CaseAdvancedResponsible AI & Advanced Practice / Agent product management specifics / #8
What should an agent do when it is uncertain mid-task?
SPARK the product is the order-exception agent at Fennimore Home, an online furniture and home goods retailer
Fennimore Home ships about 600 furniture and home-goods orders a day. Some get flagged: a bad address, an out-of-stock swap, delivery notes that don't quite line up. Marisol Guzman has run that exceptions desk for five years, and keeps a tablet mounted at the desk where the flagged queue lives.
The direct answer
When the agent hits a gap where a wrong guess would be expensive or hard to undo, it should stop and ask one narrow question instead of guessing or freezing the whole task. When the gap is small and cheap to fix either way, it should use a safe default and keep moving. The two get treated differently on purpose.
Do this, in order
Pause and ask one narrow question before a costly or hard-to-undo step, instead of guessing.Why: a wrong guess there costs far more than the few seconds a question costs.
Use a safe default and keep going when the gap is small and cheap to fix.Why: stopping for every tiny gap trains people to stop reading the agent's questions at all.
Separate "decide" from "act" so a pause has somewhere to land.Why: without a gap between the two, there is no place to put a question, no matter how sure you are you'll need one.
Ask about one field, not the whole order.Why: a narrow question gets answered in seconds; an open-ended one gets ignored or answered wrong.
Log every guess it made, not only the ones it flagged.Why: someone has to be able to find the quiet wrong ones later, not just the loud ones.
Recheck what counts as "costly" every quarter, not once at launch.Why: the mix of order types shifts, and what was rare and cheap can quietly become common and expensive.
How to answer this, stage by stage
Nobody is grading whether you can list every edge case. They're grading whether you can name the one design decision that catches the expensive ones.
Stage 1
Scope it to one real agent
Say it like this
"I'll answer this for an order-exception agent at a furniture retailer, the kind that decides how to fix a flagged order without a person touching it first."
Why this works
Turns an abstract question about "uncertainty" into one concrete moment you can design for.
Stage 2
Say your structure out loud
Say it like this
"I'll use SPARK. Situation, the job before any agent. Payoff, the habit I want it to build. Anchor, the one design decision. Risk, what breaks when the anchor is wrong. Keep out, what I won't build yet."
Why this works
Tells the interviewer you have a plan before you say a single detail.
Stage 3
Reframe the question
Say it like this
"This isn't really 'how do we make the model less uncertain.' It'll never hit zero. It's 'what happens in the ten seconds after it notices it doesn't know.'"
Why this works
Moves the conversation off model quality and onto the actual design decision being tested.
Stage 4
Give the one decision
Say it like this
"Split deciding from acting. If the next step is cheap to undo, use a safe default and log it. If it's costly or hard to undo, pause and ask one narrow question before it acts."
Why this works
This is the anchor. Concrete enough that someone could build it tomorrow.
Stage 5
Prove it with a failure
Say it like this
"Without that gate, we had a $340 order where the agent guessed 'reship everything' when the customer meant 'just credit the missing part.' Nobody caught it for weeks, because the log just said resolved."
Why this works
Shows the cost of skipping the anchor in one breath, not a slide of bullet points.
Stage 6
Say what you'd measure
Say it like this
"I'd track the wrong-guess rate on the flagged, costly subset specifically, not the overall accuracy number, since that's the one that hides the expensive mistakes."
Why this works
Shows you'd know if the design were failing before a customer told you.
Stage 7
Close on the one line
Say it like this
"Ask before it's expensive to be wrong. Default and log when it isn't. That's the whole rule."
Why this works
Leaves the interviewer with the decision, not a summary of the story.
Let's learn
Before Fennimore Home built anything, Marisol's team of four clerks handled every flagged order by hand: a bad address, a delivery conflict, an item that had gone out of stock. About 40 a day, six minutes each, close to four hours of combined work every single day.
Fennimore Home's order-exception agent looks at any order that gets flagged and decides how to fix it, without a clerk touching it first: rebook the shipping, swap the item, or issue a partial refund.
Knowledge spark: what makes a step "hard to undo"?
