated claims, and targeting nobody checked.