Learning Hub

Meet Amy: the PMM agent that doesn't guess

Erinc Karatoprak — Tue, 08 Sep 2026 11:49:43 GMT

Customer interview transcripts go into Claude with one question: what do these buyers actually care about? The answer arrives in seconds. Three themes, five benefit statements, ready to paste into the launch deck. It reads like the work of a strategist who spent a week with those customers. What you can't see is which parts came from the transcripts and which the model invented to fill the gaps. It's fast, but not accurate...

It's 18:40 on a Tuesday and Sara is still at her desk. Frustrated. Listening to call recording 9 of 14. Five to go. That's two hours if it goes well. Maybe three.

The launch messaging is due Thursday, and she knows everything she needs is already in these recordings. Real buyers, in their own words, saying why they bought, what almost stopped them, which phrase made it click. It's in there. Somewhere. So she keeps going, because she doesn't want to base her new messaging on nothing. Or on the opinion of the CEO.

So she digs. And digs. And digs. Finds a good quote, pastes it into a doc. Finds another one, adds that too. By call eleven the doc is a pile of quotes and she's lost the thread. Who said this one? A best-fit customer, or someone who churned? Those two point at opposite messaging. She doesn't know anymore. She'll sort it out later.

Later is Thursday. The messaging has to go live. So it goes live, and it's probably fine (is it?). She just can't say why.

Sara is not a real person. She is also not invented. She is assembled from our own buyer research: interviews with product marketers and product managers, all describing the same Tuesday. She is a senior product marketer at a mid-sized B2B SaaS company that inherited customer research without a background in qualitative research and without inheriting a research team. Her company can tell you the exact drop-off rate on step three of onboarding, to two decimal places, in a dashboard. It cannot tell you why a single customer bought and how they talk about a certain product.

If you have "product" in your job title, product marketer, product manager, or something close, you know her Tuesday because it is probably yours: listening to sales calls, running customer interviews, building the roadmap or the messaging from "what customers told us." That is qualitative data analysis, even when nobody calls it that.

And what Sara wants is not complicated. She wants to be fast and accurate, not one at the cost of the other: messaging she can ship this week and still defend when someone with budget authority asks how she knows.

Which is exactly the problem. So on Wednesday morning, she does the reasonable thing.

The assistant that never says "I'm not sure"

She pastes four transcripts into Claude and asks what her buyers care about.

The answer arrives in seconds. Fluent, structured, confident: three themes, five benefit statements, ready to paste into the deck. It reads like the work of a strategist who spent a week with her customers.

Then she looks for the seam: the line between what the model actually found in her transcripts and what it invented to fill the gaps. There isn't one.

One of the five benefit statements reads "cut reporting time in half." It is the sharpest line in the set for messaging, the one she would lead the launch deck with. She goes back through the transcripts. Nobody said half. One buyer said reporting used to eat most of her Friday and now it doesn't. That became a number, and the number arrived in the same confident sentence as everything else on the page. Somewhere in those nine seconds the model also decided two offhand remarks were a theme, and it did not mention that it had decided anything.

There are precise names for what she is looking at. Fluent fabrication: a plausible-sounding claim the data does not support. Silent interpretation: a judgment call the model makes without telling you. A general-purpose AI assistant commits both daily, and neither leaves a mark on the output.

Polished and true are not the same thing.

The people we interviewed felt feel it too. One product manager described his own do-it-yourself AI setup as "a bit of confirmation bias... writing a prompt from the product's hypothesis." The tool was answering the question his prompt implied. That is the villain of this story.

Sara also knows what is riding on the difference. In a previous article we told the story of a campaign built on one ungrounded positioning call: clean launch, sharp creative, almost zero pipeline, well into six figures gone. That was the human-speed version of the failure. The agentic era scales your bad execution exactly as fast as your good execution. We argued then that agents are already deciding, and the open question is what they decide on. For Sara, that question is not abstract. It is due Thursday.

The fix is not to stop using AI. She will not, and she should not. The fix is to ground the AI in verified customer truth, and to make its uncertainty visible instead of hidden.

