Six months. That’s the new half-life of any AI strategy. If you’re still running a playbook from early 2025, you’re already a fossil.
I’ve seen this firsthand. After two exits, one to Gusto and another to Gfycat, and looking at the pitches from the 200+ AI companies I’ve put my money into—including some you definitely know, like Anthropic, OpenAI, and Scale AI—I see the same pattern. The winners are moving at a speed that feels uncomfortable. The losers are holding meetings.
Everyone is building AI features. Very few are building the right ones. The noise is deafening. You have customer requests, competitor movements, and that shiny new model that just dropped last week. How do you decide what to build next? Most teams are terrible at this. They use gut feel, or worse, a committee. Both are recipes for disaster.
I want to give you the unfiltered story of how we recently shipped an AI feature at one of my companies. No polish. No corporate PR spin. Just the messy reality of going from a raw idea to a live product in today's market. This is how we navigate the chaos.
The Idea That Almost Didn't Happen
It started, as most good ideas do, with an angry customer. Not just unhappy—genuinely angry. He was the CEO of a 300-person remote company and was about to churn.
His problem was simple: his best people were leaving. He had no idea why until it was too late. "You sell us a platform for remote teams," he said, "but you can't tell me the most important thing—who is about to quit?"
He was right. We had all this data on team interactions, project velocity, and communication patterns, but we weren’t using it to solve the biggest pain point for remote leaders: retention.
The idea was born: an AI co-pilot that predicts employee churn risk. It would analyze anonymized data patterns and flag at-risk employees before they even thought about updating their LinkedIn profile. It was ambitious. And it went straight to the bottom of our priority list.
Why? Because we were already committed to three other major features. We had a roadmap. We had quarterly goals. The plan was the plan. Sticking to the plan is how you kill a company.
The Prioritization Battle: Frameworks are Mostly Useless
We have a "framework" for prioritization. Everyone does. RICE, ICE, MoSCoW—pick your acronym. They all provide a comforting illusion of objectivity. And they all fall apart when confronted with reality.
Our churn-prediction feature scored poorly. The "Reach" was hard to define, "Impact" was a wild guess, "Confidence" was low because it was R&D-heavy, and "Effort" was massive. According to the spreadsheet, we shouldn't have even considered it for another year.
But I have one rule that overrides any framework: solve the most painful problem you can find.
A high-level executive leaving costs a company anywhere from 1.5x to 2x their annual salary. For our angry customer, that was a seven-figure problem. Our other roadmap features were nice-to-haves. This was a need-to-have.
I made the call. We were going to build it. I pulled two of our best engineers off their current projects. The product manager was not happy. The engineers were terrified. It was the right decision.
This is where being a founder is different from being a manager. A manager's job is to minimize risk and stick to the budget. A founder's job is to take the right risks to create value. Sometimes that means blowing up the plan.
The Build: Blood, Sweat, and GPUs
The first two weeks were a disaster. Our initial models were no better than a coin flip. The data was noisy, and the signals were weak. We burned through GPU credits like they were going out of style.
The team was convinced it was impossible. They showed me charts, p-values, and ROC curves that all pointed to one conclusion: this can't be done with any degree of accuracy.
This is the moment where most projects die. The data says no. The team says no. The budget says no.
But I’ve learned that in the world of AI, the initial results are almost always disappointing. The magic isn't in the first model you build; it's in the iteration. It’s about finding those non-obvious data sources and features that the model can latch onto.
We started looking at things differently. Not just project completion rates, but the timing of commits. Not just the number of messages sent, but the sentiment and response times outside of normal working hours. We started feeding the model metadata from calendar events—how many meetings were being rescheduled? How many 1:1s were getting canceled?
Slowly, the accuracy started to creep up. From 50% to 60%. Then 70%. The breakthrough came when we connected to HR data and started correlating our signals with historical promotion and compensation cycles. We hit 85% accuracy on our test set.
It took us six weeks of relentless, painful iteration. That’s an eternity in the AI world. But we had a working model. It wasn’t perfect, but it was something real.
Launch, Learn, and Live to Fight Another Day
We didn't do a big, splashy launch. We rolled it out to exactly one customer: the one who was about to churn.
We told him it was an experimental feature. We told him it might be wrong. We asked him to partner with us.
Two weeks later, he called me. The tool had flagged three employees as high-risk. One was a senior engineer he considered a rockstar. He was shocked. He sat down with all three of them. He didn't mention the tool. He just had a real conversation about their career goals and frustrations.
All three were actively interviewing for other jobs. He managed to save two of them.
He’s now our biggest advocate. The feature is live for all our enterprise customers, and it’s the single biggest driver of new sales. That one painful decision to derail our roadmap ended up defining the future of our company.
What This Means For You: Trends for the Next 18 Months
This story isn't unique. It’s a template for how to operate now.
Pain-Driven Prioritization: Forget your spreadsheets. Find the most expensive, painful, urgent problem your customers have and solve that. If your AI feature isn't saving someone a million dollars or solving a problem that keeps them up at night, you’re building a toy.
Embrace the Messy Middle: The journey from a cool idea to a working AI model is ugly. It’s full of dead ends and disappointing results. Your job as a leader is to push the team through that messy middle. If you give up after the first failed experiment, you’ll never build anything meaningful.
Data is Everything: The model is not the moat. The data is. Specifically, the proprietary, hard-to-get data that no one else has. We won because we combined product interaction data with HR data. The big platform models from Google or OpenAI can’t compete with that. Find your unique data advantage.
Launch to One: Forget beta lists with 10,000 people. Find one perfect user who has the problem you’re solving and build the product just for them. Their feedback is worth more than a thousand survey responses. If you can make them a hero, you have a real business.
We are in a new era of product development. The old rules of multi-year roadmaps and cautious, committee-driven decisions are a death sentence. You have to be faster, more opinionated, and more willing to follow the pain. The AI race won't be won by the teams with the biggest models; it will be won by the teams that can ship solutions to real-world problems the fastest. Now, go break your roadmap.
Frequently Asked Questions
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.
Do all experts agree with this view?
No, and that's fine. The best ideas in business are often contrarian. I share my perspective based on my experience and data, but I encourage you to seek out opposing viewpoints and form your own conclusions.