How would you decide whether your first AI hire should be a PM or an engineer?

Pageturn reads a pitch deck a founder sends a VC and pulls out the numbers an associate actually checks: round size, market size, team background, traction. Ekaterina Larsdatter is building it alone, five years after leaving a seed fund as an associate herself, with a $180,000 pre-seed and enough runway for exactly one hire before she has to raise again. Leontyne Ferrara, a partner at Northbell Capital, has agreed to pilot Pageturn against forty real decks from this quarter's intake. Before that pilot, Ekaterina has to decide: her first hire, a product manager or an engineer.

How to answer this, stage by stage

Nobody is grading whether you can recite "hire product before engineering" and sound thoughtful. They're grading whether you can turn "who do I hire first" into a real test you'd actually run on your own company.

Let's learn

Pageturn reads a pitch deck a founder sends to a VC and pulls out the numbers an associate has to check before a partner meeting: how big the round is, what the market size claim rests on, who's on the team, whether the traction numbers hold up next to what the deck says about them.

Five steps. The fourth one, flagging what's shaky instead of stating it as fact, is the step that was still missing when Barrett was hired.

Before Pageturn, an associate at Northbell Capital spends about 9 minutes on a first pass through a deck, checking the round size against the ask, skimming the team slide, sanity checking the market number. Forty decks a week is a normal intake for one associate. That's 6 hours a week doing something a computer should be able to do faster.

Here's the turn. Ekaterina's first hire, a product manager named Barrett Kessering, made Pageturn look like it was moving. Fifteen interviews with VC associates. A roadmap. Two clean personas. None of that means anything if nothing underneath it can actually read a deck yet.

Same desk, same eight weeks of salary. Only one of these two produced something Northbell could actually pilot.

What it costs at its worst: on demo day, live against 40 real decks, Pageturn produced a usable score for 16 of them. The other 24 either errored out on an odd layout or came back wrong, including one scanned funding table that read a $40 million Series A ask as $40 thousand. Leontyne's team paused the pilot for two months. Barrett's eight weeks had cost about $20,000 in salary, and there was still no working prototype underneath any of his roadmap.

What I would leave alone: her decision to build Pageturn on top of an existing model instead of training her own extraction model from scratch. That call was right regardless of which role she hired first. Training a custom model would have taken longer than either hire question was ever going to matter for.

The lesson: "hire product before engineering" is a rule, applied the same way no matter what's actually true about your company. The real question was never a rule to follow. It was one fact to check before deciding anything.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one for the two months it took Northbell to agree to a second try.

Ekaterina Larsdatter spent five years as an associate at a seed fund before she left to build Pageturn. She could tell a real market number from a padded one in about the time it took to flip a slide. She wasn't a trained engineer, but she could write a scrappy Python script well enough to test an idea on a weekend, which is exactly what she did before hiring anyone at all.

Then an accelerator mentor told her, in the same words he used with every founder in the cohort, that you always hire product before engineering. Ekaterina had heard the same line from two other people that quarter. It sounded like the responsible thing to do. She hired Barrett Kessering.

For six weeks, that felt like real progress. Barrett interviewed fifteen VC associates about how they actually screen decks. He wrote two sharp personas, "the first pass associate" and "the diligence partner." He built a roadmap with a launch date on it. Ekaterina kept coding on nights and weekends, pulling numbers out of pitch decks with a script wired to an off the shelf model. On her three cleanest sample decks, text native, well formatted, it worked every time.

Four beats. The third one, specs piling up while nothing ran, was the one nobody was watching.

The other seven sample decks weren't clean. Scanned pages. Tables rendered as pictures instead of real text. A financial slide with the numbers laid out in a shape her script had never seen before. It broke on those, over and over, and she kept telling herself she'd get to it once Barrett's roadmap settled.

Then Leontyne Ferrara, a partner at Northbell Capital, agreed to pilot Pageturn against forty real decks from that quarter's intake. She wanted a live demo in three weeks. Ekaterina said yes before she'd fully counted how far her weekend script actually was from ready.

Demo day, live, in front of Leontyne's team. Of the forty real decks, Pageturn scored sixteen usably. It timed out on six with layouts it had never handled. On the rest, it produced numbers that looked confident and weren't. One scanned funding table turned a $40 million Series A ask into $40 thousand, stated as plainly as if it had read it off a clean spreadsheet. Leontyne's team went quiet, thanked them, and paused the pilot.

The real cost wasn't the twenty four wrong decks. It was that Ekaterina had spent her one hire slot and eight weeks and $20,000, and she still had nothing a product manager could actually judge. Barrett had never had anything real to define quality against. He'd been building a beautiful house with no foundation poured under it, because nobody had checked whether a foundation existed before he started framing the walls.

