AI for product marketing: five stages, and the two AI should not decide

What does the loop actually look like?

Builders Camp teaches product marketing as a loop rather than a launch checklist: market insight feeds narrative and positioning, which shapes product exploration, which produces a launch, which produces learning that goes back into market insight. Five stages, and the reason the loop framing matters is that most teams run it once and stop at stage four, treating the launch as the end rather than the point where you find out whether stages one and two were right.

Adoption is no longer the question. Jasper's State of AI in Marketing 2025, a survey of 500+ marketers, found 63 percent already use generative AI in their marketing and another 27 percent plan to evaluate it within six months. The same survey puts output quality (19 percent) second among the barriers to scaling it, behind data privacy (21 percent). Jasper sells an AI marketing tool, so read the adoption figure as a ceiling rather than a census; the quality barrier is the more telling number, because it is the one a vendor has no reason to overstate.

A model is uneven across those five stages, and the unevenness is predictable. It is strong wherever the task is compressing a pile of text into structure, and weak wherever the task is choosing what to be. Those two categories map almost exactly onto the stages.

Where AI earns its place, stage by stage

Stage	What a model does well	What stays yours
Market insight	Clustering reviews, tickets and interview notes into segments, needs and jobs	Deciding which segment you are actually going after
Narrative and positioning	Generating variants, stress-testing a claim with counter-arguments	The category you compete in and the claim you will defend
Product exploration	Turning a positioning statement into feature implications and open questions	Which implication is worth building
Launch planning	Timelines, roles, checklists, asset drafts, enablement material	Whether this release deserves a launch at all
Learning and iteration	Compressing adoption data and feedback into signal	Deciding what the signal means for the next loop

The pattern in the right column: every item is a choice with a cost, made under uncertainty, where being wrong is expensive and the evidence does not settle it. That is the definition of the work a model cannot take off you, however few of the total hours it takes.

Which product marketing activities should AI automate, and which should stay human?

The stage table is the loop view. Day to day, product marketing is a list of recurring activities, and the automate or keep human split runs through each one:

Activity	Automate or draft with AI	Keep human
Market and customer research	Summarising analyst notes, forum threads and reviews; tagging sentiment; clustering open-ended survey answers	Deciding which finding changes the plan
Competitive intelligence	Monitoring competitor pricing and feature pages; first drafts of battlecards	Judging what a competitor intends, as opposed to what it published
Segmentation	Proposing clusters from usage and firmographic data	Choosing the segment to serve and the ones to ignore
Voice of the customer	Grouping tickets and interview notes by theme, counting mentions	Reading the ten verbatims behind each theme before trusting it
Pricing and packaging	Modelling scenarios, comparing public price points	The number, and who gets told first
Positioning and messaging	Variants, persona adaptations, counter-arguments	The category claim and the proof behind it
Launches	Timelines, checklists, meeting notes, asset drafts	The launch tier and the go or no-go call
Experimentation	Drafting test hypotheses and copy variants	Which result is worth acting on
Sales enablement	First drafts of talk tracks, objection lists, one-pagers	The final call script a rep will say out loud

The right-hand column is short in words and long in consequence. Every entry is a choice where being wrong costs money or trust, and where the evidence does not settle it on its own. The go-to-market strategy entry applies the same split to the parts of a GTM plan.

The positioning statement, before and after

Given a product brief for a usage-analytics tool, a model produced this:

A modern analytics platform that helps product teams understand user behaviour, make data-driven decisions, and drive growth through actionable insights.

Now run the name swap test. Put your largest competitor's name in front of that sentence. It still reads as true, and so does the next three companies down. What you have is a category description written in the register of the category leader, which is what a model returns by default because that register is what it has read most.

The edit starts by deleting the category and naming the alternative you are displacing:

Most teams answer product questions in a spreadsheet, two days after they needed the answer. This gives a PM the cohort and funnel answer in the tool they already have open, without an analyst in the loop. It does not replace your data warehouse, and teams with a dedicated analytics engineer usually should not buy it.

The second version survives the name swap, because no competitor would write the third sentence. Naming who should not buy it is the single most effective differentiation move available and the one a model will never make unprompted, since nothing in its training rewards turning customers away.

