ConceptAdvancedAI Opportunity & Model Strategy / Data strategy as product strategy / #17
What does a data moat look like in an era of general-purpose models?
SPARKthe week a copycat launched with the same model underneath
Kelpstone Systems sells feeding and oxygenation recommendations to salmon farms, built on live sensor data and a general-purpose model. Marit Haugen runs product there. Anssi Virtanen manages the tanks at Northwane Aquafarms, one of Kelpstone's oldest customers, and had never needed to ask what was actually behind the numbers on his tablet, until a competitor launched offering the exact same thing for half the price.
The direct answer
Don't build the moat as the model wrapper, since any competitor with API access to the same general model can copy that in weeks. Build it as the outcome-linked archive: every recommendation logged against what actually happened, weeks or months later, across every farm and season you've ever served. Show that track record on every new recommendation as a confidence band, so a customer comparing you to a copycat can see the difference instead of taking your word for it.
Do this, in order
Log every recommendation against its real outcome, every time, without exception.Why: this is the one thing a general-purpose model wrapper cannot fake or shortcut, since it only comes from actually having been in the water long enough.
Show the track record on the recommendation itself, as a confidence band.Why: a moat nobody can see isn't a selling point, it's just an internal asset competing on faith against a cheaper claim.
Widen the confidence band honestly for thin history, instead of hiding the gap.Why: a bad early recommendation dressed up as confident does more damage to trust than a hedged one that's honest about what it doesn't know yet.
Stay narrow: don't chase every generic feature a copycat ships fast.Why: matching a wrapper competitor feature for feature spends your real advantage chasing an area where you have none.
Measure the moat itself, not just the product.Why: if the gap between your track record and a newcomer's isn't actually widening season over season, it isn't a moat, it's a head start that's shrinking.
How to answer this, stage by stage
Nobody is testing whether you can describe a general-purpose model. They're testing whether you can name the one thing a competitor with the same model still can't copy next quarter.
Stage 1
Scope it to one real product
Say it like this
"Let's ground this in Kelpstone Systems, and the week a copycat competitor launched wrapping the same general model around similar sensor data."
Why this works
Keeps the answer from becoming an abstract essay about moats in general.
Stage 2
Say the structure out loud
Say it like this
"I'll run this as SPARK. Situation, how the job gets done without you. Payoff, the habit you want to build. Anchor, the one design decision. Risk, what breaks the first time you're copied. Keep out, what you won't build yet."
Why this works
Signals a design method, not just a business-strategy opinion.
Stage 3
Reframe the question
Say it like this
"This isn't really 'what's your AI model.' Any competitor can rent the same model tomorrow. It's 'what do you have that took years to build and can't be rented at all.'"
Why this works
This is where a strong answer separates from someone who describes their tech stack instead of their defensibility.
Stage 4
Give the anchor
Say it like this
"The anchor is the outcome-linked archive. Every recommendation Kelpstone has ever made is logged against the real harvest result, weeks later. That's what a wrapper can't fake, no matter how good the underlying model is."
Why this works
This is the direct answer, made concrete as a single inspectable design decision.
Stage 5
Prove it with the compressed failure
Say it like this
"BrightCurrent AI launched the same quarter, wrapping the same foundation model around similar sensors, at half our price. If our value were just the wrapper, a prospect could switch with nothing lost. Instead, our recommendation shows four seasons of Northwane's own outcomes behind it, a three percent range. Theirs shows zero seasons, an eighteen percent range. The gap was visible, not a claim."
Why this works
Compresses the whole competitive threat into the one number a prospect could actually compare.
Stage 6
Close on the one line
Say it like this
"A data moat in the age of general models isn't the model at all. It's the years of paired recommendations and real outcomes that nobody can rent, buy, or copy in a quarter, no matter how good their prompt is."
Why this works
Restates the direct answer as a single line built to survive a follow-up question.
Let's learn
Here is what happens when a company built on top of a general-purpose model finds out, the hard way, what it's actually selling.
Before Kelpstone, a farm's feed and oxygenation decisions came from gut feel, a generic equipment manual, and a paper logbook, with the real result, harvest weight and mortality, only known months later, remembered rather than measured. Kelpstone's first version wrapped a general model around live sensor readings and returned a feed adjustment number, cutting the feed-conversion ratio at Northwane Aquafarms from 1.35 in the first season to 1.28 by the second, an early, genuine win.
A day with no shortage of effort, and no way to compare this season's guess against last season's actual result.