So we built the colleague Sara needed. And we packaged her so you can run her yourself.

Meet Amy

Amy is a senior product marketing manager you run inside Claude or Gemini. You set her up once, tell her where your company context, and from then on she works with you on positioning, messaging, pricing, launches, sales enablement, and measurement. She is direct, she gives you a clear recommendation with the reasoning behind it, and she does not let weak strategy through.

Plenty of AI assistants/agents will do those tasks. What makes Amy different is what she refuses to do while doing them:

Every factual claim carries a confidence label: HYPOTHESIS, EMERGING, or VALIDATED. She never presents a hypothesis as validated, and she never upgrades a label silently.
Confidence is a count, not a feeling. Evidence, to Amy, means a finding, a verbatim customer quote, and a count: said by 7 of 12 churned accounts. A claim without that behind it is a hypothesis, and she says so out loud.
She names the seam. When part of an answer is evidenced and part is assumed, she tells you which is which, instead of handing you a fluent blend of both.
Missing research never stops the work. Thin data? She keeps going, from clearly labeled assumptions, with a list of the research that would confirm or kill each one. She never refuses, and she never pretends an assumption is a fact.
She pushes back. She will not say "looks good." If work is strong, she names why. If it is weak, she names where, and what to fix.

Sara's first session with her sounds like this. 👇🏻

Sara: Should we lead the launch messaging with the integration story?

Amy: HYPOTHESIS, and I would not lead with it yet. The March win/loss report supports speed-to-first-value as the decision driver (said by 6 of 9 interviewed buyers: "we picked the one that was live before the pilot ended"). Integrations came up twice, both times as a checklist item, not a reason. If you want integrations to lead, that is a research question, and I can scope it.

None of this discipline is a personality quirk. It is build on the Bubble Research Methodology, the same method our product runs on, shipped inside Amy as part of her operating instructions: how research is planned, what counts as evidence, how findings get analyzed and reported. Not a promise in a README. The doctrine the rest of her behaviour hangs on.

What you can hand her on day one

Sara's problem this week is messaging. Yours might be pricing, a launch, or the battle card sales keeps asking for. Nine capability areas ship with her:

Customer insight and ICP. ICP profiles and persona cards built from research, a confidence label per section, and an explicit negative ICP: who you do not target.
Positioning and messaging. Positioning run backwards from your real customers, and a messaging house where every message traces to the positioning canvas.
Pricing and packaging. Recommendations built on value, competition, and willingness-to-pay, treating pricing as a positioning signal.
Launches. Honest tier classification, wave planning (learn before you amplify), one-page briefs, backward-planned timelines.
Sales enablement. Drop in a call transcript and she extracts the objections, the buyer language, and what landed. Or flip it: she plays the skeptical buyer while you pitch. Battle cards come with honest competitor strengths.
Strategic sparring. Hand her a strategy and ask her to find the weakest link: hidden assumptions, claims without proof, premature scaling moves.
Measurement and retros. Success defined before the initiative starts, vanity metrics pushed back on, retros that end in playbook changes.
Content and copy. Launch copy, press materials, product pages: always in the persona's language, always tracing to the messaging framework.
Frameworks on demand. Eight PMM frameworks (JTBD, SPICED, Dunford positioning, category design, and more), applied to your context rather than recited to you.

The plan: three steps

Step 1: Add her to your Claude or Gemini. We wrote step-by-step setup guides for both. You create a project, paste her instructions, upload her knowledge files. One-time setup, about 15 minutes, in the browser, nothing to install.

Step 2: Bring her your reality. A short State doc tells her your company, your ICP, your positioning, your open bets. That is her memory between chats. Then feed her the real material you already have: research reports, call transcripts, the objection a rep heard yesterday. She treats every one as data and routes it into messaging, ICP, and enablement.

Step 3: Let her keep getting better. Amy improves through a feedback loop: we ship regular updates to her instructions and knowledge, announced in our newsletter. Download the changed files, replace them in your project, and she is current.

That is the whole plan. No procurement, no integration project, no new tool to learn ,and no invoice: Amy is free, included the moment you sign up for Bubble's free plan. She lives inside the LLM app you already use.

