Describe how you would size an AI opportunity before knowing whether it is technically feasible.

Plumbline is Anvilrow Technologies' tool that reads a stack of blueprint PDFs for a construction project and drafts a full material takeoff and labor estimate, a bid draft a general contractor's estimator checks before it goes out the door. Leofric Cartwright, the product manager who owns Plumbline's feasibility and sizing calls, found out how far a clean number can travel before anyone tests it. Dorothea Kingsley, Anvilrow's VP of Growth, sized a new market at fourteen million dollars and carried it straight to the board. Ysabel Everhart, who owns Everhart Restoration Group, just wanted a bid she could trust before her busiest season started.

How to answer this, stage by stage

Nobody is grading whether you think Dorothea's math was good. They're grading whether you can turn "size it before you know if it works" into a specific, defensible number, or two.

Let's learn

Plumbline sits behind one upload button on Anvilrow's estimating dashboard, the button a general contractor's estimator clicks after dragging in a stack of blueprint PDFs. A few minutes later, a full material takeoff and a labor estimate come back: how much lumber, how much concrete, how many sheets of drywall, priced and ready for the estimator to check before the bid goes out.

On new-construction projects, clean CAD-native blueprints straight out of an architect's design software, Plumbline is checked hard. Against 600 real sheets, matched by hand against a certified estimator's own takeoff, it lands within 5 percent of the right quantity 95 percent of the time. That's the number on Anvilrow's homepage, and it's a real, earned number.

Renovation and retrofit projects run on a different kind of document entirely. Blueprints for an existing building get scanned, marked up with hand-drawn revision clouds where something changed mid-project, stamped by two or three different permit reviews, sometimes carrying two different scale bars from two different renovation eras on the same sheet. Nobody had ever run Plumbline against a real batch of them.

One dot on this chart had actually been checked when Dorothea built her slide. The other two were still just a guess wearing a decimal point.

Here's the turn. Dorothea Kingsley, Anvilrow's VP of Growth, wanted to size the renovation-bid opportunity for the board. She took the 95 percent number, the only accuracy figure she had, and carried it forward onto a market of general contractors whose blueprints Plumbline had never once been tested against. The mistake wasn't the market math. Renovation-focused contractors really are a big market. The mistake was letting one clean number stand in for a technical question nobody had actually answered yet.

Four things on the slide. Three of them are real math. None of them says the accuracy number had never been checked on this kind of document.

At its worst, this is Sixtus Fairfax's engineering team six weeks into a ten-week roadmap slot, building screens and an upload flow for a capability the model can't yet reliably run. It's Ysabel Everhart, who owns Everhart Restoration Group, four weeks from her busiest bid season, holding a promised pilot that only works cleanly on two of the six renovation jobs she tried it on. And it's a board still carrying fourteen million dollars in projected revenue that nobody has told them is built on a number nobody checked.

The lesson: an opportunity that depends on an untested model capability isn't a market-sizing problem wearing a technical costume. It's a technical question wearing a business suit. The job isn't picking a confident number. It's finding out, on real documents, what's actually true before the number ever leaves the room.

Now here is the same thing as a story

Stage five above compresses this into four sentences. Here's the ten weeks underneath, the part a stand-up answer skips.

Leofric Cartwright can tell, from the first three sheets of a blueprint set, roughly how good a week Plumbline is about to have. Four years into owning the model's feasibility and sizing calls, he reads a set of drawings the way an old mechanic listens to an engine: not for what's wrong yet, just for what kind of thing he's looking at.

Dorothea Kingsley runs growth at Anvilrow, and she's good at her job in a specific way: she can turn the rough shape of an opportunity into a number a board will actually act on. Ten weeks before the quarter she'd usually build a growth deck for, she pulled up Plumbline's own numbers page. Ninety-five percent, on new-construction takeoffs. She multiplied that against the renovation-contractor market, a market Anvilrow's sales team had been asking about for a year, and landed on fourteen million dollars in new revenue within eighteen months, if the self-serve motion that worked for new-construction customers worked here too.

She didn't invent the ninety-five percent. She just never asked whether it applied to a document nobody had shown her yet. Nobody in the room asked either. The board approved the number on a Tuesday in April, and by Thursday, Sixtus Fairfax's engineering team had a Q3 slot: ten weeks to ship renovation-bid support.

Neither person here did anything careless. One of them just had the only copy of the assumption underneath the figure.

Dorothea, wanting a flagship account to point to, called Ysabel Everhart that same week. Ysabel owns Everhart Restoration Group, a renovation-focused general contractor two towns over, the kind of company the fourteen-million-dollar number was supposed to represent by the thousand. Dorothea promised her a working pilot in time for the fall bid season, ten weeks out. Ysabel, who'd spent years pricing renovation jobs by hand, said yes on the spot.

For six weeks, Sixtus's team built. A new upload flow for renovation projects. A revised results screen. Nothing about the model itself, because nobody had asked them to touch it. Why would they. The number said ninety-five percent.

In week four, in an ordinary Tuesday planning meeting, Cassander Woodrell, three weeks into the job, asked the question nobody else in the room had thought to say out loud: "Wait, has anyone actually run this on a real renovation set yet? Like an actual scanned one?" The room went quiet in the specific way a room goes quiet when the answer is no and everyone realizes it at the same moment.

Four of these five weeks looked completely fine, if the only thing anyone was watching was the roadmap burn-down.