Here's the turn: the recommendation itself was never the hard part to copy. Any team with access to the same general-purpose model and similar sensor feeds could produce a similar-looking number within weeks. What Kelpstone actually had, four seasons of every past recommendation logged against its real outcome, took years to build and could not be rented, bought, or prompted into existence overnight.
Knowledge spark: why can't a competitor just copy the wrapper and catch up fast?
Wrapping a general-purpose model around live data is genuinely quick to build. What it can't shortcut is time. An outcome-linked track record only accumulates one season at a time, on real farms, with real fish. There's no way to compress four years of harvest data into four weeks of engineering.
At its worst, a company mistakes its model wrapper for its moat, invests in matching every feature a copycat ships fast, and quietly lets its real advantage, the years of outcome data, sit unused and unshown, while a cheaper competitor claims "the same AI" and nobody can immediately prove otherwise.
Feed-conversion ratio at Northwane Aquafarms, by season with Kelpstone
The model wrapper alone doesn't produce this curve. The curve comes from three extra seasons of outcomes a copycat simply hasn't lived through yet.
The choice I would take back
In the first season, Kelpstone didn't show any confidence measure on its recommendations, just a single clean number. That felt right when there was only one season of data behind everything anyway, so a confidence band would have looked identical across every farm. It stopped making sense once some farms had years of history and others had none, because the same clean number now hid a real difference customers deserved to see.
What I would leave alone: the underlying general-purpose model itself doesn't need to be replaced or fine-tuned into something proprietary. It's a fine engine. The moat was never going to live inside the model; trying to build a custom one from scratch would spend years chasing something the outcome archive already solves better.
The lesson: if the whole product could be rebuilt by a stranger with an API key in a month, the product isn't the moat. Whatever took years and can't be rented is.
Now here is the same thing as a story
The short version above is what you'd say defending the product roadmap to a nervous board. Read this one for how a copycat's launch made the invisible thing visible.
For eight months, the best part of Anssi's morning at Northwane Aquafarms was a single number on a wall-mounted tablet by the feed tanks: a recommended adjustment, updated hourly from live oxygen and temperature readings.
Four things on one screen. Only one of them, the confidence band, is the part a competitor genuinely can't fake.
Anssi trusted the number early, cautiously, the way anyone trusts a new tool. By the second season, he'd stopped double-checking every routine adjustment by hand, the way he once did with the generic manual, and started saving his own judgment for the unusual days, storms, sudden temperature swings, the cases the tool flagged as uncertain itself.
Then, in the fourth season, a message arrived from a rival, BrightCurrent AI, offering "the same AI-powered feed optimization" at half Kelpstone's price. No single dramatic moment triggered the worry inside Kelpstone. It built slowly, over a few weeks, as more farms mentioned the pitch in passing during renewal calls.
Nobody at Kelpstone decided, on any single day, to start building a moat. The outcome archive just kept quietly accumulating in the background of doing the job well.
Marit's team could have panicked and rushed to match BrightCurrent's price, or its feature list. Instead, they pulled up what four seasons of paired recommendations and real harvest outcomes at Northwane actually looked like next to what a brand-new competitor could show.
Two products built on the same kind of engine. Only one of them had anything behind the number besides the engine.
Recommendation uncertainty band, by track record
Same kind of model underneath both. A completely different amount of real-world proof standing behind each number.
The model was never the moat. It was the part every competitor could rent by Tuesday. The moat was three extra seasons of proof nobody could rent at any price.
Marit's team made one deliberate call about what not to build: BrightCurrent was fast-following with a general chatbot feature, answering any fish-health question a farmer typed in. Kelpstone left that alone entirely, since matching it would spend real engineering time chasing an area where Kelpstone had no actual advantage over a wrapper.
Two things Kelpstone could have built. Only one of them was actually theirs to defend.
Access to a good model is common now. The other axis is the one that actually separates anyone.
Anssi renewed with Kelpstone that season, not because of loyalty, but because the confidence band on his own tablet told him something BrightCurrent's pitch never could: four years of Northwane's own fish had already told this system what worked.
What I'd tell myself, watching that first clean recommendation ship with no confidence band at all: the number felt complete the day it shipped, because there was nothing yet to compare it against. It stopped being complete the moment a competitor could match the number but not the years standing behind it, and by then, showing that difference should have already been built in.
SPARK, the anchor made visibleNot a pitch for a fancier model. SPARK is what shows exactly which design decision is the actual moat.
S
Situation. The job, done without you.
Anssi managing feed and oxygenation by gut feel, a generic manual, and a paper log, with real outcomes known only months later.