Some fixes can be reversed with one more email. A refund can't be un-sent once the money moves, and a reshipped order can't be un-shipped. Those are the ones worth pausing for. A typo'd zip code fixed the wrong way just gets fixed again.
The agent now resolves all 40 a day in under ten seconds each. About 34 of those are unambiguous, an obvious zip-code fix, a clear backorder swap, and the agent nails nearly all of them. The other six are genuinely unclear: conflicting delivery notes, two possible substitute items, a customer message that names two different addresses.
Wrong resolutions per week on ambiguous orders
The pause-and-ask design cuts wrong resolutions on ambiguous orders by more than 90%, for a subset that's about 15% of the daily queue.
We did not lose accuracy on the easy ninety percent of orders. We lost the six a day nobody was watching.
At its worst: a $340 order where the customer's note actually meant "just credit me for the missing drawer pull," and the agent guessed "reship the whole cabinet" instead. Nobody caught it. It only surfaced weeks later, when a teammate reopened that exact ticket for an unrelated reason.
The decision I would take back
We built "decide" and "act" as a single step with no pause in between, since it made the system simpler to ship and the model was right most of the time anyway. That was fine while the ambiguous share of orders was small and the fixes were cheap. It stopped being fine once a wrong guess meant real money moving before anyone could look.
What I would leave alone: the 34 unambiguous fixes a day. Adding a pause there would just recreate the manual busywork the agent was built to remove.
The lesson: an agent that's usually right doesn't need a smarter guess. It needs a place to say "I don't know" before it acts, not a smarter way to hide that it didn't.
Now here is the same thing as a story
The short version above is what you'd say defending this design to Fennimore's fulfillment leadership. Read this one for how the gap actually got found.
Marisol Guzman has run Fennimore Home's exceptions desk for five years. She can tell a genuine delivery conflict from a customer's typo in about two seconds, just from the shape of the note.
Before any agent, every flagged order went through the same four steps, by hand, six minutes at a time.
For the first months after launch, the agent was the best part of Marisol's day. She'd open five or six resolved tickets at random each morning, just to see how it was doing. It was always fine.
So she went to two spot-checks a week. Then none. The log only ever said "resolved," never how sure the agent had been, so there was nothing left to check against.
Nothing dramatic happened at week 4 or week 7. The wrong-guess rate on ambiguous orders just kept climbing, quietly, with nobody watching.
Wrong-guess rate on ambiguous orders, week 1 to week 9
Nobody was tracking this line, because the log had no field for "how sure was the agent."
Then a colleague reopened that $340 ticket, not because anything had flagged it, but because a customer called back about something unrelated and the old note happened to be attached to the same thread.
Same order, same ambiguous note. One version guesses and ships wrong. The other pauses for one question and ships right.
Marisol spent two full days pulling every "reship" decision from the past two weeks, 84 orders, and rechecking each one by hand. That was the visible cost. The real cost was nine weeks of quiet drift nobody had a number for.
It was never really about the model getting worse. The agent's overall accuracy barely moved. Marisol's trust had exactly two settings: "the log is fine" and "recheck everything," and there was nothing in between for the log to earn back on its own.
The engineers who built the agent had merged decide-and-act into a single step in the very first design meeting, because it shipped faster and the model was right almost every time back then. Nobody had asked for a pause. It looked like one extra hop for no reason.
The redesign didn't make the model smarter. It gave the gap somewhere to land before the order shipped.
Replayed with the fix in place: the same $340 note comes in, the conflicting instructions trip the "costly and unclear" gate, and the agent sends one line back through the existing order-update channel: "Just to confirm: credit the missing part, or reship the full order?" The answer comes back in about 40 minutes. The right fix goes out. No loss, no two-day recheck.
Only the last one, a costly next step, is what actually earns a pause. The other three can usually take a safe default.
The old design handed Marisol a dial with no markings on it: trust it a little, trust it a lot, no way to tell which. The new one hands her a switch that already tells her which six orders a day are worth a second look.
I built the fast path because it shipped faster and the model was right almost every time. It took a colleague stumbling onto a stale ticket, something that could just as easily never have surfaced at all, to see that "right most of the time" was never the same promise as "safe to never check."