Leofric ran the spike that should have run back in April. Two weeks, forty real renovation blueprint sets, two hundred forty pages, pulled from Anvilrow's own customer files. He also, on a hunch, ran fifteen sets of historic-restoration drawings, hand-drafted originals from buildings old enough that nobody had ever digitized the plans. The renovation number came back at sixty-one percent. The historic number came back at thirty-four.

The model wasn't guessing randomly. It was doing something specific, and once Leofric looked closely, obvious in hindsight: it read a hand-drawn revision cloud, the scribble an architect draws around a part of a drawing that's changed, the same way it read a real wall boundary. Both are just closed loops on a scanned page. Nothing in training had ever told it the two were different.

By week six, with four weeks left before Ysabel's bid season, Leofric had to tell Dorothea, Sixtus, and the board three things at once: the model couldn't reliably run renovation bids yet, not at the quality the fourteen-million number assumed. Sixtus's team had spent six of ten weeks building around a capability that wasn't actually there. And Ysabel was going to find out, one way or another, before her season started.

The decision Leofric would take back sits earlier than any of that. Anvilrow had expanded Plumbline twice before, once into a new state's building codes, once into a new trade's material list, and both times the ninety-five percent held, because both times the underlying documents were the same clean CAD files the model had always seen. Two clean wins in a row had quietly taught everyone the number just travels. Nobody had ever built a rule saying it wouldn't always.

So here's what changed. Every sizing memo that reaches the board or becomes a roadmap commitment now has to carry two numbers, not one, each tied to a named scenario, plus a dated spike that resolves which one is real. Dorothea's redo, run properly this time: if a lightweight fine-tune closes the gap on renovation documents, the opportunity is eleven million in eighteen months, close to the original. If it needs a heavier pipeline, handwriting-aware extraction plus a mandatory human check before any renovation bid ships, it's three million in eighteen months, but could reach twenty million in three years once the review cost comes down. The two-week spike, run in week two this time instead of week six, is what tells you which one you're actually building toward.

Ysabel got a real answer in week two of the redo, not week six: the heavier pipeline was needed, her bids would carry a short human check for now, and Plumbline would still save her most of a day per bid, just not the fully hands-off version she'd been promised. She kept the pilot. She just kept it with an accurate picture of what it actually was.

What I'd tell myself, sitting in that April board meeting where nobody asked if the number applied: two wins in a row isn't a pattern. It's a coincidence you haven't been charged for yet.

GUARD, checked against one clean number and a ten-week clock

This was never really about whether Dorothea's math was good. It was good. GUARD is for naming who actually pays when a number nobody's tested reaches a roadmap before the thing it's about does.

Four of these five steps happened exactly as designed. The missing one, a completed spike, was never built into the path at all.

Four lines. None of them apologize for not knowing yet. They just say what's actually known, and what isn't.

And if you want to be sure it really works, try it somewhere else

Same five letters, a claims-coding tool instead of a bid tool, and this time the untested documents are surgical operative notes instead of scanned blueprints.

Codewright, built by Marrowbank Health Analytics, reads a clinician's note and assigns the billing codes for a claim, so a practice can submit it without a coder starting from scratch. Anatoly Cavanaugh owns Codewright's coding model the way Leofric owns Plumbline's.

Same three branches Anvilrow faced with Plumbline, a different document entirely. The honest branch is still the one that survives an actual spike.

On primary-care visit notes, checked against 800 real claims and a certified coder's own assignment, Codewright gets the full code set right 92 percent of the time. Sylvester Millhouse, who runs revenue-cycle partnerships, sized the opportunity of expanding into orthopedic surgical coding at nine million dollars in new revenue within twelve months, carrying the 92 percent number forward the same way Dorothea carried Plumbline's.

Same rank, mapped onto Codewright: size it as two numbers. Surgical operative notes bundle several procedures into one note, and deciding which ones are separately billable and which are folded into another code is a judgment call nobody had tested the model against. A two-week spike against fifty real orthopedic notes found the real accuracy at 58 percent, not 92, because bundling logic doesn't behave anything like a single visit's diagnosis code. If a lightweight fix, flagging bundling questions for a human coder to check, closes enough of the gap, the opportunity is seven million in twelve months. If it needs a separate model trained specifically on surgical bundling rules, it's two million near term, climbing toward fifteen million in three years.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: never let a single point estimate leave the room without naming which feasibility scenario it assumes. A number with no range attached is a guess wearing an estimate's clothes.
Cost: no budget for a full two-week spike this quarter. Run a scoped one-day test on 10 to 15 real documents instead of zero. A small real sample beats an extrapolation from a completely different population.
The model got better, for real: say Plumbline's renovation accuracy jumps to 90 percent next quarter after a training pass on scanned documents. That's a reason to update the range and lift the mandatory-review requirement, on purpose, with a new eval run backing it up, not a reason it should have been assumed at 95 percent from day one.

Where people run it wrong.
They treat sizing as a market-research exercise only, and never connect the number to whether the model has actually been tested on the real documents it would need to read.
They let whichever number reaches the board first become the anchor everyone quietly designs around, even after a spike proves it wrong.
They wait for a blown deadline or an upset customer to reveal the gap, instead of gating any external promise behind a dated feasibility check.

How to use it live. Before answering, ask out loud: "has anyone actually run this on real renovation blueprints yet, or are we assuming it behaves like new-construction ones?" Naming that split buys you a few seconds of thinking time, and it's most of the real answer.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0