Describe three ways to communicate uncertainty without displaying a number.

Northfield Community Lending underwrites small-business loans. Devon Ashworth is a senior credit analyst there, ten years into reading loan files. LoanScope is the AI tool that reads each application and recommends approve, decline, or a closer look, before a human ever opens the file.

How to answer this, stage by stage

Look, this isn't a trivia question about knowing three techniques exist. It's about proving you'd build them in an order that survives a real engineering quarter.

What unblocks what, and what ships first

Here's what I'd actually build, in order, for Northfield's underwriting queue.

All three do the same job. They cost very different amounts to build well.

Before LoanScope, Devon's team read every small-business file cold, an average of forty minutes per application, checking revenue trends, debt load, and owner credit history by hand. LoanScope reads the same file in under ten seconds and cut the average time to a first read to about six minutes.

Here's the turn: the extra speed was never the risk. The risk was that every recommendation, clean file or genuinely messy one, came out in the exact same confident sentence, with nothing in the wording to tell Devon's team which files actually deserved a second look.

Each box needs something the one before it doesn't. Skipping ahead is how a team ends up shipping the hardest one first, badly.

At its worst, shipping the expensive reasoning trail first, before anyone had confirmed hedge language even worked, means burning a whole quarter on a feature analysts still can't parse, while the cheap fix that would have caught real problems sits unbuilt.

Not every file needs the expensive treatment. Most need nothing more than an honest sentence.

What I would leave alone: Devon's own weekly calibration meeting, where the analyst team compares notes on files LoanScope got wrong. That process doesn't need a number or a UI change, it needs the room it already has.

The lesson: the cheapest honest sentence usually catches more real problems than the fanciest feature nobody's tested yet.

Now here is the same thing as a story

The short version above is what you'd say defending the build order to Northfield's engineering lead. Read this one for how the mixed-signal file almost slipped through.

Devon Ashworth had underwritten small-business loans at Northfield for a decade, and could smell a shaky revenue projection before he'd finished the first page. The morning in question was a Tuesday, nothing unusual about it, the queue moving at its normal pace.

A landscaping company's file came through. LoanScope's recommendation read: "Recommend: Approve." The exact same sentence, same confident tone, it used for every clean file that morning.

Buried in the landscaping company's file were two different years of reported revenue that didn't reconcile, the kind of thing a careful read catches in minutes. LoanScope's recommendation never flagged it, because nothing in its wording ever changed based on how mixed a file's signals actually were.

Devon's team could have fixed the wording in an afternoon. Instead, nobody had built anything at all yet.

A routine quarterly audit, pulling ten random approved files, found the reconciliation gap six weeks later. The loan hadn't defaulted yet. It was still, technically, fine. But nobody at Northfield had caught it on purpose, and that was the part that worried Devon's manager more than the gap itself.

Devon's team shipped the cheapest fix within two weeks of the audit. The expensive one came three months later, tested and working.

ORDER, in one screenNot a style guide for wording. ORDER is what tells you which of the three to build this sprint.

A reasoning trail is four separate parts. Building all four badly is worse than shipping none of them yet.

The recap, one line per letter: outcome is an analyst who checks the right files harder without a bare number, reversibility puts hedge wording ahead of anything the model has to explain about itself, dependency is history data and structured rationale gating the pricier two, evidence is a cheap comprehension test before the expensive build, and rank is hedge language, comparable cases, reasoning trail, cheapest first.

And if you want to be sure it really works, try it somewhere elseSame five letters, a city permits office instead of a lender. A different dependency breaks the second story.

Ashgrove City Permits Office uses PermitAssist, an AI tool that checks building permit applications against code and flags likely compliance issues before a human reviewer looks. Harlan Cobb is the permits reviewer who handles the flagged queue. Mapped onto ORDER: outcome is a reviewer who spends more time on genuinely borderline applications, not an average review time that looks good on paper; reversibility still puts hedge wording first, cheap to write and cheap to withdraw if it reads wrong to applicants.

The dependency step works in reverse here. At Northfield, engineering readiness gated the order. At Ashgrove, legal readiness does: showing a comparable past permit as evidence requires the city attorney's office to confirm that citing another applicant's file, even anonymized, doesn't create a fairness claim if the two permits get decided differently. That review had to clear before the second-ranked technique could ship at all, regardless of how fast engineering could build it.

Same four parts. "Similar past permit" is the one that needed a legal sign-off before it could ship at Ashgrove.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "hedge language first, comparable cases second, a full reasoning trail last, cheapest and safest to undo first," and stop.
Cost: there's no budget this quarter for the history data pipeline the second technique needs. Say so honestly, and ship the hedge language alone, since it needs nothing else to start working today.
The model gets better, for real: if LoanScope's underlying accuracy genuinely improves, the hedge wording should get more confident on more files, not disappear entirely, since even a strong model will still meet a genuinely unusual case sometimes.

Where people run it wrong.
They build the flashiest technique first because it demos well, not because it's cheapest to test and safest to undo.
They use one wording for every file, clean or mixed, so the language stops carrying any real signal at all.
They skip the cheap comprehension test and find out the wording didn't work only after the expensive feature is built on top of it.

How to use it live. When someone asks for ways to show uncertainty, ask yourself: which of these ships without new engineering? Say that one first, out loud, then build up from there.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0