Grounding the answer in a real, unglamorous day keeps the anchor from floating free of any actual person.
P
Payoff. The habit you want built.
Anssi stops re-deriving every routine call from scratch and trusts the recommendation on the roughly 80 percent of days that are ordinary, saving his own judgment for the unusual 20 percent.
The habit, not the model's accuracy number, is the actual product being shipped.
A
Anchor. The one design decision.
Every recommendation ever made is logged against its real outcome, and that archive drives a visible confidence band on every new one.
This is the hardest step, and the direct answer to the whole question: the archive, not the model, is the moat.
R
Risk. What breaks the first time you're copied.
BrightCurrent AI launches the same model wrapper at half price. The anchor survives because the confidence band makes the difference visible instead of a matter of trust.
Designing the anchor to survive being copied is what separates a real moat from a temporary head start.
K
Keep out. What not to build yet.
A general fish-health chatbot, the feature BrightCurrent fast-followed with, deliberately left alone.
Chasing a feature where you have no real advantage spends the same time the actual moat needs to keep compounding.
The recap, one line per letter: situation is Anssi's day without Kelpstone, payoff is trusting the routine calls and saving judgment for the unusual ones, anchor is the outcome-linked archive driving a visible confidence band, risk is a copycat launching the same wrapper and the anchor surviving it visibly, and keep out is leaving the generic chatbot feature to the competitor chasing it.
And if you want to be sure it really works, try it somewhere elseSame five letters, a funeral home network instead of a fish farm. A completely different kind of outcome to track.
Grouse Hollow Funeral Home uses an AI tool to suggest service arrangements and scheduling based on a family's stated preferences. Mapped onto SPARK: situation is a funeral director building an arrangement by memory and instinct, with no way to know afterward how well a similar arrangement was received by other families in the past. Payoff is the director trusting the tool's suggested defaults for routine logistics, saving their own judgment for the emotionally sensitive parts of the conversation. Anchor here isn't outcome data in the fish-farm sense, since there's no equivalent "harvest result," so the real anchor is a different kind of proprietary signal: years of anonymized post-service feedback from families, linked back to which specific choices were made, which a brand-new wrapper competitor has never collected and can't buy. Risk is a low-cost competitor launching the same general-model-powered suggestion engine with no feedback history behind it, and the anchor surviving because Grouse Hollow can show, honestly, which of its suggestions are backed by hundreds of families' feedback and which are new and untested. Keep out is a generic grief-support chatbot feature a competitor ships fast, deliberately left alone since it's not where Grouse Hollow's real advantage lives.
A different kind of outcome entirely, and the same shaped answer: show the years behind the suggestion, don't just make the suggestion.
Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "the moat is the outcome-linked archive, not the model, and it has to be visible on the product itself or it isn't doing any work," and stop.
Cost: no engineering time to build a proper confidence-band display before the next renewal conversation. Say so honestly, and share the underlying numbers directly with the customer in the meantime, rather than letting the moat sit invisible.
The model got better, for real: if a newer general-purpose model closes some of the accuracy gap on its own, that's a legitimate reason to lean harder on the outcome archive as the differentiator, not a sign the moat has stopped mattering.
Where people run it wrong.
They describe their model or their prompt engineering as the moat, when a competitor with API access can rebuild that part in weeks.
They collect years of outcome data and never surface it to the customer, leaving the real advantage invisible and unused.
They chase every feature a fast-following competitor ships, instead of investing that time in the one thing that actually compounds.
How to use it live. The moment an interviewer asks about a data moat, ask yourself: if a well-funded competitor got access to the exact same general-purpose model tomorrow, what would still take them years to build? Answer that, and you've found the anchor.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework fits a design question like this, and what's its one-line job?
Tap to flip
ANSWER
SPARK: design against the failure before you build. (Swapped in for the flip-family slot, since this is a design question, not a perturbation.)
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Anssi Virtanen, who manages the tanks at Northwane Aquafarms, and Marit Haugen, the Kelpstone product lead who had to defend the company's actual advantage when a copycat launched.
3 · THE HABIT
What did Anssi stop doing once he trusted Kelpstone's routine recommendations?
Tap to flip
ANSWER
He stopped re-deriving every ordinary feed and oxygenation call from scratch, saving his own judgment for the unusual, flagged-as-uncertain days instead.
4 · THE ANCHOR
What is the one design decision this answer says is the actual moat?
Tap to flip
ANSWER
Logging every recommendation against its real outcome across every farm and season, and showing that track record as a visible confidence band on every new recommendation.