SPARK, one design decision at a timeNot a policy essay. SPARK is what forces the uncertainty question down to one buildable anchor.
S
Situation. The job before any agent.
Four clerks hand-resolving 40 flagged orders a day, six minutes each, entirely on paper judgment.
Grounds the anchor in a real workflow that already existed.
P
Payoff. The habit worth building.
Clerks stop hand-resolving obvious swaps, but never stop being the one who signs off on money-moving or hard-to-reverse fixes.
Names the habit the design is actually shipping, not just the time saved.
A
Anchor. Split decide from act.
Before a costly or hard-to-undo step, pause and ask one narrow question. Otherwise, use a safe default and log the guess.
The single decision the whole answer hangs on.
R
Risk. What breaks when it's wrong anyway.
Even after asking, a customer can answer wrong. Since a person answered, the agent has real cover, and the disagreement gets logged for review, not silently absorbed.
Proves the anchor survives its own worst case.
K
Keep out. Not on day one.
No full confidence dashboard, no auto-retry loop, no tone or sentiment reading. Just the one gate on cost and reversibility.
Shows judgment instead of a wish list bolted onto the launch.
Keeping these out of day one is what makes the anchor buildable in a sprint instead of a quarter.
The recap, one line per letter: situation is the four-clerk manual desk, payoff is clerks staying out of the obvious swaps but in on the costly ones, anchor is the pause-before-costly rule, risk is a wrong answer still getting logged for review, and keep out is everything fancier than that one gate.
And if you want to be sure it really works, try it somewhere elseSame anchor, a veterinary clinic instead of a furniture warehouse. Completely different stakes, same one rule.
Dr. Priya Nair runs a scheduling agent at her veterinary clinic that reschedules or escalates same-day appointment requests sent in by pet owners through a booking app. Mapped onto SPARK: the situation is a front-desk assistant juggling a paper schedule and a ringing phone. The payoff is the assistant no longer touching routine reschedules, a nail trim moved a day, a check-up bumped an hour. The anchor is the same rule: bumping a routine visit is cheap and reversible, so the agent just does it and logs it; cancelling a chemo follow-up or a pre-surgery fast is costly and hard to undo, so the agent pauses and sends one question to the front desk first. The risk: if the front desk answers wrong under pressure, the exchange is still logged, so the clinic can see the miss instead of it vanishing into a calendar change. Keep out: no auto-rebooking across multiple pets, no waitlist prediction, not on day one.
A different clinic, a different calendar, and the exact same fork in the road: routine gets a default, irreversible gets a question.
Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "pause before it's expensive to be wrong, default when it isn't," and stop there.
Cost: there's no budget this quarter for a fancy confirmation UI. Send the one question through whatever channel already exists, a text thread, an order-update email, since even a plain-text question beats a silent guess.
The model gets better, for real: if overall accuracy improves, that's still not a reason to drop the gate on the costly subset. A better average model can still guess wrong on the one case that costs $340 to get wrong.
Where people run it wrong.
They build one blanket confirmation step for every flagged case, which trains people to stop reading the agent's questions at all.
They wait for a confidence score to look scientific instead of asking the plainer question: is this cheap or costly to undo.
They log only what the agent flagged as uncertain, not every guess, so the quiet wrong ones on the "confident" side stay invisible.
How to use it live. When someone asks what an agent should do when it's unsure, answer the question behind the question first: unsure never fully goes away, so ask what happens in the next ten seconds. Then give the one rule: cheap and reversible, default. Costly or hard to undo, ask.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework fits "what should an agent do when it's uncertain mid-task," and what's its one job?
Tap to flip
ANSWER
SPARK: situation, payoff, anchor, risk, keep out. Its job is to design against the failure before you build, not describe the model's confidence.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Marisol Guzman, who has run Fennimore Home's order-exceptions desk for five years and can spot a real conflict in a customer's note in about two seconds.
3 · THE HABIT
What did Marisol stop doing because the agent seemed to work?
Tap to flip
ANSWER
She stopped spot-checking resolved tickets, from five or six a morning down to zero, since the log only ever said "resolved" with nothing to check it against.
4 · THE SWITCH
What's the two-setting switch in this story?
