Take something on your roadmap for next quarter. Now ship it by Friday.

If that sentence made you flinch, keep reading.

The flinch is the old speed talking, and the old speed is gone.

Most companies overpay for AI and underuse it at the same time:

Routine work routed to the most expensive model, the whole effort pointed at whatever trended last week, and then surprise that the bill grew and the lead didn’t.

Nobody knows where this technology ends up. You don’t need to.

These moves pay off either way.

#1 - Start with the “Friday Test”

On a Monday, take one task you had penciled in for next quarter. Ask what it would take to ship it by Friday. Then ask what it would take to ship it today?

The distance between the timeline you assumed and the one available is the AI opportunity, measured in one exercise.

Run it on a real task and list what blocks the faster date.

• A person
• An approval
• A missing data connection

The blocker on that list is rarely the technology anymore.

The drafting, the code, the analysis, the parts that used to take weeks, a model now handles in an afternoon.

What remains is process: meetings, approvals, handoffs.

The list you just wrote is your roadmap’s real schedule, and every item you clear converts straight into shipping speed. Five minutes, and you know what to fix first.

#2 - Use workhorse AI for most work. Pay frontier costs when the decision matters

“Too cheap to meter” was the promise about nuclear power in the 1950s. For a while a lot of people assumed AI would go the same way, that intelligence would get so cheap it would basically be free.

Until recently, this felt like a safe bet.

The cost of any given capability keeps dropping fast. But the frontier keeps moving, and frontier-grade reasoning stays expensive, because the best models think harder, and thinking harder burns real money in GPU time.

Yesterday’s magic gets cheap, but today’s magic stays pricey. It’s entirely possible the best intelligence gets more expensive for whoever can afford it, at least for a while.

The good news is you often don’t need the smartest model anyway. Open models and the tier right below the frontier have gotten startlingly good.

As of mid-2026, that means Claude Sonnet instead of Opus or Fable 5, GPT-5 Mini instead of GPT-5, Gemini Flash instead of Pro.

Most of what your company asks an AI to do, drafts, summaries, research, routine analysis, runs fine on models that cost a fraction of the best one.

The smartest model earns its bill only on decisions where being wrong is expensive: a pricing change, a term sheet, a key hire.

The price of old intelligence falls. The price of the best intelligence doesn’t.

Sort the work once, then revisit monthly, because the tiers keep moving.

The teams that win on cost match the model to the job instead of sending everything to the top shelf.

#3 - Build so the next model upgrade is free

Every few months a better model ships. That’s the most reliable fact in this field, so build to cash in on it instead of getting broken by it.

In practice:

• Connect your data first: when an AI system disappoints, the cause is usually that it can’t see your information or act in your systems, not that it isn’t smart.
  
• Keep the structure thin: chains of forty or fifty AI agents passing work down a line demo beautifully, then break the day a new model doesn’t fit.
  
• And make the model a setting you can change, not a foundation you built on. A simple check: if switching models would take your team more than a day, the structure is too thick.

We learned the data lesson in our own shop.

We run an agent named Cove inside Cape Fear. She lives in our Slack, reads our deal pipeline in Notion, and watches the inbox, so nothing slips through the cracks of a small team moving fast.

Working alongside her taught us the constraint is access, not intelligence: what she can see, and what she’s allowed to do.

#4 - Point AI at the lead you already have

Your company does something competitors can’t or won’t.

Usually it’s one of two things: data nobody else has, or customer relationships that took years to earn.

That’s where AI pays first. It multiplies whatever you feed it, and your advantage is the one input competitors can’t copy.

Ask which part of that edge gets faster, cheaper, or wider with AI.

Aim there and ignore everything else until it’s working.

One of our portfolio companies spent years becoming the system of record for an industry most software companies overlook.

Its data on how that industry actually runs, day to day, order by order, exists nowhere else. Point AI at that and it produces answers no rival can match, however smart their model is.

The model is rentable. The data isn’t.

The companies getting the most from AI are widening a lead they already had. The ones chasing whatever is trending this week are funding someone else’s learning curve.

#5 - Ask whether anyone wanted it before AI made it easy to build

The last two years are littered with AI features nobody asked for, such as chatbots bolted onto products that needed better search, or copilots for tasks no customer ever complained about.

They shipped because they got cheap to build, but cheap to build is not a reason.

AI makes a good business faster and a bad business fail faster.

No technology has ever made a market want something it doesn’t.

So run the kill question before the kickoff: did customers want this before it became cheap to build?

The evidence is usually lying around already, in support tickets asking for it, workarounds customers cobbled together, the request that keeps surfacing on sales calls.

If none of that exists, the demand is theoretical, and the project should be too. The cheapest AI project is the one you kill in a meeting.

Speed used to be the differentiator. Now everyone has it, and direction is the only thing left to get right.

The accelerant doesn’t care which way you’re pointed. It will take you to product-market fit or to the wall, twice as fast, and it doesn’t know the difference.

No technology, however novel, can make a market want something it doesn’t.

Build something people want. Then pour it on.
_____

We spend our days backing founders who aim this kind of speed at an advantage they already own:

Software becoming the operating layer of industries that still run on spreadsheets and legacy systems; AI pointed at data nobody else has; national security-critical defense tech that gets fielded, and the infrastructure of the orbital economy.

If that's you, hit reply and tell us what you're building. We read everything, and the patterns across founders are half of how this thesis gets sharper.

richard@capefear.vc