The decision Ekaterina would take back sits in a fifteen minute conversation with her accelerator mentor, months before any of this. He told her, the way he told every founder, to hire product before engineering. She didn't ask him the one question that mattered: could she, right now, already build a working prototype herself. She couldn't. Her own 42 hours on three clean decks was the honest answer sitting in front of her the whole time.

Run it again. Ekaterina hires an engineer first instead, a contractor who's built extraction pipelines before. Seven weeks, full time, building the parsing, the layout handling, the fallback for scanned tables, the scoring logic. He reaches 34 of 40 real Northbell decks scored usably, 85 percent, including scanned and image heavy ones her weekend script never touched. The cost: about $20,200, almost exactly what Barrett's eight weeks had cost. Only now there's a real prototype under it. Ekaterina brings on a PM the day it exists, to define what "good enough" means for a VC and decide when to show it. The second demo with Leontyne, six weeks later, goes well enough that Northbell keeps using it on their live deal flow.

Same money both times. One hire produced a roadmap sitting on nothing. The other produced code a real VC could actually use.

What I'd tell myself, sitting in that fifteen minute conversation with the mentor: general advice is a dial, turned the same way for every company that walks through the door. The question I actually needed to answer was never "what does the standard advice say." It was one fact about my own company, and I already had the evidence sitting in my own weekend folder.

BOUND, for turning "PM or engineer" from a hunch into an equation

Not a way to sound thoughtful about hiring order. BOUND is what forces "who do I hire first" into real weeks, real dollars, and a test you can actually run, instead of repeating advice that sounds right at every company and fits none of them exactly.

Ekaterina's real number landed near the low end of the range, once the right role was actually in the seat.

The test in one picture. Whether a judge exists only matters once something exists to judge.

Three things worth naming directly, since this is where the real judgment sits. The alternative Ekaterina's advisor floated and she turned down was hiring one generalist who could do a bit of both PM and engineering work. It lost because she interviewed for it and found nobody senior enough at either skill, within her runway and budget, to do both well, and a person who's mediocre at both would have been slower than a strong engineer alone. The AI specific failure mode here is silent wrongness: the extraction pipeline doesn't error when it misreads a scanned table, it just states a confident number that happens to be wrong, the $40 million read as $40 thousand. The guardrail is the confidence threshold from the knowledge spark above, paired with a golden set of real decks checked before every version ships. And the trade off is real and accepted on purpose: building quietly for 7 weeks before showing Northbell anything again meant giving up the look of fast, visible progress that a roadmap and fifteen interviews can produce in a single board update, in exchange for something actually trustworthy the second time Leontyne's team looked at it.

And if you want to be sure it really works, try it somewhere else

Same five letters, a fabric mill instead of a VC firm, and this time the swing fact points the other way.

Weftguard is Gorrick Fentanu's AI tool for garment manufacturers: photograph a roll of fabric before it's cut, and it flags snags, dye streaks, and weave flaws before a bad roll turns into a bad order. Bellfort Mills has offered him a pilot: 200 real fabric roll photos from their current run, results back in two weeks. Gorrick is asking Ekaterina's exact question, on a much smaller floor.

Different floor, same test. The answer flips because the fact underneath it is different.

Run BOUND on it. Break it down, same two halves: someone to build a defect flagging prototype against real photos, someone to decide what catch rate is actually good enough for Bellfort's tolerance. Own the numbers: Gorrick already has a part time contract engineer, a former colleague, who's built a rough prototype in his spare hours. On Bellfort's first 50 sample photos, it already catches about 70 percent of real defects. Use the range: for Gorrick, the honest range isn't 6 to 10 weeks of engineering. It's 2 to 3 weeks of product judgment, because the prototype exists, and nobody has said whether 70 percent is good enough, or what false alarm rate a premium garment line versus a cheaper one can tolerate.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: hire the engineer first unless someone inside the company can already build a working prototype, that one fact decides it, not the round size or the pressure to show progress.
Cost: no budget for a second hire this quarter. Have the founder or an existing technical contractor build the prototype part time first, and only bring on a dedicated hire, either role, once there's something real to hire against.
The model got better, for real: say a newer, cheaper extraction model cuts Pageturn's build cost in half. The hire order barely moves, because someone still has to decide what "good enough" means for this specific product, and a better model doesn't answer that question on its own.

Where people run it wrong.
They apply "hire product before engineering" as a rule, without checking whether anything exists yet for the product hire to define.
They let a PM's early visible output, roadmaps, interview decks, personas, stand in for real progress on a product that still can't do the one thing it's for.
They size the prototype phase off a guess instead of a real number, and never check that number against what it actually took someone to build.

How to use it live. Before naming a hire, ask yourself one plain question out loud: "can anyone here already build a first working version of this, today, not eventually." Whichever answer comes back honestly is usually the exact hire order an interviewer is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0