How would you talk a founder out of an AI feature without damaging your relationship?

Rivermint builds DeckForge, a tool that turns a startup's raw notes into a full pitch deck. Yannick Belrose, Rivermint's founder, wants to ship PitchCoach: a feature that listens live during a founder's rehearsal call and tells them, in real time, how the investor on the other end is actually reacting. Ilenia Draskovic is the PM who owns DeckForge's roadmap, and she does not think PitchCoach should ship as pitched.

How to answer this, stage by stage

Nobody is scoring whether you can recite SPARK's five letters. They're scoring whether you'd survive the actual meeting where a founder's favorite idea is on the table.

Let's learn

DeckForge is a tool that turns a startup's raw notes into a full pitch deck, so a founder isn't staring at a blank slide template at midnight before a fundraise.

Before any redirect technique, this is how it usually goes. Fast yes, or flat no. Neither one actually serves the goal underneath the pitch.

Before Ilenia built a real approach for this, Rivermint had a habit: whatever AI feature Yannick pitched in a Monday meeting got waved through in that same meeting, because saying yes fast felt like alignment. Over the past year, six of his ideas shipped that way. Half of them got quietly shelved weeks later once they didn't actually work, PitchCoach's own precursor, a "confidence score" overlay, burned three engineering weeks before investors said plainly they didn't trust a number with no reasoning behind it.

Now the same kind of conversation runs differently. When Yannick pitched PitchCoach, Ilenia didn't say yes and didn't say no in the room. She asked what result he actually wanted, ran the blind test that week, and came back with the number above plus a scoped alternative. The whole redirect took one 25 minute meeting.

Here's the turn: the extra questions Ilenia asks aren't about doubting Yannick's instincts. His instinct, that founders need to know if their pitch is landing, was right the whole time. What needed catching was the specific bet that a model could read something a microphone was never going to reliably hear.

At its worst, this costs a full sprint of two engineers plus a founder who now associates his own AI PM with "no." That's what happened with the earlier confidence-score feature: three weeks built, quietly abandoned, and a real dent in how much Yannick trusted the next thing DeckForge's team told him.

What I would leave alone: not every founder AI idea gets this much scrutiny. Yannick also wanted deck templates to auto-suggest a color palette from a company's existing brand assets, cheap to build, low risk if it's wrong, easy to quietly improve later. That one just got built the week it was pitched. The redirect is for bets where being wrong is expensive, not for every idea that comes out of a Monday meeting.

The lesson: a founder's instinct about the goal is usually right, even when their proposed feature isn't. Arguing with the feature loses the relationship. Getting curious about the goal, then bringing real evidence and a real alternative, keeps it.

Now here is the same thing as a story

The short version above is what you'd say out loud in the room. Read this one for what it actually felt like to sit across from a founder and not flinch.

Ilenia Draskovic had been Rivermint's AI PM for a year and a half, and she was good at the parts nobody sees: writing eval sets, tracking which model version shipped when, catching drift before a customer did. Founder conversations were the part she hadn't fully worked out yet.

Yannick pitched PitchCoach on a Tuesday, standing at the whiteboard, genuinely excited. "Every founder who's ever pitched wishes they knew, in the moment, whether the investor was actually buying it. We can build that." The room nodded. It was a good instinct, said with real conviction.

For a moment, Ilenia almost said yes on the spot, the way the team usually did. Instead she asked the only question that mattered. "What happens for a founder if this works exactly the way you're picturing it?"

Yannick answered without hesitating: "They stop guessing. They know, live, whether to change tack mid-pitch." That was the real goal. Nobody had said it out loud in those words before.

Four beats, in order. Skipping the second one is what turns this into an argument instead of a conversation.

She didn't argue with him in that meeting. She asked for a week to test it properly, and he agreed, mostly because nobody had ever asked him for evidence before, only opinions.

The test itself was almost boring to run: 40 old rehearsal recordings where the team already knew, from the founder's own follow-up notes, how each real investor call had actually gone. She had the sentiment model score all 40 live, blind to the outcome. Then she checked it against a baseline that did nothing but guess "lukewarm" every single time, since most real reactions land somewhere in the middle anyway.

The model matched reality 17 times. The baseline, which never listened to a single word, matched reality 21 times.