If you use Bubble

Everything above works standalone. Paste or upload a research report and Amy reads it the way the methodology demands: she audits the evidence behind the findings she is about to build on, and she surfaces the report's flagged assumptions for you to confirm instead of letting them slide through.

If you run your research on Bubble, one more step makes her better. Connect your Bubble account through the bubble connector in Claude and Amy reaches your company's Bubble workspace (where your research lives) directly: she can list, search, fetch, and verify reports herself, citations included, read-only. Ask her to fetch the latest win/loss report and tell you what it changes in your messaging, and she comes back with the answer, the quotes, and the counts, without you shuffling files.

Same Amy. Shorter distance between a question and the evidence.

One honest boundary either way: Amy does not run your customer interviews (yet). When the research she needs does not exist, she does not guess. She drafts the research request that would close the gap: the decision that is blocked, the current hypothesis, and what result would change the recommendation.

The method she runs on

Everything Amy does traces back to the Bubble Research Methodology, and the shape of it is worth thirty seconds, because it is also the answer to the question someone with budget authority will eventually ask: how do you know?

Why it exists. Most people who do research at work were never trained for it, and you should not need a research degree to defend a finding. The methodology gives product (marketing) people just enough foundation to do genuinely good research, and to survive scrutiny when real money is on the line.

How it works. Codebook thematic analysis with semantic coding: a fixed framework decides the shelves, and the codes are built from your transcripts, in your customers' own words, each anchored to a verbatim quote. AI does the mechanical passes. A human stays in the loop at exactly two checkpoints: audit the evidence, resolve the flagged assumptions.

What it produces. Findings you can defend. Every theme traces down to real quotes, every citation is checked against its source, and belief is kept separate from evidence with confidence labels.

The full methodology is a 52-page whitepaper, and we distribute it for free. If you want the foundation under Amy, or under your own research practice, download it here.

Sara's Thursday, twice

Version one. She ships the nine-second messaging answer. The deck is clean, the meeting goes fine, the campaign runs. The cost arrives quarters later, in a budget review, when someone asks which of those confident guesses actually came from customers, and the room goes quiet. We wrote last time about what that silence is: it is what a failed launch sounds like.

Version two. She walks in with three messages, and under each one: the claim, the customer's own words, the count, the confidence label. Someone with budget authority asks how she knows. She opens the quote. The debate moves from whose opinion wins to what the evidence says, and that is a debate she now runs. Nobody in that room calls what she does a supporting role.

The people in our research describe that second version in their own words. The professional one: "from sharing hearsay to presenting data." And the quieter, personal one: "I did my research, it's not just I was coming up with random ideas."

That second Thursday is the point of all of this. The person with "product" in their job title stops producing opinions and starts owning the customer truth their team runs on, for the humans deciding today and the agents deciding tomorrow.

The agents are already deciding. Amy is how your corner of the business decides on truth.

How to collect, analyze, report and act on customer insigths

anne@bubble-research.ai (Anne van der Kammen) — Sun, 12 Jul 2026 22:03:22 GMT

Every product (marketing) decision is only as good as the customer truth behind it. We believe getting that truth right shouldn't be daunting... and it shouldn't take months. That's why we built a customer research process that's practical, simple, and scalable enough to actually run in your role.

The Agentic Era Rewards Truth, Not Speed

steven@bubble-research.ai (Steven Schroeyens) — Sun, 12 Jul 2026 21:40:37 GMT

Why roughly 95% of companies saw no measurable return on generative AI in 2025, and the two questions every CEO racing to go agentic should bring to their next board meeting. One of them is "Do we actually have the fundamentals?"

Grow to ten million euros in annual recurring revenue (ARR). Ten people. Maximum. That's what our business plan says.

Every time I say that out loud, a little voice in my head does the math and looks at me like I've forgotten how companies work. And that little voice is right. Ten humans do not run a €10M tech business. The headcount doesn't add up, and it's not supposed to. The rest of the company is an agentic workforce. Agents doing the work, making decisions, moving things forward, at a scale and speed ten people never could.