5 · THE OLD DECISION
What decision would you take back?
Tap to flip
ANSWER
Shipping the first season's recommendation with no confidence measure at all, since a clean single number looked fine before any farm had a real track record to compare against.
6 · THE NUMBER
Fill in the blank: Kelpstone's recommendation carries an uncertainty band of about ±___ percent, against BrightCurrent's ±18 percent.
Tap to flip
ANSWER
3 percent, backed by four seasons of Northwane's own outcome data versus BrightCurrent's zero.
7 · THE REPLAY
Same copycat launch, the confidence band already built and visible. What changes?
Tap to flip
ANSWER
Anssi and other customers can directly compare the two products' track records instead of taking Kelpstone's word for it, and the real advantage becomes a visible reason to renew rather than an invisible internal asset.
8 · CROSS PRODUCT TRANSFER
Section 4 answers this same question again for a different product. Which product, and what's the equivalent anchor?
Tap to flip
ANSWER
Grouse Hollow Funeral Home's service-arrangement tool. The equivalent anchor is years of anonymized post-service family feedback linked to specific past choices, in place of a fish farm's harvest outcomes.
Check yourself Score: 0 / 0
Multiple choice
1. According to this answer, what is Kelpstone's actual moat?
A. The general-purpose model it prompts.
B. Its lower price compared to competitors.
C. Years of recommendations logged against their real outcomes across farms and seasons.
D. Its wall-mounted tablet hardware.
Show hint
Look at the direct answer and the anchor step.
Show answer
C. The model wrapper can be copied in weeks. The outcome-linked archive took years and can't be rented or copied that fast.
True or false
2. True or false: this answer recommends Kelpstone match BrightCurrent's general chatbot feature to stay competitive.
True
False
Show hint
Look at the "keep out" step.
Show answer
False. Kelpstone deliberately leaves the generic chatbot feature alone, since matching it would spend time on an area with no real advantage over a wrapper.
Fill in the blank
3. Fill in the blank: feed-conversion ratio at Northwane Aquafarms improved from 1.35 in season one to ___ by season four.
Show hint
Look at the line chart tracking feed-conversion ratio by season.
Show answer
1.12. A steady improvement each season as the outcome-linked loop accumulated more of this specific farm's history.
Short answer, where it wouldn't matter
4. Name a part of Kelpstone's product where building a custom, proprietary model instead of using a general-purpose one genuinely wouldn't be worth it, and say why.
Show hint
Look at "what I would leave alone."
Show answer
Model answer: The underlying model itself. It's a fine engine already, and building a custom one from scratch would spend years chasing something the outcome archive already solves better.
Short answer, apply it yourself
5. Think of a product you use built on a general-purpose AI model. What could a well-funded competitor with the same model still not copy quickly?
Show hint
Think about something that takes real time or real users to accumulate, not just engineering effort.
Show answer
Model answer: A language-learning app's years of data on which specific corrections actually helped real learners retain a concept, tied to each learner's own history, not just a general grammar-checking model.
Short answer, work the number
6. If BrightCurrent AI operated for two full seasons before Kelpstone responded, roughly how would their uncertainty band likely change from the initial ±18 percent?
Show hint
Look at how Kelpstone's own uncertainty band narrowed across its first four seasons.
Show answer
Model answer: It would likely narrow meaningfully, perhaps into the low double digits, since two seasons of real outcome data would start closing part of the gap, though probably not all the way to Kelpstone's four-season number yet.
Before you close the answer
Why this works
Tests whether you can name a real, defensible design decision as the moat, or whether you'll describe the model itself as if that were still rare.
Follow-up traps
"Couldn't a competitor just buy outcome data from farms directly?" Response: they'd need years of farms actually using their product to generate it, and by the time they had a comparable archive, Kelpstone's own would have grown further too, so the gap doesn't close, it moves.
"Isn't a confidence band just a UI detail, not a real moat?" Response: the band is the visible proof of the moat, not the moat itself. Without showing it, the years of outcome data are just an internal asset nobody outside the company can act on.
If pressed
The outcome-linking process itself matches each recommendation to its result using a 90-day lag window after harvest, since mortality and weight data trickle in from different systems at different times, and the confidence band only updates once a full season's worth of matched pairs clears review.
From U2xAI Academy
From answering questions to owning outcomes.
A live workshop where you ship a working AI agent, defend a launch decision, and walk away with a portfolio recruiters can't wave off, not just more questions to study.