The counter-argument pass

The syllabus for Product Marketing with AI puts stress-testing positioning with counter-arguments in the same module as crafting it, and that pairing is the practical core of using a model here. A prompt worth keeping:

Here is our positioning statement and the three proof points behind it. Argue against it as the head of product at our biggest competitor, then as a sceptical buyer who already has a spreadsheet that works. For each objection, say what evidence would settle it and whether we published that evidence.

The last clause is what makes it useful. Most positioning fails not because the claim is wrong but because the proof is missing, and a model comparing claims to published proof will find that gap reliably. Where a claim has no proof, you have two options: publish the proof, or cut the claim. Softening it is the third option and it is the one that produces the copy everyone ignores.

For the research that feeds this, Perplexity for product managers covers pulling competitor claims with sources attached, which matters here because competitor copy is evidence of what they say and not evidence of what is true. The gap between those two is usually where your positioning lives.

What a model gets wrong about launches

A model plans every release as a launch. Asked for a go-to-market plan, a model produces the full apparatus: announcement post, email sequence, enablement deck, social plan, press outreach, internal briefing. It does this for a feature that changes an export format and for a product line that changes the company, because nothing in the request told it which one this was, and nothing in its defaults asks.

Deciding what deserves a public launch against a lightweight release is an explicit module in the Product Marketing with AI syllabus, and it is the judgment that protects the thing you actually want to protect: attention. Launch four small releases loudly and the fifth one, which mattered, arrives to an audience that has learned to skim you. State the tier in the prompt, and ask for a lightweight plan by default.

When should you not use AI in product marketing?

Some product marketing work should not go through a model at all, even as a first draft, because the draft anchors the decision:

• Strategic choices. The segment, the category and the wedge. A model's first answer is the average of what it has read, and an average is the opposite of a position.
• High-stakes calls in real time. A discount a rep asks for mid-negotiation, a response to a competitor's launch the same morning. The context that decides these is in the room, not in the prompt.
• Messages whose job is trust. The customer email after an outage, the note to accounts affected by a price change. Readers can tell when an apology was generated, and that costs more than the time saved.
• Bias-sensitive targeting. Lookalike audiences and exclusion lists can quietly encode who gets offered what. A person should sign off on who is excluded and why.
• Cultural and emotional nuance. Humour, idiom and tone in a market you do not know. Use a native speaker, not a translation with confidence.
• Anything that manufactures proof. Customer quotes, reviews, case study numbers and testimonials must come from real customers.

The last item is also law in the United States. The Federal Trade Commission's final rule banning fake reviews and testimonials, announced on August 14, 2024, covers reviews and testimonials that misrepresent that they are by someone who does not exist, such as AI-generated fake reviews. "Fake reviews not only waste people's time and money, but also pollute the marketplace and divert business away from honest competitors," said FTC Chair Lina M. Khan in the announcement.

The four edits to make on every AI marketing draft

• Run the name swap test, and rewrite anything that survives it with a competitor's name attached.
• Delete one segment. A model given three audiences will write for all three, which produces copy that is precise for nobody.
• Replace every adjective with a number your product actually produces, and where no number exists, cut the sentence rather than keep the adjective.
• Add the sentence about who should not buy this. It costs one line and does more differentiation work than the three above it combined.

Where this sits alongside the rest of product work

Positioning that cannot be tested is a hypothesis with good typography. If the product is early enough that the segment itself is unproven, how to validate a startup idea is upstream of any of this, and From Idea to Launch with AI covers the same ground as a 1 week bootcamp: validating the problem, tightening scope, and picking distribution moves that get first users without an audience. If the question is what happens after launch, whether the loop is compounding or leaking, that is growth rather than marketing, and Growth for Product Managers covers retention loops, acquisition loops and cohort measurement across 2 weeks.

Product Marketing with AI is the direct match for this page: 1 week, 2 live sessions and 6 microlessons, across five modules that follow the loop from market insight to learning and iteration, with AI used as a copilot at each stage rather than as the decision maker.

See the Product Marketing with AI bootcamp

For the mechanics of keeping these prompts reusable rather than retyping them every launch, ChatGPT for product managers covers the setup. And a habit worth building on top of all of it: save the positioning statement you shipped with, then re-run the name swap test on it six months later. Categories move, competitors copy the good line, and a statement that passed in March frequently reads like everyone else's by September.