So my co-founders and I spent the last year staking our company on a single bet: that agents can carry the load.

And somewhere in that year we realised the bet isn't really about the agents at all.

The change nobody gets to opt out of

Let's start with what's actually happening, because it's bigger than my company or yours.

Every board on earth has "become agentic" somewhere on the 2026 agenda. This isn't a fad you can wait out. Gartner expects that by 2028, at least 15% of day-to-day work decisions will be made autonomously by AI agents, up from essentially zero in 2024, and that a third of enterprise software will have agentic capabilities baked in. The shift is happening independently of what any one of us decides. Agents are going to be making decisions inside your business. Many of them. Faster than your people, and in far greater numbers.

That's the change. And it comes with a stake attached, the way real changes always do: agents are already drafting the work your people rubber-stamp, and the autonomous share is rising. The only open question is what those decisions will be grounded on. Opting out doesn't remove the stake, either; it only changes how quickly the bill arrives.

The reason 95% of it failed

Start from something every strong company already knows: growth begins with a deep understanding of your customers. The best operators make almost every decision a customer-centric one, most of all in product and go-to-market, which happens to be exactly where most companies spend most of their money. So when we talk about making good decisions, whether it's decisions made by humans or agents, these need to be grounded in deep customer insights and truth.

Now the uncomfortable part.

In 2025, MIT's Project NANDA published The GenAI Divide: State of AI in Business: 52 structured interviews, 153 survey responses from senior leaders, and a review of 300+ real AI initiatives. The headline number went everywhere: roughly 95% of organizations got zero measurable return on their generative-AI investment. Billions spent, a rounding error back. (The authors themselves call it a directionally-accurate six-month snapshot, not a final verdict, and it's drawn its share of pushback, so take the exact figure with a grain of salt. The direction is harder to argue with.)

But don't stop at the number, because the instinct it triggers, the technology just isn't ready yet, is the wrong lesson. Read the why and a different story shows up. MIT found the failures weren't about model quality at all. They were about what the authors call the learning gap: tools that couldn't retain context, couldn't plug into real workflows, couldn't improve over time. Buyers who partnered with vendors and integrated deeply succeeded far more often than teams building disconnected tools of their own: internal builds succeeded only about half as often (roughly 33% vs 67%).

And don't write this off as a 2025 story that newer models have since solved. S&P Global's 2025 survey found firms abandoning most of their AI projects at more than double the prior year's rate, and through 2026, a year of fast model progress, the numbers barely moved: 2026 surveys from PwC and Foundry landed in the same place, with most enterprises still unable to show a return. Capability had stopped being the bottleneck: the frontier models converged, and the gap moved into the workflow, not the model. That's the tell. If a year of dramatically better models doesn't move the number, the number was never about the models.

Gartner tells a compatible story from the agentic side: it predicts over 40% of agentic AI projects will be cancelled by the end of 2027, naming escalating cost, unclear business value, and inadequate risk controls. Different words from MIT's learning gap, but they rhyme: projects die on integration, ownership, and value, not on how smart the model is.

Here's the through-line I read across both, and I'll own it as my read, not their finding. When a system isn't grounded (no reliable source to reach, no one owning the truth it acts on) it fills the gap with a guess, delivered with full confidence. A tool that cannot retain context has to re-guess that context on every call. MIT names the mechanism a learning gap; Gartner counts the wreckage in cancelled projects. I call the thing underneath both the confident guess. That, not AI hype and not slow models, is the failure mode.

Left to its own devices, an agent fills every gap it hits with whatever its training data suggests, a plausible-sounding assumption, or a deep search across whatever the open web happens to serve up that day. It doesn't hesitate. It doesn't flag the seam between what it knows and what it invented. It hands you fluent, self-assured, beautifully-formatted output, and some unknowable fraction of it is made up.

I want to be honest here: we nearly walked into this ourselves. Early on it's intoxicating to watch an agent produce a polished answer in seconds. You have to keep reminding yourself that polished and true are not the same word. Everyone chasing agentic is up against the same failure mode, and most don't know it yet.

