ConceptIntermediateDesigning for Uncertainty & Trust / Designing for failure and graceful degradation / #20
Describe the difference between a retry that helps and a retry that just costs money.
PICK the same pixels, sent again, expecting a different answer for free
Fenwick Mutual runs an AI damage-assessment tool that scores photos of storm damage for auto and home claims. Teodor Vance is a field inspector who stands in customers' driveways all day, tablet in hand, submitting photos the tool either scores or kicks back.
The direct answer
A retry helps only when something new enters the system between attempts, a fresh photo, a network connection that came back, more time for a genuinely unsure score to resolve. A retry that resubmits the exact same input to the same model, expecting a different result, is not a retry. It is the same question asked again for money.
Do this, in order
Diagnose why a submission failed before deciding whether to retry it.Why: this single check is what separates a retry that can possibly help from one that can't.
Only retry when new information can enter the next attempt.Why: a recaptured photo has new pixels. A resubmitted one doesn't, and the model has no reason to answer differently.
Route input-quality failures straight to a human or a recapture prompt, never back through the model again.Why: glare, ambiguous damage, and genuinely blurry shots will fail the same way every time, since the input itself is the problem.
Cap blind retries at zero for any failure category that's already been shown not to improve on resubmission.Why: each blind retry is a real inference cost with no chance of a different outcome, which makes it pure waste, not caution.
Track success rate on identical resubmissions, split by failure reason.Why: this is the one number that tells you which failure categories are worth retrying and which never will be.
Keep retrying network or timeout failures automatically, since those are genuinely transient.Why: unlike a bad photo, a dropped connection really can succeed on a second try with no new input needed.
How to answer this, stage by stage
Nobody is grading whether you know what a retry is. They're grading whether you can tell, for one specific failure, if trying again could possibly change anything.
Stage 1
Scope it to one photo, one failure
Say it like this
"I'll ground this in Fenwick Mutual's damage-assessment tool, and one specific case, a storm-damage photo that fails to score."
Why this works
Keeps the answer from floating into a generic "retries are good" or "retries are bad" debate.
Stage 2
Name your structure out loud
Say it like this
"I'll use PICK. Position, my call up front. Impact, who feels each kind of error. Cost asymmetry, which one is actually expensive. Kill criteria, what would change my mind."
Why this works
Signals you're about to commit to something, not hedge for two minutes first.
Stage 3
Take your position before any hedging
Say it like this
"A retry only helps when new information enters between attempts. Same photo, same model, same input, that's not a retry, that's the same question asked twice."
Why this works
This is PICK's whole test: commit to a position before explaining it.
Stage 4
Name who feels each kind of error
Say it like this
"A blocked retry costs a few seconds and one recapture prompt. A blind retry loop costs three inference calls, nine extra seconds, and still lands on a human's desk anyway."
Why this works
Turns an abstract tradeoff into two costs you can actually compare.
Stage 5
Point at the real cost asymmetry
Say it like this
"The recapture prompt looks slower in the moment, one extra tap. The blind retry loop is the one actually burning money, three calls for a photo that was always going to fail the same way."
Why this works
This is the heart of PICK: naming which cost is hidden and expensive versus visible and cheap.
Stage 6
Give your kill criteria and close
Say it like this
"If a failure category ever shows a real success rate on identical resubmission, above two percent, say, I'd revisit it. Until then: a retry that helps changes the input. A retry that doesn't is just a second bill for the same answer."
Why this works
Shows the position isn't stubborn, it's evidence-based, and restates the direct answer in one breath.
Let's learn
What actually happens the second time an AI model is shown the exact same photo it just failed to score?
Fenwick Mutual's tool scores a photo of storm or accident damage in seconds, instead of a claim sitting in a queue for a human adjuster to review by hand. Before it, every claim waited two to four days for a person to look at the photos.
The old system skipped straight from the first box to a blind third attempt, missing the check that actually matters.
