# Product Marketing Should Own Your Company's AI Knowledge > The honest playbook for AI in product marketing: which prompts work, where they break, and why PMM should own the company's AI knowledge. - Source: https://calven.ai/resources/ai-for-pmm/how-to-use-ai-for-product-marketing - Author: David Kolinek - Published: 2026-07-18 - Last updated: 2026-07-18 - Tags: AI for PMM, product marketing, MCP, prompts, positioning, competitive intelligence --- Nearly every product marketer now has a chat window open in a second tab, and almost none of them get more than a competent first draft out of it. Adoption is total. The payoff is thin. The honest answer to how a PMM should use AI comes in two parts, and most teams only see the first. Part one: treat AI as a fast first-drafter on the work you already own, and stop expecting a chatbot to know your company. What separates a useful draft from a generic one is not the prompt. It is grounding. Part two is the job most teams miss. [AI is reshaping the role](/resources/ai-for-pmm/how-ai-is-changing-product-marketing), and product marketing already owns the knowledge every other team's AI is starving for, so putting that knowledge in front of the whole company's AI is part of the job now. ## Which product marketing tasks can AI actually handle today? AI reliably handles first-pass structuring and drafting. It does not do the judgment. That line, drafting yes and judgment no, is the whole honest boundary, and it holds across nearly every use case a PMM reaches for. Adoption already reflects it: **96% of B2B marketers use AI**, per [Demand Gen Report's 2026 B2B Trends Research Report](https://www.demandgenreport.com/industry-news/feature/demand-gen-reports-2026-b2b-trends-research-report-is-live/52002/). The question stopped being whether to use it and became what to trust it with. Hand it raw material and it shapes that faster than you can. Paste a competitor's homepage and pricing page in with a prompt like _analyze their positioning and target segments, then flag where they claim an advantage we should counter_, and you get a structured first-pass teardown in seconds. Ask for _three positioning options using the competitive alternatives, unique attributes, and value_, the components April Dunford's framework leans on, and it hands you drafts to react to faster than a whiteboard, which beats staring at a blank [positioning statement](/resources/product-marketing/how-to-write-a-positioning-statement). Drop in twenty call transcripts and it clusters the recurring pains and objections at a pace no human matches. Point it at closed-deal notes and it sorts them by [win and loss reason](/resources/product-marketing/win-loss-analysis) into a first-pass categorization you sharpen instead of build. Even competitive research holds up as a starting point, which is why so many PMMs [reach for ChatGPT or Claude to do it](/resources/comparisons/chatgpt-claude-competitive-intelligence). None of that is judgment. It is structuring, extraction, and a first draft. The PMM still decides which option is right, which theme matters, which objection is signal and which is noise. Use AI for the pass that used to eat an afternoon, and keep the decision for yourself. ## Why do those prompts keep breaking down? Every one of those prompts breaks at the same seam: it assumes knowledge the model does not have. Your real positioning, your actual ICP, a competitor's current pricing, the correction you made last week. These are not prompt-quality problems you can write your way out of. They are structural, and there are only a few of them. The model has no grounding in your company, so the positioning it drafts is a plausible average any competitor could claim, and the persona it critiques your value prop as is a training-data stereotype, not the buyer you built from real calls. It has no memory across sessions, so last week's corrected framing is gone the moment you open a fresh chat. It cannot monitor anything, so it never notices a rival changed its pricing or a deal just closed, which is how a [battle card](/resources/competitive-intelligence/battle-cards-101) goes stale three weeks before anyone checks it. And when it does not know, it does not stop. It invents a competitor feature, a pricing tier, a stat, and in front of a live deal that confident fabrication is a liability, not a draft. Not one of these is fixed by a better prompt. > After you have re-pasted the same positioning, the same ICP, and the same competitor notes into one fresh chat after another, you stop believing the problem is your prompt. The model is not short on intelligence. It is short on your context. Once that lands, the skill worth building stops being a cleverer prompt and becomes grounding the AI in what your company already knows, so nobody has to paste it in again. > > — David Kolínek, Co-founder at Calven There is a second, quieter failure underneath. Even a great prompt collapses the day you forget to paste the latest positioning, and the context does not travel. It lives in the PMM's head and nowhere else, so the sales rep, the BDR, and the CS manager who never had it get the ungrounded version by default. ## What actually fixes it: grounding AI in one shared source of truth If every failure traces to missing grounding, the fix is to stop supplying the context by hand and put it where the AI reads from every time. One place holding your real