Analyze a vertical AI product and identify its actual defensibility.

Driftgauge lets a farmer photograph a leaf and get back a likely disease and a treatment window. Osei Vantongeren manages a farm co-op of 60 growers and has watched two copycat apps launch and stall in the last year alone.

How to answer this, stage by stage

Six moves, and the third names the thing most people get wrong on this question.

Let's learn

Driftgauge takes a photo of a crop leaf and returns a likely disease along with a treatment window, calibrated to the region's usual disease timing.

Before Driftgauge, growers in Osei's co-op relied on a county extension agent's occasional visits, roughly once every three weeks, to catch a disease early enough to treat cheaply.

With Driftgauge, growers photograph a leaf the same day they notice something off, and get a likely answer within seconds instead of waiting up to three weeks for a visit.

The turn: the real question was never whether Driftgauge's model was good. It was whether the thing making it good could be copied by anyone with a checkbook.

At its worst: Osei's team spends the next roadmap cycle polishing the app's interface and tuning the model, exactly the two things a well-funded rival can match within a single growing season, while the actual moat quietly goes unprotected.

What I would leave alone: continued investment in the app's UI and general usability. It won't ever be the moat, but a clunky app still loses growers to a nicer-looking rival, so it still deserves steady, unglamorous attention.

The lesson: a vertical AI product's real defensibility usually lives in the boring, slow-to-build asset sitting behind the flashy model, not in the model itself.

Now here is the same thing as a story

The short version above is what you'd say in the room. Read this one for how close the second copycat actually got.

Osei has managed this co-op for seven years and knows which of his 60 growers will call at the first sign of trouble versus who waits until it's serious.

The third step, local retraining on regional photos, is the one a rival genuinely can't shortcut.

The first copycat app launched eight months ago, built on a nearly identical licensed base model. It matched Driftgauge's general accuracy within weeks, since the underlying technology was never scarce to begin with.

Only one of these four branches is actually slow and expensive for a rival to walk down.

The trigger was small: a second copycat, this one better-funded, offered growers in a neighboring co-op a steep discount to switch. Two of Osei's growers asked about it directly.

By month six, the rival had matched the model but was still visibly behind on anything requiring years of regional history.

Osei checked the rival's app on a disease pattern specific to his region, a blight that shows up in a distinctive two-week window tied to local humidity. The rival's app, trained on generic data, missed the timing entirely and suggested a treatment window three weeks too late.

The model and the UI cluster in the easy-to-copy, low-protection corner. Only the photo archive sits in the corner that actually matters.

With that gap named plainly to the co-op's growers, both of the two who'd asked about switching stayed, since the rival's cheaper price didn't survive contact with one real, regional disease case.

Three of these four take years to build. The fourth, the base model, could be swapped out entirely without anyone noticing.

The old roadmap ranked "improve the model" first every quarter, because it was the most visible, most demoable piece of work. The new one ranks "grow and protect the regional data pipeline" first, since it's the piece nobody else can shortcut.

I let engineering time default to model improvements for two years because that's what every roadmap review meeting asked about first. It took watching a copycat match our accuracy in weeks, while still missing a disease timing pattern only five years of local data could teach it, to see the model was never what growers were actually staying for.

ORDER, in one screenNot a list of buzzwords. ORDER is what forces you to rank by what a rival actually can't buy.

Five letters, and reversibility is the one that actually separates a real moat from a marketing slide.

The recap, one line per letter: outcome is protecting margin and retention, reversibility is the photo archive at sixty months versus the model at two, dependency is contracts keeping farmers contributing photos, evidence is the rival's real six-month timing miss, and rank puts data and contracts ahead of the model.

And if you want to be sure it really works, try it somewhere elseSame five letters, a textile mill instead of a farm co-op.

Fennscape runs an AI fabric-defect detection system for a regional textile mill network. Ines Vartholomew leads quality assurance across four mills using it.

Mapped onto ORDER: outcome is protecting defect-catch rates and buyer trust against a rival mill network adopting similar AI. Reversibility ranks a licensed base vision model as fast to copy, roughly two months, against three years of mill-specific defect photos tagged with which ones actually caused a returned shipment, which a rival can't rebuild quickly. Dependency: the tagged defect archive depends on keeping the same inspection staff long enough to label returns accurately, which depends on staff retention more than any AI feature. Evidence: a rival mill network licensed a similar vision model and matched raw defect-detection accuracy within ten weeks, but still missed a mill-specific defect pattern tied to one particular loom's wear signature. Rank: one, the tagged, mill-specific defect archive. Two, inspector retention and labeling consistency. Three, the vision model itself.

Swap crop photos for loom-specific defect photos, and the exact same ranking holds for a textile mill.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "the moat is five years of tagged regional data and distribution contracts, not the model, rank protecting the data pipeline first," and stop.
Cost: there's no budget this quarter to expand the photo-tagging pipeline. Say so honestly, and start by simply not letting any current tagging effort lapse, since even holding the existing archive steady beats losing ground to a rival actively building one.
The model gets better, for real: if Driftgauge's detection model genuinely gets more accurate next release, that's still not the moat, a better model available to everyone raises the floor for every competitor equally.

Where people run it wrong.
They treat "we have an AI model" as the defensibility, when the model is usually the least scarce part of the whole product.
They rank roadmap work by what's most demoable in a meeting, not by what's genuinely hardest for a rival to copy.
They panic and discount prices the moment any copycat appears, instead of checking whether the copycat can actually reach the parts of the product that took years to build.

How to use it live. When someone asks about a vertical AI product's defensibility, ask yourself: if a well-funded rival launched tomorrow with a similar model, what's the one thing they still couldn't touch for years. That's the real answer, not whatever's on the pitch deck's technology slide.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0