Now most photos score instantly. On a failure, the tool automatically resubmits the identical photo up to three times before giving up and routing it to a human anyway.
Here's the turn: the extra attempts were never actually fixing anything for a whole class of failures. A photo that fails because of glare or a genuinely ambiguous damage pattern will fail the exact same way every single time, since nothing about the photo changed between attempt one and attempt three.
Cost per failed photo, blind retry vs diagnose-then-act
Three times the cost and three times the delay, for a photo that landed on a human's desk either way.
At its worst, Fenwick pays for thousands of pointless resubmissions a month while claimants wait through a delay that bought nothing.
One of these looks like the slower option. It is not the expensive one.
The decision I would take back
Fenwick's engineers merged "detect a failure" and "retry automatically" into a single step, skipping the check for why the photo failed at all. That made sense when most early failures really were transient network drops, and a blind retry usually did help. It stopped making sense once glare and ambiguous damage became common failure reasons that no amount of resubmitting could ever fix.
What I would leave alone: network timeouts and dropped connections should still auto-retry without asking a human anything. That failure really is transient, and the same photo really can succeed the second time.
The lesson: a retry only ever buys you something new to look at. If nothing changed, you didn't retry, you just paid twice for the same wrong answer.
Now here is the same thing as a story
The short version above is what you'd say defending this fix to Fenwick's claims-technology council. Read this one for how the waste actually got found.
Teodor Vance has inspected storm damage for Fenwick Mutual for four years, standing in driveways with a cracked tablet, snapping roof and siding photos while a homeowner watches from the porch. He can usually tell hail damage from wear damage before the tool even finishes scoring.
Knowledge spark: why would resubmitting an identical photo ever change the model's answer?
It mostly wouldn't. A model given the exact same input tends to return the exact same, or nearly the same, output. The only real exceptions are small amounts of randomness in how some models generate an answer, and that randomness almost never turns a genuinely unclear photo into a clear one.
For months, Teodor noticed his queue never actually moved faster on cloudy, glare-heavy days, even though the tool was supposedly "smart" about retrying. He assumed it was just a harder day for photos. He never timed it, so he never noticed the nine extra seconds per failed shot, three times over, quietly stacking up behind every glare-blown roof.
Nine seconds spent proving something that was true from the very first attempt.
A finance-side audit, run to understand why inference spend had crept up faster than claim volume, pulled a sample of logs and found it: nearly a third of all model calls were identical resubmissions of photos that had already failed once, for reasons that had nothing to do with a bad connection.
The claim wasn't taking three tries to solve. It was taking one try, paid for three times, before anyone admitted a person needed to look at it.
The audit traced it to one merged step: the system checked "did this fail," not "why did this fail," and treated every failure as worth an automatic second and third try.
Three failure types cluster in the same corner. Only one belongs anywhere near "retrying helps."
With the redesigned check, a failed photo is first classified: glare or ambiguous damage routes straight to a human reviewer with no further model calls, blur triggers an on-screen prompt asking Teodor to retake the shot, and only a genuine network timeout auto-retries. Run the same driveway forward: the glare-heavy roof photo goes straight to a human on the first failure, at a third of the cost and none of the wasted seconds.
The old system asked the model the same question three times and called that effort. The new one asks once, then asks the right next question.
I built the retry logic to fail safe, when in doubt, try again. It took an inference-spend audit to see that "try again" and "actually try something different" were never the same move.
PICK, in one screenNot a debate about whether retries are good. PICK is what tells you which one is a real second chance.
P
Position. The call, before any nuance.
A retry helps only when new information enters between attempts. Same input to the same model is not a second chance, it's the same question paid for twice.
The hardest step: commit to the rule before hedging into "it depends."
I
Impact. Who feels each kind of error.
A blocked, diagnosed failure costs a claimant one extra tap to recapture a photo. A blind retry loop costs Fenwick real inference money and the claimant nine extra seconds, for the same eventual outcome.