She brought the number back to Yannick before she brought anything else. Not a slide deck of objections, one printed page with the two bars on it. "Your instinct was right, founders do need to know if it's landing. This specific way of telling them isn't true yet. Here's what is." She showed him the corpus of 600 funded decks and the structural gaps it could actually, checkably catch.

He pushed back once, hard. "Investors expect this kind of thing now, I've seen competitors demo it." She didn't argue the feature. She stayed on the goal. "If a competitor's version is real, it'll show up in results we can check the same way I checked ours. Ours isn't there yet, and I don't want to hand a founder something that's wrong with total confidence."

The redirect doesn't only need to work in the calm meeting. It has to survive the one where he's annoyed.

He never had a number for what would make him trust a redirect instead of hearing "no." He had a feeling with two settings: dismissed, or heard. The blind test number didn't flip that feeling by itself. Being asked what he actually wanted, before any of it, did.

Back at the whiteboard three weeks earlier, on the old confidence-score feature, the decision that got made wasn't unreasonable at the time. Saying yes fast to a founder's idea is a defensible habit when the ideas have mostly been cheap and safe. It stopped being defensible the moment one of them needed a technical bet nobody had checked.

Same underlying goal, split into what could ship honestly now and what still needs real evidence before it ever does.

Here's the replay, same founder, same conviction, but with the question asked before the reaction. Ilenia scopes the content-scoring feature that same week. It ships in eleven days. Yannick demos it to the next batch of founders himself, unprompted, because it's a thing that's actually, checkably true.

One version of this story ends with three burned weeks and a founder who trusts his PM a little less. The other ends with a shipped feature and a founder who brought the next idea straight to her, instead of around her.

What I'd tell myself, standing at that whiteboard: the goal was never the enemy. Only the fastest, flashiest guess at how to get there was. Ask what's underneath the pitch before you have an opinion about the pitch itself.

SPARK, run for the meeting where you're about to disagree with the person who signs off on your roadmapNot a script for winning an argument. SPARK is what keeps the relationship intact when the honest answer is still no, not yet.

The actual move this answer is about. Four parts, in this order, every time.

The recap, one line per letter: situation is the greenlight-on-conviction habit written down honestly, payoff is the habit of chasing the goal instead of the pitch, anchor is the reflect-evidence-alternative sequence, risk is the trust that's actually on the table either way, and keep out draws the line at bets cheap enough not to need this much care.

And if you want to be sure it really works, try it somewhere elseSame five letters, a customer support inbox instead of a pitch deck. The judgment call this time is whether a model can hear anger in someone's voice.

Halvard Petrenko founded Northloom, a customer support platform. He wants EmotionDetect: a feature that flags a caller as "angry" in real time from their voice, so agents get routed a warning before they even pick up. Sunniva Torbin, Northloom's AI PM, ran the same kind of test Ilenia did. Mapped onto SPARK: situation is that agents currently rely on their own ears, and miss maybe one in five genuinely upset callers in a busy queue. Payoff is the same habit again, chase the goal, not the pitch. Anchor is a different, reliable feature: flag callers whose call has already been escalated twice before, a real, checkable signal, instead of guessing at emotion from tone, which drops sharply in accuracy across different accents and speech patterns the training data barely covered. Risk is that Halvard, like Yannick, could hear this as a flat no and stop bringing ideas forward. Keep out is that Northloom's simple keyword flag for the words "cancel" or "refund" stays exactly as it is, a plain rule doesn't need this conversation at all.

The same fork applies whether the pitch is about a pitch deck or a support inbox. The goal decides the branch, not how exciting the pitch sounded.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Skip straight to the anchor, say "reflect the goal, then bring evidence and an alternative," and stop.
Cost: no time to run a real blind test before the meeting. Say so honestly, ask for a week, and don't guess at a number you don't have.
The model got better, for real: say a future sentiment model actually clears the baseline on a real test. Revisit the original idea then, the redirect was never a permanent no, it was a "not on this evidence."

Where people run it wrong.
They argue the feature's merits instead of asking what goal it's actually chasing, which turns a conversation into a debate nobody wins cleanly.
They cave to conviction because conflict feels worse than a wasted sprint, then eat the wasted sprint anyway, later and bigger.
They bring a flat no with no built alternative, which reads as blocking even when the underlying concern is completely correct.

How to use it live. The moment a founder or exec pitches you a feature you don't trust, don't respond to the feature at all. Ask what result they're actually chasing. That question buys you real thinking time, and more often than not, it's where the honest path forward was hiding the whole time.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0