A vendor offers 90 percent of what you need. How do you evaluate closing the last 10 percent yourself?

Harborview Insurance processes auto claims. FieldPilot is the vendor tool that reads a submitted accident report and estimate, and pulls out the standard fields automatically. Renata Sooklal runs claims operations, and one near miss on a handwritten estimate is what forced her to actually price the gap instead of guessing at it.

How to answer this, stage by stage

Nobody is scoring whether you'd like to close the gap. They're scoring whether you can price it honestly enough that the answer might come back "don't."

Let's learn

Every month, Renata Sooklal's claims team spent roughly 560 hours simply retyping paperwork a vendor tool was supposed to have already handled.

FieldPilot reads a submitted accident report and estimate and pulls out the standard fields, date, VIN, adjuster notes, automatically. On about 2,000 claims a month, it gets roughly 1,800 of them clean. The remaining 200, mostly handwritten damage estimates from small independent repair shops and a handful of non-standard state forms, still need a person to type them in by hand, about 14 minutes each.

The last 10 percent is rarely one thing. It's usually a handful of different messy cases, each with its own real cost to close.

Here's the turn: the manual typing itself was never really the problem, the team had already built that into their staffing. The real question showed up the first time someone asked whether Harborview should just build its own handwriting extractor to close the gap for good, and nobody had actually priced what that would cost against what the gap was already costing.

At its worst, skipping this arithmetic means sinking six figures into a custom handwriting extractor for a slice of claims that was already being handled, reliably enough, by one person's typing, and still not fully closing the gap, because messy handwriting stays messy no matter how much engineering you throw at it.

What I would leave alone: a genuinely rare edge case, like claims involving classic or vintage cars, under half a percent of volume, doesn't earn its own custom build no matter how annoying it is when it shows up. Route it to a person and move on.

The lesson: "the last 10 percent" isn't one number, it's usually several different gaps with different real costs. Price each one against what it already costs to live with, and let the arithmetic say no when it says no.

Now here is the same thing as a story

The short version above is what you'd say out loud in the room. Read this one for what actually pushed Renata to run the real numbers instead of trusting a feeling.

Renata Sooklal could tell a real coverage dispute from a simple paperwork error before she'd finished her coffee, four years running claims operations at Harborview will do that. FieldPilot had been live for a year, and her team had settled into a rhythm: 1,800 claims a month moved through clean, and the 200 that didn't got quietly typed up by a rotating pair of processors.

The mistake wasn't ignoring the gap. It was treating every part of it as equally worth building for.

A new claims director had floated the idea of building a custom handwriting model to close the gap for good, and it sounded reasonable, ten percent felt small enough to mop up quickly. Nobody had actually priced it. Then, on a Tuesday afternoon, a handwritten estimate came in with a smudged final digit.

Nobody designed a catch for this. It worked because a reviewer happened to double back on the file.

FieldPilot's fallback OCR read the smudged estimate as 2,400 dollars instead of 8,400. The claim moved into the auto-approval queue, and a reviewer only caught it because she happened to reopen the file to check an unrelated detail. It rattled the team enough that Renata finally asked for real numbers instead of a feeling.

A two-week sample of 200 flagged claims showed the gap breaking into a few real shapes: handwritten estimates, a couple of out-of-state forms, and a small number of second-language submissions. Engineering estimated six weeks to build a rules-plus-model patch for the handwriting piece alone, roughly 75,000 dollars, plus 15,000 for curating enough real examples, 8,000 for a proper eval set, and 22,000 a year to keep it maintained as handwriting styles and forms kept drifting.

Four real line items, not one round number. Each one had to be priced on its own.

Against 200 manual claims a month at 14 minutes each, roughly 24,000 dollars a year in loaded labor cost, the 120,000 dollar first-year build didn't come close to paying for itself, even before counting the 22,000 dollars a year it would keep costing after that. Building it made sense as an idea. It stopped making sense the moment real numbers sat next to each other.

Harborview's gap was common but not costly enough yet, and the vendor had no roadmap fix in sight, which is what kept the honest answer at "don't build."

Here's the replay: with the real numbers in hand, Renata kept the gap routed to her team, added a second reviewer check specifically for estimates over 5,000 dollars, and set a standing quarterly recheck of the math. Six months later, the gap's volume hadn't grown enough to change the answer, and the near miss never repeated, because the new check would have caught it even without a model.

What I'd tell myself, standing in that room when the handwriting-model idea first came up sounding so reasonable: ten percent is not a size, it's a placeholder for several different gaps, and none of them deserve a build decision until you've actually priced what they cost to live with.

BOUND, run on Harborview's vendor gapOne line per letter, if you want to say it fast under pressure.

The recap, one line per letter: break it down is the equation, build cost against living-with-it cost, own numbers is the real 200-claim sample behind every figure, use a range is 95 to 140 thousand dollars instead of one falsely precise number, nail the sanity check is comparing it to a headcount cost, and direction is the gap's volume, the one thing that would flip this decision if it changed.

And if you want to be sure it really works, try it somewhere elseSame five letters, a permit reviewer's vendor gap instead of a claims processor's.

Efe Adaramola reviews building-permit applications at Millbrace Permits, which sells software cities use to process them. A vendor tool pre-checks applications for missing required documents and clears 92 percent of submissions cleanly. The remaining 8 percent are applications bundled as single scanned PDFs instead of separate files, which the vendor tool can't split apart. Mapped onto BOUND: break it down is engineering time to build a PDF-splitting step, plus a small eval set of misfiled bundles, plus light ongoing tuning as new scanner software changes bundle formats. Own numbers: a one-month sample shows 340 affected applications, each costing a reviewer 9 minutes to manually split and refile. Use a range: build cost lands between 30,000 and 45,000 dollars, far lower than Harborview's gap, because splitting a PDF is a much narrower problem than reading handwriting. Nail the sanity check: 340 applications a month at 9 minutes is only 51 hours, about 15,000 dollars a year in labor, so even the low end of the build estimate takes two years to pay back. Direction: the swing factor here is whether the city's own scanning vendor fixes bundling on their end first, which would erase the gap for free. Efe's team decides to wait and recheck in two quarters rather than build, since the vendor fix is plausible and cheap to wait for.

Same method, a different shape of picture: Millbrace's decision plotted as a timeline instead of a tree, since the real story here is when to recheck, not which branch to take.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "price the build cost against the cost of living with the gap, in real numbers, and only build if the build cost clearly wins," and stop.
Cost: no time to pull a real sample before a decision is due. Say so honestly, and make pulling that sample the very next step, not a guess dressed as an estimate.
The model got better, for real: a new vendor release closes half the gap on its own. Rerun the numbers, since the remaining gap is now smaller and the build case gets weaker, not stronger.

Where people run it wrong.
They estimate the build cost and never price the cost of living with the gap, so there's nothing to compare it against.
They treat "10 percent" as a size instead of asking what specific cases make up that ten percent.
They skip the sanity check, so a 120,000 dollar build against a 24,000 dollar problem ships without anyone noticing the mismatch.

How to use it live. The moment someone says "the vendor gets us most of the way there," ask what the missing piece actually costs to live with, before asking what it would cost to build. That number alone usually settles the argument.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0