AI in the PMM Workflow: What Actually Works (And What Reps Spot Instantly)
Turn this article into takeaways for your work.
Each assistant summarizes the article only for you and suggests best practices for your work.
Every PMM tool now claims AI. Most of them regurgitate the same marketing copy a competitor's intern wrote, then ship it back to you as "competitive intelligence." Your reps can smell an AI-written battlecard from across the Slack channel. Customers spot AI-written one-pagers before they finish the headline.
The PMM job is judgment, voice, and customer truth. Three things AI is genuinely bad at.
But there are five places it saves you a full day a week. The trick is knowing which five, and being honest about the rest. I've watched a top AE tear up an "AI-generated battlecard" in front of me, hand it back, and ask if we could "send the real one." So this guide is opinionated. The goal isn't to use more AI. The goal is to use AI for the parts of PMM that are pattern-matching, and to keep human hands on the parts that aren't.
Where AI Actually Helps
Five workflows, in rough order of ROI.
1. Win-loss transcript analysis. This is the highest-leverage use of AI in PMM, full stop. Twenty Gong calls used to mean two weeks of listening, tagging, and trying to remember what the prospect said in call 14. Now you paste 20 transcripts into Claude with a structured prompt and get coded themes in 30 minutes. You'll still validate with reps before you ship anything, but AI compresses the synthesis from weeks to hours.
2. Competitive intel summarization. Reading 14 G2 reviews, the last three earnings transcripts, and a competitor's new pricing page is grunt work. Drop them into Claude and ask: "What changed in the last 90 days, what objections did reviewers raise, what new features did they ship?" You get a cleaner starting point than any "AI competitive intel" tool will sell you for $40K a year. Note: the input is the source material. The output is your own writeup, not a forwarded summary.
3. Message-test variants. You're running an A/B headline test on a launch page and need 8 variations. Generating them by hand takes an hour and produces six near-duplicates because your brain ruts. Generating them with AI takes four minutes and gives you eight directions, half of which are bad on purpose because variance is the point. You judge them as a human. AI just expanded the search space.
4. Sales asset first drafts. Battlecard skeletons. FAQ stubs. Objection-handling outlines. The structural skeleton of a one-pager. AI gets you to a v0 in 15 minutes that would have taken 90. Then you rewrite 70% of it. The savings is in not staring at a blank doc, not in the words AI generates.
5. Briefing memos. A messy product spec needs to become a launch brief for marketing, sales, support, and CS. Paste the spec, paste your launch brief template, ask AI to draft. You'll cut half of it and rewrite the rest, but you skipped the part where you re-read the spec three times trying to figure out where to start.
That's the list. Notice what's not on it.
Where AI Breaks (And Reps Notice)
These are the failure modes I see PMMs get burned on. Each one costs trust with sales, and trust is what gives PMM leverage.
Positioning judgment
AI averages everything. Positioning is a sharp choice.
When you ask Claude "what's the best positioning for our product," it will give you something that sounds reasonable and is therefore wrong. April Dunford's whole point is that positioning requires picking a competitive alternative, picking who you're best for, and explicitly leaving everyone else behind. AI hates leaving anyone behind. The model wants to please. Positioning requires the courage to disappoint.
Anti-pattern: AI-generated positioning statements that say things like "the leading platform for forward-thinking teams." That's not positioning. That's the noise positioning is supposed to cut through.
Brand voice
Tone is taste. Taste is not a prompt.
You can feed Claude 30 of your best blog posts and ask it to write in your voice. It will produce something 70% there. The 30% that's missing is the part that makes a reader stop scrolling. Brand voice is the result of hundreds of small judgment calls: which adjective to cut, when to break a rule, when a half-sentence lands harder than a full one. AI smooths all of that out.
I'm not saying don't use AI for drafts. I'm saying do not let AI ship copy under your brand without a human last pass. Readers can tell. Especially the ones you want.
Customer truth
AI invents quotes. Customers don't talk like that.
If you ask a model to "write a customer testimonial about our onboarding," you'll get something that sounds like a customer testimonial. It will not sound like an actual customer, because actual customers say "yeah, the implementation took longer than we thought, but the SE was great" and AI says "the platform's intuitive interface accelerated our team's time-to-value."
Reps have a sixth sense for this. So do prospects. The minute someone reads a quote that sounds AI-written, every quote on the page becomes suspect.
Anti-pattern: AI-generated "customer voice" anything. Quotes, case study openers, testimonial pull-quotes. Pull from real call transcripts or don't use them.
Original POV
If your launch narrative could've been written about a competitor, you don't have a narrative.
AI is trained on everyone else's launch copy. Asking it to write yours produces the average of everyone else's launch copy. The launches that move pipeline are the ones with a sharp opinion the category isn't already saying. That opinion has to come from your customer interviews, your win-loss data, your founder's actual belief about why this matters now. Not from a prompt.
The Gong + Claude Combo for Win-Loss
This is the single workflow that justifies AI's existence in PMM. Here's the flow I run.
Step 1. Pull the transcripts. Grab 15-20 closed-lost call transcripts from Gong (or Chorus, or Fathom). Filter for deal stage = "Proposal" or later, so you're getting calls where the buyer actually evaluated. Strip names and company identifiers. You can do this manually for 20 calls or use Gong's export with a small script.
Step 2. Paste into Claude with a structured prompt. Long context matters here. Claude handles 15-20 transcripts in one session better than ChatGPT, which is why I default to Claude for synthesis and ChatGPT for variant generation. The prompt structure I use:
You are analyzing 18 closed-lost call transcripts from sales calls.
Each transcript is delimited by "=== TRANSCRIPT N ===".
