Mental Models for the AI EconomyProduct Strategy
Roadmapping When the Model Gets Better Every Seven Months
METR found AI task horizons doubling roughly every seven months. How to plan in capability bands, write the test before the feature, and stop building what the next model gives away.
A roadmap is a bet on what you can build and what customers will want. In AI products the first half of that bet moves under you. A feature that is impossible in January can be a weekend project by summer, and a feature you spent a quarter on can become a free checkbox in someone else's model.
There is at least one useful measurement of how fast the ground moves. In March 2025 the research group METR measured the length of software tasks frontier models could finish half the time. The best models managed tasks that take a skilled person about an hour, and that horizon had doubled roughly every seven months for six years. Nobody should treat that as a law. But it is a far better planning assumption than “the model we have now is the model we will have.”
When does your feature become possible?
Three habits that replace the fixed roadmap
Plan in capability bands, not dates. Sort ideas by what they need from the model: works today, works sometimes, needs a model that does not exist yet. The middle band is where most of the product work is: review steps, fallbacks and limits that make a mostly-working model safe to ship. I wrote about that review work in the sign-off bottleneck.
Write the test before the feature. For every “Later” idea, keep a small fixed set of real examples and the answer you need. Run it against each new model release. The week it passes, you know before your competitors do. This is the same discipline as the agent rules in Harness²: capture what good looks like, then let evidence move things.
Ask what survives the next model. If a better model would make your feature trivial for everyone, it is not a moat. Spend your scarce effort on what a model cannot give a competitor: your customers' data, workflow and trust, the argument in When Code Is Cheap, What Still Protects a SaaS?.
Now, Next, Later — or Rethink
What I tell product teams
Shorten the horizon you commit to and lengthen the one you watch. Commit to a quarter. Watch eighteen months. Keep the bets small, as in The Best Product Decision Might Be a Smaller Bet, so that when the curve jumps you can move with it instead of finishing a plan that was right a year ago. And remember who decides what is worth building. Models make more things possible; they do not tell you which ones matter. That is still the product manager's job.
Sources
Across the series
- The European Accessibility Act Is Live. Here’s What It Asks of Your Product.
- Your Visitors Already Said No. Is Your Site Listening?
- Seats, Usage or Outcomes: Pricing an AI Product Without Losing Your Margin
- Your AI Chat Box Is Probably Unusable With a Screen Reader
- Can You Train AI on Your Customers’ Data? Ask Them First
- Accessibility Testing: A Green Scan Is Not a Finished Journey
- AI Data Retention: ‘Not Used for Training’ Does Not Mean Deleted
- Agent Evaluation: The Demo Worked. Can You Ship It?
Get the next essay the day it publishes.
A few a month, written from real builds. Unsubscribe in one click.
Performance marketing and search leader who builds — 15+ years in digital growth and product (AT&T/DirecTV, Microsoft, Employbridge, Site Impact), now running his own properties with AI. These essays are field notes from that loop: what AI actually changes about work, product, and the economics underneath. Based in South Florida.
- Product StrategySeats, Usage or Outcomes: Pricing an AI Product Without Losing Your Margin
Seat pricing fell from 21% to 15% of SaaS companies in a year while hybrid jumped to 41%. Why AI breaks the seat, how to model margin before you launch, and when outcome pricing is honest.
9 min read · interactive - Product StrategyThe Best Product Decision Might Be a Smaller Bet
When the evidence is thin, conviction is not a substitute for a learning plan. A practical framework for reversible decisions, meaningful experiments, and stopping rules.
8 min read · interactive - Product StrategyWhat Do You Build When AI Can Do the Work?
AI runs the instructions. It doesn't write your why, own your state or decide where the person takes over. Those are the parts still worth building.
12 min read · interactive - Product StrategyWhen Building Gets Cheap, Product Judgment Gets Expensive
Claude Code expands the space of possible products. The differentiator is no longer a convincing specification—it is choosing, testing and shipping the right experience.
8 min read · interactive - Product StrategyAgent Evaluation: The Demo Worked. Can You Ship It?
A product leader's release contract for AI agents: representative tasks, final-state checks, critical failure gates and monitored pilots—with a reliability microscope and an interactive release desk.
9 min read · interactive - Web & AI SearchBrand Mentions Are the New Backlinks
Across 75,000 brands, web mentions correlated three times more strongly with AI visibility than backlinks. What that means, where to earn mentions, and how to track them.
10 min read · data - Web & AI SearchHow to Build a Site-Wide Entity Graph
Stable @ids, hub pages, about and mentions, sameAs, and one registry that keeps markup, links and llms.txt from drifting — a seven-step build guide.
11 min read · interactive - Web & AI SearchTopical Authority: 30 Connected Pages Beat 300
Authority is the shape of your knowledge, not the size of your archive. How to map a topic, assign one owner per question, and link hub, spoke and sideways.
9 min read - Web & AI SearchGoogle Maps Heatmaps: What Your Local Ranking Really Looks Like
Why one rank is misleading, how geo-grid footprints expose opportunity, and the truth behind the observed shared-building filter — with an interactive map model.
8 min read · interactive