How do you assess whether a competitor's capability is a moat or a thin wrapper?

Ground it in NordVantage Telecom, which runs GridSense, an AI outage-prediction system. Callum Ferro runs product there. A rival, Aeloria Systems, claimed to predict outages two hours earlier than the industry standard, and NordVantage nearly rebuilt a defensive feature before anyone ran the numbers on what that claim would actually take to copy.

How to answer this, stage by stage

Nobody is scoring whether you sound alarmed by a rival's press release. They're scoring whether you can show your work before deciding how alarmed to actually be.

Let's learn

Here's what happens when a competitor's scariest-sounding claim turns out to be the cheapest thing in the world to copy. Aeloria Systems announced it could predict grid outages two hours earlier than the industry standard. NordVantage Telecom, which runs its own outage-prediction system called GridSense across forty thousand utility poles, treated the claim as an emergency.

Nobody at NordVantage checked which of these two boxes Aeloria's claim actually belonged in.

Leadership ordered a defensive feature within the week: a copycat "two-hour early warning" flag, rushed to ship before anyone estimated what it would actually take to build it properly, let alone what it would take for Aeloria to have built theirs.

Here's the turn: when Callum Ferro finally ran the numbers, months later, Aeloria's actual pipeline turned out to be public weather feeds and customer-uploaded grid schematics, run through a prompted foundation model, no proprietary sensor network at all. Copying it properly would have taken three to five engineer-weeks. NordVantage had spent eleven months on a rushed, unreliable copy that never should have taken more than a month.

At its worst, skipping the estimate doesn't just waste engineering time. It spends eleven months of a team's attention chasing a rival's shadow, time that could have gone toward NordVantage's own real advantage instead.

What I would leave alone: I wouldn't tell NordVantage to never react quickly to a competitor. Some real moats do need a fast response. The fix isn't slower reactions across the board, it's one estimate, done first, that tells you whether speed is actually warranted.

The lesson: a claim that sounds like a moat and a claim that actually is one look identical in a press release. The only way to tell them apart is to price out what it would take to build it yourself.

Now here is the same thing as a story

The short version above is what you'd say defending this method under interview pressure. Read this one for what the rushed response actually cost on the ground.

The alert sits as a single red banner across the top of Beatrix Halloran's dispatch screen, and for eleven months before any of this, she'd trusted it completely. GridSense's outage predictions had never once sent a crew somewhere for nothing, and she'd stopped second-guessing it entirely by her third month on the job.

One of these two foundations took a decade to build. The other took a data-sharing agreement and an API key.

When leadership rushed the copycat "two-hour early warning" feature to answer Aeloria's claim, it shipped without the validation GridSense's original predictions had gone through. Within its first month live, it cried wolf twice, dispatching crews to substations with nothing wrong at either one.

Four weeks from launch to a dispatcher who'd stopped trusting the newest flag entirely.

By the fourth week, Beatrix had stopped acting on the new early-warning flag the moment it fired, waiting instead for GridSense's original, validated prediction to confirm it before dispatching anyone. The trouble was, the doubt didn't stay contained to the new feature. She caught herself, twice that same month, double-checking an original GridSense alert she would never have questioned three months earlier.

Three ordinary components. None of them took ten years to build.

When Callum Ferro finally ran the estimate, eleven months in, he broke the replication cost into its real parts: a prompt and orchestration layer over a foundation model, public weather data integration, and a monitoring dashboard, each one to two weeks of a small team's time. He sanity-checked it against how long NordVantage's own leadership had spent debating whether to partner with Aeloria instead of competing: longer than the estimate itself.

The old decision, to merge "should we respond" and "how do we respond" into a single rushed motion, had been made in a leadership meeting during the first week of the scare, when the instinct to look fast felt more urgent than the instinct to look carefully. That instinct made sense for exactly one week. It stopped making sense the moment "urgent" had already quietly been true for a month with no cost attached to spending four more days on an estimate.

The replay: same Aeloria announcement, same leadership pressure to respond fast, but the estimate runs first, before any build decision. Three to five engineer-weeks to copy the actual pipeline, sanity-checked against the length of the debate itself. NordVantage either matches the feature properly in a month, or, more likely, spends that month confirming Aeloria has no exclusive data deals and redirects the engineering time toward GridSense's own real advantage instead. Beatrix never sees an unreliable flag, and her trust in the original, validated prediction never gets tested by a rushed feature that had nothing to do with it.

What Callum took from it wasn't "always move fast" or "always move slow." It was that the decision to rush and the decision to estimate are not actually in tension, the estimate itself only takes days, and skipping it is what turned a four-week problem into an eleven-month one.

BOUND, the five letters that price out a claimNot a gut call on how scary a headline sounds. BOUND is the arithmetic that tells you.

The recap, one line per letter: break it down is the cost equation for data, integration, and distribution, own numbers is grounding each assumption in Aeloria's own public materials, use a range is three to five weeks instead of a false-precision single figure, nail the sanity check is comparing the estimate to how long the internal debate itself took, and direction is naming the one hidden data-exclusivity assumption that would flip the whole answer.

And if you want to be sure it really works, try it somewhere elseSame five letters, a textile mill instead of a power grid. A different old decision breaks this one.

A textile manufacturer's quality-assurance team heard a rival supplier claim an AI fabric-defect scanner that "catches flaws no human inspector can see." Mapped onto BOUND: break it down is separating the claim into camera hardware cost, the vision-model training cost, and the cost of building a defect-labeled dataset large enough to train it well. Own numbers assumes commodity high-resolution line-scan cameras, since the rival's marketing photos show standard industrial hardware, not a custom sensor. Use a range gives six to ten engineer-weeks to build a comparable model, assuming access to a similarly sized labeled defect dataset, which is the real open question. Nail the sanity check compares that to the eighteen months the rival actually spent building their own defect-labeled dataset from scratch, a number obtained from a former employee's public conference talk, suggesting the model itself is cheap but the dataset behind it took far longer than the marketing implies. Direction is that the single assumption moving this estimate most is whether a comparable labeled dataset can be bought or licensed rather than built fresh, since buying one collapses the timeline back down to weeks, while building one from scratch is the actual multi-year moat. The old decision here isn't a merged step, it's a workaround: engineers on the QA team, distrusting an earlier defect-scanner rollout that gave no reasons for its flags, had built a private spreadsheet cross-checking every flagged bolt of fabric by hand, a shadow process that would have made any new AI tool's real accuracy nearly impossible to measure cleanly.

This is what a real moat actually looks like laid out in parts, not a marketing claim.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "price out what it would cost a competent team to copy the claim before reacting to how it sounds; cheap to copy means it's a wrapper," and stop.
Cost: there's no time or budget for a deep estimate before a leadership decision. Do the rough version in a day, using only public materials, rather than skipping it entirely.
The model gets better, for real: if a rival's capability genuinely turns out to depend on years of proprietary data you can verify, that's the moment to treat it as a real moat and respond accordingly, not a reason to distrust every future estimate.

The same five-step shape, run on fabric instead of power lines.

Where people run it wrong.
They react to how a claim sounds instead of pricing out what it would actually take to copy.
They give a single confident number instead of a range, which invites false precision about a rival's internals nobody outside the company can fully see.
They skip the sanity check and never notice when the estimate itself doesn't survive an obvious comparison.

How to use it live. The moment an interviewer asks you to judge whether a rival's capability is real, ask yourself: what would it cost me, in weeks, to build the same thing? Say the number, and let that number do the judging instead of the headline.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0