knowledge, positioning, messaging, personas, ICP, competitors, win-loss, so answers come from your truth rather than the training-data average. That shared base is also the memory the chat window never had: a correction made once persists, and every future answer inherits it. Connecting any AI tool to that source is now standardized. Anthropic introduced the [Model Context Protocol in late 2024](https://www.anthropic.com/news/model-context-protocol), an open way to plug an AI tool into a knowledge source without a bespoke integration for each one, and within a year [OpenAI, Google, and Microsoft had all adopted it](https://blog.modelcontextprotocol.io/posts/2025-11-25-first-mcp-anniversary/). It is how a governed knowledge base becomes something any model can query, and querying verified context reins in a model's tendency to invent an answer. This is the shape of what Calven builds. Its agents ground in a shared knowledge base, the One Brain, and answer from your company's real, current knowledge instead of the model's average. The [Competitive Intelligence Agent](/platform/research#competitive-intelligence) tracks a rival's site and pricing and folds what changed into the dossier, so the teardown stays current. The [Positioning Agent](/platform/messaging#positioning) works from your actual market and audience intel, not a generic frame. It is the same principle Calven's own marketing runs on, agents working against a shared repository, the practice of running a team as [marketing as code](/resources/thought-leadership/marketing-as-code). ## Why product marketing should own the company's AI knowledge Product marketing should own it because it already owns the canonical version, and every other team's AI is describing the same company from a fragmented copy. Sales has one version of the pitch, support another, the website a third, a new hire's chatbot a fourth assembled from public scraps. Point an ungrounded model at that and it describes the company inconsistently, because there is no single truth to read. Enablement vendors like Highspot and Glean have named the same need from the sales side, then located that truth in the enablement platform. It does not live there. It lives in product marketing. Look at what the function demonstrably owns: positioning, messaging, the [personas and ICP](/platform/audience#icp) built from real buyer conversations, the competitive intel, the win-loss patterns, the product truth. That is precisely the knowledge every team's AI needs and does not have, and [product marketing](/resources/product-marketing/what-is-product-marketing) is the only function that already maintains the correct version of it. Which makes the shared base a PMM asset, and making it AI-available a PMM responsibility. Not a stretch of the mandate. The most direct expression of it. The objection writes itself: product marketing is already stretched thin, and this sounds like a second job on top of the first. It is not. It is the first job done once instead of every session, since the context a PMM re-pastes into chat after chat is the same context the base holds permanently. Governance cuts the same way: knowledge copied into a dozen tools nobody can audit is the exposure, and one governed source is the control. The first move is small. Put your own positioning and messaging where your AI reads them, start with a single agent, and let the rest of the company follow the proof. ## What changes when every team works from the same brain Once that knowledge is exposed through MCP, every team's AI answers from the same source instead of guessing, and the company stops contradicting itself. The content team's AI drafts on-brand because voice and messaging are inputs, not something it approximates. A sales rep's AI answers a competitive question with your real positioning and the current battle card, not a fabricated feature that falls apart on the call. A BDR's AI personalizes against the actual ICP instead of a generic segment. The product team grounds a roadmap call in real market and competitive intel rather than the loudest opinion in the room, and a CS manager's AI answers a renewal question with accurate product truth instead of last year's deck. Same brain, five teams, one company speaking with one voice. Messaging moves, so holding that together is a continuity problem. When it changes, Calven flags the affected assets for review, so a battle card or landing page that has drifted off the current position gets caught instead of spreading the old line for another quarter. That is what turns a pile of connected agents into an [AI product marketing platform](/resources/ai-for-pmm/ai-product-marketing-platform) rather than a folder of clever prompts. Both jobs start with AI, and only one scales. The PMM who learns to prompt better optimizes a private tool that stops working the moment they forget to paste the context. The PMM who grounds the company's knowledge and exposes it becomes the reason every team sounds like the same company. Fluvio's 2026 research found **82% of product marketers now face an expectation to use AI, and 72% say it raised the output expected of them**, per [Fluvio's Product Marketing Hiring Trends Report](https://www.fluviomarketing.com/blog-summary/2026-product-marketing-hiring-trends-report-ai-is-hiding-whats-still-broken). The bar rose. Prompting harder does not clear it. Owning the knowledge the whole company runs on does.