Let me make that concrete, because I've watched it cost real money. A team we work with built an entire campaign around a message they were sure would land: a positioning call made from gut and a few confident voices in a room, not from anything their customers had actually said. The creative made by agents looked sharp at first sight. The launch was clean. And it produced almost zero pipeline, because the message answered a problem the buyers didn't really have. The copy turned out to be generic. The buyers saw right through it. That one ungrounded decision cost them well into six figures: a confident guess, delivered beautifully. And note that this was a human failure, made at human speed; the agents only executed the creative.

Why is that guess so much more dangerous now than it was a year ago? That was one team, one campaign, one bill. When an agent makes the same kind of call from the same kind of assumption, that flawed logic runs across thousands of decisions, automatically, downstream, at machine speed. The agentic era doesn't just scale your good execution. It scales your bad execution just as fast, and hands you the bill quarters later, when someone in a budget review asks what the project actually returned and the room goes quiet.

That silence is what a cancellation sounds like. It's what 95% sounds like.

What a grounded agent looks like

So picture the other side of that divide.

Picture an agent that, when it hits a gap, doesn't guess. It reaches into a repository of verified customer truth: real findings, anchored to real quotes from real customers, each carrying a confidence label that says how much weight it can bear. It answers. And every claim traces back to something a customer actually said.

A product marketer hands an agent the next launch. Instead of inventing benefit statements no buyer has ever uttered, it drafts positioning built from the exact phrases churned customers used on their way out, each line traceable to the interview it came from, each carrying a confidence label that says how many customers actually said it. Or a sales development (SDR) agent personalizing outreach: not a plausible-sounding pain scraped from a job title, but the real objection three lost deals raised last quarter, quoted. Win/loss, churn interviews, ideal customer profile (ICP) refinement, messaging, battlecards (the competitive one-pagers sales teams carry), willingness-to-pay: these are the surfaces where go-to-market teams already spend, and every one of them is a place an agent either grounds in what customers said or makes something up.

That's not a fantasy feature. It's how the technology was designed to work. Retrieval-augmented generation (RAG), grounding a model in an external, trusted source instead of its own training data, was shown years ago to produce measurably more factual, grounded output, with sources you can inspect and cite. The mechanism has existed the whole time.

If you've built with RAG, you're already objecting: we wired up retrieval and it still hallucinated. You're right, and it's the most important thing to be clear about. RAG the technique is commoditized, a retrieval step anyone can bolt on. Point it at an unvalidated dump of documents and it will still lie confidently, because it's grounding in noise. The differentiator was never the retrieval. It's the corpus: whether what the agent reaches is curated, quote-anchored, confidence-labeled, and traceable to a real source, or a pile of PDFs nobody validated. Naive RAG over a messy corpus is a confident guess with a citation stapled to it. The moat is the quality and structure of what you ground in, not the act of grounding.

On that side of the divide, your fundamentals stop being academic hygiene and become the thing that lets the business move fast without making things up. Confidence labels and quote-level traceability turn into a governance layer: everyone can see how much an insight can bear and trace it back to the customer's own words. That's what makes it safe to let an agent act. For those of us building in Europe, that traceability isn't a nice-to-have either. When an agent makes a claim about a person, or acts on customer-interview data, provenance is what lets you show where a decision came from and honor an access or erasure request: the difference, under GDPR, between an agent you can account for and one you can't.

And this is the flip, the counterintuitive heart of the whole thing:

Model limitations make your fundamentals more valuable, not less.

The more decisions you hand to agents, the more leverage a single well-grounded, validated insight carries, because it now shapes hundreds of automated decisions, not one person's Tuesday. The researcher who used to "produce reports people skimmed" becomes the owner of the customer-truth layer the whole business runs on, humans and agents alike. Rigor was never the boring part. In the agentic era, rigor is the moat.

The agent matters less than what you feed it

This is the part everyone gets backwards. They spend the budget on the agent: the model, the orchestration, the demo that dazzles the board. They spend almost nothing on what the agent stands on.