Tap to flip
ANSWER
Trusting the log completely, or rechecking everything by hand. Once the near miss hit, there was no middle setting left to fall back on.
5 · THE OLD DECISION
What decision would you take back?
Tap to flip
ANSWER
Merging "decide" and "act" into one atomic step at launch, so there was never a place to insert a pause, even after guesses started getting shaky.
6 · THE NUMBER
Fill in the blank: the wrong-guess rate on ambiguous orders climbed from ___% to ___% over nine weeks.
Tap to flip
ANSWER
18% to 35%. Nobody had a number to watch, since the log never recorded how sure the agent had been.
7 · THE REPLAY
Same $340 order, redesigned agent. What changes?
Tap to flip
ANSWER
The conflicting note trips the costly-and-unclear gate, one question goes out, the answer comes back in about 40 minutes, and the right fix ships. No loss, no two-day recheck.
8 · CROSS PRODUCT TRANSFER
Section 4 answers this again for a different product. Which product, and what's the anchor there?
Tap to flip
ANSWER
Dr. Priya Nair's veterinary scheduling agent. Same anchor: default on a routine reschedule, pause and ask before cancelling something like a chemo follow-up.
Check yourself Score: 0 / 0
Multiple choice
1. Why does the agent pause and ask before a costly or hard-to-undo step, instead of just guessing like it does elsewhere?
A. Because the model can't produce a guess for those cases.
B. Because a wrong guess there is expensive and can't easily be reversed, unlike the cheap, reversible cases.
C. Because customers always prefer being asked a question.
D. Because it's required for every flagged order by policy.
Show hint
Look at the anchor in Stage 4 of the walkthrough.
Show answer
B. Cost and reversibility, not raw confidence, are what decide whether the agent asks or defaults.
True or false
2. True or false: this answer recommends adding a confirmation question to all 40 flagged orders a day.
True
False
Show hint
Look at "what I would leave alone."
Show answer
False. Only the roughly six costly, ambiguous orders a day get a question. The other 34 keep their instant safe-default resolution.
Fill in the blank
3. Fill in the blank: the near miss involved an order worth $___.
Show hint
Look at "what that costs, at its worst" in Section 1.
Show answer
340. The agent guessed "reship the whole cabinet" when the note meant "credit the missing drawer pull."
Short answer, where it wouldn't matter
4. Name a kind of flagged order where pausing to ask would NOT make sense.
Show hint
Look at "what I would leave alone."
Show answer
Model answer: An obvious zip-code typo. It's cheap to fix either way and doesn't need a question before the agent acts.
Short answer, apply it yourself
5. Pick a product you use yourself. What's one habit it built in you that you'd stop doing if it got a little worse?
Show hint
Think of something you stopped double-checking because it was reliable for a while.
Show answer
Model answer: Many people stop proofreading autocomplete text messages after months of it being right, then a garbled one goes out before they notice.
Short answer, the number
6. If the ambiguous share of orders doubled from 6 a day to 12 a day, would the same one-question design still hold up? Why or why not?
Show hint
Think about what the design actually depends on: the count of costly cases, or the fraction of the day they take.
Show answer
Model answer: Mostly yes. Twelve extra questions a day, answered in tens of minutes each, still leaves the 34-plus unambiguous fixes untouched. It would only strain if the answer channel itself got slow.
Before you close the answer
Why this works
Tests whether you reach for a smarter model when the real gap is architectural: a place for the agent to pause before it acts, not a better guess dressed up as confidence.
Follow-up traps
"Why not just make the model more confident before shipping?" Response: confidence never reaches 100%, so you still need a policy for whatever's left; a pause step is cheaper than chasing the last few points of accuracy forever.
"Doesn't asking a question slow the agent down too much?" Response: only for the roughly 15% of exceptions that are genuinely ambiguous. The other 85% keep their ten-second resolution, so overall throughput barely moves.
If pressed
The actual trigger isn't one confidence score. It's two flags combined: would reversing this cost more than about fifty dollars or a phone call, and does the model's own certainty sit in the middle third of its range. That combination is what keeps the gate from firing on merely rare inputs, only on rare-and-costly ones.
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.
