Founders don't pick the wrong customer on purpose, or even in one clean mistake. They do it slowly, over months, while the deck gets prettier and the pipeline stays empty.
One founder I worked with spent six months building for customers that didn't exist. Every demo got "interesting!" and "oh cool!" The only thing those customers had in common was being polite. The deck kept improving. The pipeline stayed at zero.
Pick the wrong buyer and you're burning cash on a product nobody will pay for.
The slow-motion failure pattern
You pick a market because it's big. You talk to a few friendly people. You hear "interesting" and decide that means traction. You build, you ship, and then you find out the buyer has no budget and no reason to switch. So you "expand the ICP," which is startup-speak for panic.
I call the warning signs death flailing. It's that phase where a company starts chasing vanity metrics and magical saviors. The sales guru. The former-FAANG exec. The PPC god. Nobody can say who the ideal customer is without using the words "everyone" and "ubiquitous."
And the whole time, the product stays the same.
I built a document management system for doctor's offices once, back when the head of sales had print-product experience and somehow that seemed like a natural fit for selling software. Our biggest client was an oncologist. He loved that Lamborghini, or maybe it was a Ferrari — it's been a while and honestly it doesn't matter which, that's not the part that stuck with me. The part that stuck with me is that every single time the services bill came due, he never had the money. Great car. No check. We kept extending him, kept doing "one more month," because on paper he looked like exactly who we wanted: an established practice, real patients, real budget. Except the budget was theoretical right up until we needed it to be real.
I learned to always get goal-based payments with a substantial amount up front before starting any contract. Not because people are dishonest. Because "looks like a great customer" and "is a great customer" are two completely different measurements, and only one of them is about money actually showing up on time.
Pick a fight you can actually win (sounds obvious, rarely happens)
Geoffrey Moore's "beachhead" idea from Crossing the Chasm sounds obvious ...until you try to do it. Pick a narrow segment you can actually reach, where you can credibly win, and where the pain is sharp enough that people will move. Ezpz.
Founders avoid this because a tight segment makes their addressable market look tiny. Most investors want a story they can believe. Not a TAM slide. A believable story means you've had the same conversation 42 times and heard the same pain in the same words. A tight beachhead gives you the same conversation 42 times. That's the evidence investors actually want to see.
If the segment is too broad, every interview becomes a new universe with new rules. You can't compare anything. You can only guess.
Where AI helps (and where it lies to you)
AI generates options fast. It also hallucinates with confidence unless you box it in.
I've used smaller, specialized models that outperformed larger general-purpose ones in real product work. Not because they were cosmically smarter. They were trained closer to the problem and didn't wander off into awful "helpfulness". Same idea here: give AI narrow tasks with explicit outputs.
Scenario generation
Have the model generate 20–50 target-customer scenarios. Scenarios like "day in the life before", "day in the life after", what triggers purchase intent, what have they tried already that didn't work.
Workflows and though processes will get your creative juices flowing.
Assumption mapping (Lean Startup style)
Map leap-of-faith assumptions like you mean it:
- Value hypothesis: what changes for them?
- Growth hypothesis: how do you reach the next 10 buyers?
- Budget owner: who can actually sign?
- Switching cost: what breaks if they adopt you?
AI can draft these fast. You still have to pick the ones that are real enough to test. And you'll definitely need to do some refining but again, this should get you going.
That oncologist would have sailed through most of these on paper. Budget owner: obviously him, he's the doctor. Value hypothesis: obviously strong, the practice needed the system. What none of that captures is whether the money moves when it's supposed to move. You don't map that from a distance. You test it by sending an invoice and watching what happens.
Cluster analysis
If you have interview notes, cluster by:
- urgency language ("need this now" vs "cool idea")
- workflow friction
- regulatory constraints
- toolchain compatibility
- reachability (channels you can actually use)
Don't let the model cluster on vibes. Give it structured fields.
Where boring software beats AI
A chatbot will hallucinate your research notes six weeks from now. Use a spreadsheet or a real CRM.
Use boring tools:
- A spreadsheet or CRM for who you contacted, who replied, and what segment they're in
- A scoring sheet with explicit weights
- Market sizing math with sources and reachable account counts
- Forms, scheduling links, transcripts, and raw notes stored somewhere you can audit
If you can't show me the raw notes, you're just telling a story.
Here's a scorecard I've used. Boring on purpose:
Beachhead Score (1-5 each, weighted)
Urgency (x3): Are they actively trying to solve it?
Reachability (x3): Can we get in front of them reliably?
Authority (x2): Can our contact buy or drive a buy?
Budget (x2): Is there money already allocated?
Credibility (x2): Do we have a believable right to win?
Switching cost (x1): How painful is adoption?
Competition (x1): Are we up against an entrenched default?
Argue about the weights. That argument is the actual work.
The part you can't outsource: talking to humans
I worked for a startup where the CEO had a full-on cult of personality. People would do things they knew were wrong to stay aligned with The Vision. I'm allergic to that. I ask too many questions and I don't clap on command.
The same thing happens in customer interviews. You want to hear "yes." You want to feel momentum. So you nudge the witness. You ask leading questions like:
"Wouldn't it be helpful if...?"
That's pitching.
You need questions that force concrete answers:
- "What did you do the last time this happened?"
- "What did it cost you?"
- "Who else got involved?"
- "What did you try before, and why did you stop?"
- "If I disappeared tomorrow, what would you do instead?"
Stop talking. Let them fill in the gaps.
The worksheet and prompt pack
If you want a lightweight system, make a one-page worksheet and a prompt pack that feeds it.
- Worksheet: segments, assumptions, scorecard, interview log
- Prompt pack: scenario templates, assumption templates, interview question generator, clustering instructions
If you want to be extra disciplined, keep the prompts in a repo and version them. Treat them like code. Because they are.
Pick a beachhead you can reach. Prove the pain is real. Then commit long enough to get signal instead of exploring forever.
-Sethers