The general fix is grounding: give the agent a trusted external source to reach instead of its own training data, and make someone own it. That principle is well established, and it isn't mine. The bet I'm making is narrower: that for the decisions product and go-to-market teams hand to agents, the source worth grounding in is a customer-truth layer: insights stored atomically, each one a finding plus its evidence quote, its tags, its confidence label; connected, queryable, and wired into the tools where decisions actually happen. Grounding is the principle. Customer truth is where I'm putting my chips.

Connect that layer to the applications running your agents, and every insight they reach is quote-anchored and verified against the source interview, so they reason from evidence instead of filling gaps with guesses. The agent stops being a liability and becomes a conduit for verified customer truth. MIT's buy-versus-build finding points the same way: teams that integrated a partner's system succeeded about twice as often as teams that built their own. That's the system we built at Bubble, and it's the reason I can put that absurd number in our business plan.

There's a sharper edge here for anyone who builds software. Gartner now puts up to $234 billion of enterprise-application software spend at risk from what it calls agentic arbitrage by 2030, as agents complete the work across systems and the interface stops being the differentiator; the market repriced that in real time when a February 2026 selloff erased roughly $285 billion in SaaS value in about 48 hours. Gartner's survivors are the vendors who capture and keep their customers' knowledge, and the deepest version of that knowledge (this extension is my read, not Gartner's) is what your customers actually said.

The strongest version of the claim I'm entitled to make

Here's the most honest thing I can offer, and I want to be precise about what it is and isn't: we run on this ourselves. Not proof: the €10M number is a target, not a result, and I'd be doing the exact thing this article warns against if I dressed a projection up as evidence. What it is, is conviction with skin in the game.

The €10M-with-10-people bet only works if our agents aren't guessing, so we grounded them in the same customer-truth layer we sell. When an agent makes a call about a segment, a pain, a churn risk, it's reaching a real quote with a confidence label, not improvising from training data. If I'm wrong about grounding, I don't get to watch it fail in a slide deck. I watch it fail live, in my own company. That's the strongest version of the claim I'm entitled to make: not this works, but I've bet the company that it does.

That's also why I think this is the moment, and not a year from now. Three things had to line up: a genuine board-level shift (agentic is real, not hype), a market that's been underserved (customer research built for everyone except the product marketers and product managers who need it most), and a defensible, EU-native system built specifically for them. Remove any one of those and it doesn't hold: a clever tool with no urgency, or urgency with nothing underneath. Right now all three hold at once.

The two questions to bring back to your board

If you take one thing from this, don't take a product. Take a diagnostic. Walk into your next leadership meeting and put two questions on the table:

1. Do we actually have the fundamentals: decisions rooted in verified customer truth, with the evidence and confidence to back them? Not slides. Not a wiki nobody reads. A living, validated, traceable layer of what your customers actually said.

2. Do our agents have access to them? Because a truth layer your agents can't reach is a truth layer that doesn't exist as far as the decisions are concerned.

Sit with what each "no" costs you.

If the answer to the first question is no, the agentic shift will take your weak fundamentals and scale them: bad decisions and poor execution, automated, downstream, faster than any review cycle, straight into your growth and your margins. And sitting the transition out doesn't avoid that bill; it just arrives more slowly, in your valuation and your next raise.

There's no version of this where the fundamentals don't matter. The agentic era just raised the stakes on getting them right, and shortened the time you have to do it.

Agents are already drafting the decisions your people sign off on, and their share is only growing. The question left is whether you've given them the truth to decide on.

Sources referenced: MIT NANDA, "The GenAI Divide: State of AI in Business 2025"; Gartner press release, "Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (June 2025); Gartner press release, "$234 Billion in Enterprise Application Software Spend Is at Risk from Agentic AI" (July 2026); Lewis et al., "Retrieval-Augmented Generation" (2020). Return-on-AI figures draw on S&P Global Market Intelligence's 2025 Voice of the Enterprise survey, PwC's 2026 Global CEO Survey, and Foundry's 2026 State of the CIO. Market corroboration of the February 2026 SaaS selloff draws on contemporary reporting including CIO Dive.