Turns "retries cost money" into two comparable, named costs.
C
Cost asymmetry. Which one is actually expensive.
The recapture prompt looks like the slower option to a claimant. The blind retry loop is the one silently burning money at scale, three calls, every time, for photos that were never going to pass.
Names which cost is hidden and which is visible, and optimizes against the hidden one.
K
Kill criteria. What would change the call.
If identical-resubmission success on a failure category ever clears about two percent, that category is worth auto-retrying after all. Below that, it's just cost.
Shows the position is evidence-based, not stubborn.
Four conditions. If none of them are true, the retry was never going to help.
The recap, one line per letter: position is that only new information makes a retry real, impact is a tap versus a silent inference bill, cost asymmetry is the blind retry being the expensive one despite looking automatic and free, and kill criteria is watching for any failure category where resubmission genuinely does help.
Four questions, asked once per failure, instead of one blind loop repeated three times.
And if you want to be sure it really works, try it somewhere elseSame four letters, a farm's crop-disease scanner instead of a storm-damage photo. Same wasted loop, a different field entirely.
Thornbank Agritech makes a phone app that scans a photo of a crop leaf and flags disease risk. Anouk Ferreira runs a mid-sized vineyard and scans suspicious leaves most mornings before the day heats up. Her app, like Fenwick's, used to auto-retry any failed scan up to three times against the same photo.
Mapped onto PICK: position is that a blurry or backlit leaf photo will fail identically no matter how many times it's resent, so the fix is a recapture prompt, not a resend; impact is Anouk losing a few seconds to retake a shot versus Thornbank quietly paying for three scans of a photo that could never pass; cost asymmetry is the same one, the automatic-feeling option is the expensive one, not the one that asks the farmer to do something.
Same shape as Fenwick's driveway, a vineyard instead of a windshield.
Identical-resubmission success rate, by failure reason
Only network timeouts clear the kill line. Every photo-quality failure sits far below it, exactly where the retry should stop.
Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "a retry only helps if new information enters between attempts, otherwise you're paying twice for the same answer," and stop.
Cost: there's no time to rebuild the retry logic this sprint. Say so honestly, and start by turning off blind retries for the failure categories already proven not to improve on resubmission, that alone stops most of the waste.
The model gets better, for real: if the underlying model's accuracy improves and fewer photos fail at all, that's still not a reason to bring back blind retries on the ones that still do fail, since an improving model doesn't change whether the same input produces the same answer.
Where people run it wrong.
They treat "add a retry" as a free safety net, without ever checking whether the failure could possibly resolve differently.
They measure retry success by whether the claim eventually gets handled, not by whether the retry itself changed anything.
They let a genuinely helpful case, like network drops, justify auto-retrying everything else by association.
How to use it live. When someone asks you this question, ask yourself one thing first: between this attempt and the next one, does anything actually change? If the honest answer is no, it isn't a retry, it's a repeat.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework fits a "when does a retry actually help" question?
Tap to flip
ANSWER
PICK: position, impact, cost asymmetry, kill criteria. Commit to a rule, then show which side is actually expensive.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Teodor Vance, a Fenwick Mutual field inspector who has assessed storm damage for four years, tablet in hand, in customers' driveways.
3 · THE POSITION
What's the one-sentence rule for whether a retry helps?
Tap to flip
ANSWER
A retry helps only when new information enters between attempts. The same input to the same model will not produce a meaningfully different answer.
4 · THE COST ASYMMETRY
Which option looks slower, and which one is actually expensive?
Tap to flip
ANSWER
The recapture prompt looks slower to the claimant, one extra tap. The blind retry loop is the expensive one, three inference calls that were never going to pass.
5 · THE OLD DECISION
What decision would you take back?
Tap to flip
ANSWER
Merging "detect a failure" and "retry automatically" into one step, skipping the check for why it failed, which made sense back when most failures really were transient.