For each transcript:
1. Identify the deal stage where the loss happened (discovery, demo, proposal, late-stage)
2. Code the primary loss reason using these categories:
- Price / TCO concern
- Feature gap (specify which)
- Competitor preferred (specify which, why)
- Internal champion lost / org change
- Bad fit / disqualified
- Timing / no decision
3. Pull 1-2 verbatim quotes that support the coding (preserve exactly)
After all transcripts, produce a synthesis:
- Top 3 themes by frequency
- For each theme: 3-5 verbatim quotes, deal-stage distribution,
and which competitors came up most often
- Surprises: anything that contradicts our current narrative
Do not invent quotes. If a transcript doesn't support a theme, say so.
Step 3. Validate with reps before shipping anything. This is non-negotiable. Take the top 3 themes Claude surfaced and run them past 3 AEs. Two questions: "Does this match what you're hearing?" and "What's missing?" Half the time, the rep will tell you a theme that didn't show up in the transcripts because reps know context the transcript doesn't capture (like the budget conversation that happened over text).
Step 4. Ship the synthesis with sourced quotes. Every theme in your final writeup gets attached to a real verbatim quote and a call ID. No paraphrased "customers tell us..." nonsense. If you can't quote it, you can't claim it.
This is the workflow that took me from "we should do win-loss" to "we ship a win-loss readout every quarter." AI compressed the analysis. Reps validated the truth. The combination is what makes it credible to leadership.
The "AI-Written Battlecard" Trap
Reps spot AI battlecards instantly. Three reasons.
Generic comparisons. AI produces "we are easier to use, more flexible, and better for growing teams." Every competitor's AI battlecard says the same thing. A real battlecard says "when they bring up X, here's the trap question; when they ask about Y, here's the demo move; when they cite the Acme case study, here's the followup." That's deal context AI doesn't have.
No inside knowledge. Real battlecards include things like "their head of sales just left for a Series B in February" and "they don't support Workday SSO out of the box; you'll need this exact answer." AI doesn't know the head of sales left. It has no signal on the SSO gap unless you feed it both, in which case you're already doing the work.
Missing the deal context reps need. A battlecard's job is to win a specific kind of deal: the mid-market deal where the prospect already evaluated the cheaper competitor and is hesitating on price. AI doesn't know that's the deal. It writes a generic comparison page.
What to do instead: Let AI draft the structure (sections, headers, comparison table skeleton). You fill in the trap-questions, the landmines, the names of competitor reps who close differently, the recent product changes. AI gets you a v0 in 10 minutes. The rest is human. Your reps will use the result. They'll throw the AI-only version away.
The Practical Stack
What a PMM IC should actually run with.
| Tool | What it's for | Why it's on the list |
|---|---|---|
| Gong or Chorus | Call recording + transcripts | The data layer for win-loss and message validation |
| Claude | Long-context synthesis (win-loss, comp intel) | Best for 20+ transcript analysis in one session |
| ChatGPT | Quick variants, headlines, FAQ stubs | Faster iteration on shorter tasks |
| Crayon or Klue | Competitive monitoring (raw signal) | Use as input feeds, not as your output |
| Notion or Coda | Human-edited final assets | Where the AI draft becomes a real asset |
What to skip: any "AI battlecard generator" or "AI launch narrative" tool that promises end-to-end automation. The output is exactly the slop your reps and customers will reject. The savings the vendor promises evaporate the first time a deal stalls because the rep brought a generic comparison page to a sharp prospect.
A 30-Day Plan to Adopt AI Without Becoming Slop
Don't try to AI-ify everything at once. The PMMs who get this right pick one workflow, get it sharp, then add the next.
Week 1. Pick one workflow and build a reusable prompt. Win-loss is the highest-ROI starting point. Build the prompt structure above. Run it on 10 transcripts you've already analyzed manually so you can compare AI's output to your own and calibrate. Save the prompt to a Notion doc with notes on what worked.
Week 2. Ship one AI-assisted output. Get rep feedback. Log what failed. Take the win-loss synthesis to a quarterly business review or sales kickoff. Watch how reps respond. Note every place a rep pushes back, every theme that didn't survive validation, every quote that felt off. This is your training data for round two.
Week 3. Build a prompt library for your top 3 recurring tasks. Win-loss synthesis. Competitive landscape summary. Launch brief draft. Each gets its own prompt template, refined based on what week 2 taught you. Store them somewhere your team can use them, like Notion, a shared doc, or a Claude project.
Week 4. Define your team's "no-AI zones." Write them down. Mine: positioning documents, founder narrative, customer quotes, tier-1 launch copy, anything with my name on it that goes to a customer. The point isn't to ban AI from drafting. The point is to be explicit that final voice and judgment stay human, and that anyone on the team can call out an AI-written customer quote without it being awkward.
By day 30 you've got two or three workflows running cleanly, a prompt library, and a clear line between AI-drafted and human-shipped. That's the entire goal.
Closing
AI is leverage for the boring 60% of PMM work. Synthesis, summarization, drafting, variant generation. That's the stuff that's pattern-matching at scale. Use it. Save the day a week.
The 40% that wins deals — positioning, voice, customer truth, original POV — is still your job. Reps can tell. Customers can tell. Search engines are starting to tell.
The PMMs who treat AI as a force multiplier on judgment win. The ones who treat it as a replacement for judgment ship slop, lose rep trust, and get quietly worked around. Pick the first one.
If you're touching the ACE Framework, most PMM AI use cases sit in Analyze (win-loss, competitive synthesis) and Generate (asset drafts, message variants). Resist the urge to frame positioning as a Generate task. It's not pattern completion. It's a sharp human choice about who you're for and who you're not.
Learn More
Principal Product Marketing Strategist