6 · THE NUMBER
Fill in the blank: a blind retry costs about 18 cents and ___ extra seconds, versus 6 cents and 3 seconds for diagnose-then-act.
Tap to flip
ANSWER
9 extra seconds. Three times the cost and three times the delay, for a photo that ends up with a human either way.
7 · THE KILL CRITERIA
What evidence would change this answer's position?
Tap to flip
ANSWER
A failure category showing identical-resubmission success above roughly two percent. Only network timeouts clear that line in this story.
8 · CROSS PRODUCT TRANSFER
Section 4 answers this again for a different product. Which product, and what's the position there?
Tap to flip
ANSWER
Thornbank Agritech's crop-disease scanner. Same position: a recapture prompt beats a blind resend, since a blurry leaf photo fails identically every time it's resubmitted.
Check yourself Score: 0 / 0
Short answer, name the reversal
1. What old design decision does this answer take back, and why did it make sense when it was made?
Show hint
Look at "the decision I would take back."
Show answer
Model answer: Merging failure detection and automatic retry into one step, with no check for why the failure happened. It made sense while most failures were transient network drops, a genuine case where retrying helped.
Multiple choice
2. Which of these failures is the strongest candidate for an automatic, blind retry?
A. A photo with heavy glare hiding the crack pattern.
B. A genuinely ambiguous damage pattern the model can't confidently classify.
C. A request that failed because the network connection dropped mid-upload.
D. A blurry photo caused by the inspector's hand shaking.
Show hint
Ask which failure is a property of the input versus a property of the connection.
Show answer
C. A network drop isn't a property of the photo. The same upload really can succeed the second time, unlike glare, ambiguity, or blur, all baked into the same pixels every time.
True or false
3. True or false: resubmitting the exact same photo to the same model will usually change a low-confidence score into a confident one.
True
False
Show hint
Look at the knowledge spark about identical inputs.
Show answer
False. The same input tends to produce the same, or nearly the same, output. Small randomness rarely turns a genuinely unclear photo into a clear one.
Fill in the blank
4. Fill in the blank: in the audit that found this problem, nearly ___ of all model calls were identical resubmissions of photos that had already failed once.
Show hint
Look at the story's audit paragraph.
Show answer
A third. That's a large share of total inference spend going toward retries that could never have changed the outcome.
Short answer, apply it yourself
5. Think of an app that made you tap "try again" on something that was never going to work differently the second time. What should it have asked you instead?
Show hint
Think of a failed upload, a rejected document scan, or a search with no results.
Show answer
Model answer: Most people can recall a "try again" button on a failure that was really about the input itself, a bad photo, a wrong file type, where the honest next step was a different action, not a repeat.
Short answer, where it wouldn't matter
6. Name a failure type in Fenwick's tool where an automatic retry genuinely is the right call.
Show hint
Look at "what I would leave alone."
Show answer
Model answer: Network timeouts and dropped connections. Those are genuinely transient, unrelated to the photo itself, so the same upload really can succeed on a second try.
Before you close the answer
Why this works
Tests whether you understand that a model's output depends on its input, not just whether you know the word "retry."
Follow-up traps
"What if the model has some randomness built in, so retries do sometimes succeed?" Response: track that success rate directly. If it clears the kill-criteria threshold for a given failure type, that type earns an automatic retry. Below it, the randomness is noise, not a real second chance.
"Doesn't blocking retries just mean more claims land on a human faster, overwhelming the review team?" Response: those claims were always going to land on a human, the blind retries only delayed that handoff by nine seconds and charged for it. Blocking them speeds up the handoff, it doesn't create new review work.
If pressed
Fenwick's actual fix logs a per-failure-category resubmission-success rate on a rolling basis, and only failure categories above the two percent line stay on the automatic-retry list, everything else is reviewed quarterly instead of assumed permanently safe to retry.
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.