ConceptAdvancedDesigning for Uncertainty & Trust / Onboarding users to probabilistic products / #18
What is the right amount of friction in AI onboarding?
PICK the product is Callmark, an AI sales assistant that drafts follow-up emails and scores deal risk after every call
Callmark listens to sales calls, drafts a follow-up email, and scores how risky the deal looks. Dorian Kessler is an account executive who closes mid-market deals, and Callmark can be set to send those drafted follow-ups on its own, with nobody reading them first.
The direct answer
There's no single right amount of friction across an entire onboarding. Put almost none on anything reversible, previewing a score, reading a draft, editing a line. Put real, deliberate friction on the one or two decisions that are genuinely hard to undo, like letting the AI send something to a client with nobody reading it first. Spreading friction evenly is the actual mistake, not the amount of it.
Do this, in order
Concentrate real friction on irreversible, high-stakes actions, and remove it everywhere else.Why: uneven friction protects the moment that actually matters instead of taxing every moment equally.
Make the one hard confirmation specific and legible, not a generic "are you sure" dialog.Why: a vague warning gets clicked through from habit; a specific one names what could actually go wrong.
Measure setup completion and time-to-first-value for the low-friction paths.Why: this is where added friction has a real, visible cost in lost adoption.
Measure client-facing incidents for the high-friction path.Why: this is where removed friction has a real, hidden cost that a setup-completion number would never show.
Watch for warning fatigue if the confirmation gets used on too many actions.Why: the friction only works if it stays rare enough to still mean something.
How to answer this, stage by stage
Nobody's grading whether you picked "more friction" or "less friction." They're grading whether you can find the one decision friction is actually protecting.
Stage 1
Scope it to one real decision
Say it like this
"I'll answer this for Callmark's onboarding, specifically the moment where a rep can turn on auto-send for drafted follow-up emails."
Why this works
Turns a vague question about "friction" into one concrete design decision.
Stage 2
Say your structure out loud
Say it like this
"I'll use PICK. Position, my pick up front. Impact, who feels each kind of error. Cost asymmetry, which one is hidden. Kill criteria, what would change my mind."
Why this works
Signals you're about to commit to a position, not hedge for two minutes.
Take a position, before any reasoning
Say it like this
"Almost no friction on anything reversible. Real friction on the one thing that isn't: letting the AI send something to a client with nobody reading it."
Why this works
Commits to an answer immediately, which is exactly what a tradeoff question is testing.
Stage 4
Name who feels each error
Say it like this
"Too much friction costs Dorian time and costs Callmark a rep who quits onboarding halfway through. Too little friction, on the one risky action, costs Dorian an actual client relationship."
Why this works
Puts a person on both sides of the tradeoff instead of comparing two abstract numbers.
Stage 5
Find the asymmetry
Say it like this
"One extra click on a preview is cheap and you feel it immediately. One skipped confirmation on auto-send is hidden until a client complains, and by then the damage is already done."
Why this works
This is the heart of PICK: the two errors are not the same size, and saying so is the actual insight.
Stage 6
Name the rejected alternative
Say it like this
"I considered a single confirmation style used everywhere, so every action gets the same 'are you sure' dialog. I rejected it, because people learn to click through a dialog that shows up constantly."
Why this works
Proves this was a real choice between real designs, not the only idea that came to mind.
Stage 7
Name the kill criteria, and close
Say it like this
"If we saw reps under-preparing for calls because a preview-only risk score felt like a guarantee, I'd add friction there too. Until then, friction stays concentrated on the one action you can't take back."
Why this works
Shows the position isn't stubborn; it would change given real evidence.
Let's learn
Say we build a tool that drafts a follow-up email after every sales call, and can be set to send those drafts on its own.
Onboarding Dorian to Callmark could look like five confirmation screens for every single action, previewing a score, editing a line, opening the call summary, each one asking "are you sure?" Or it could look like zero friction anywhere, including the one setting that lets an unread email reach a real client.
Knowledge spark: what's warning fatigue?
When a warning shows up so often that people stop reading it and start clicking through automatically. The warning still exists. It just stopped doing its job.
Neither extreme is really about "the right amount." Both are about putting friction in the wrong place.
Every step up to the last one is reversible. The last one is the only step that actually needs a real pause.
Setup completion rate, by number of confirmation steps
Uniform friction applied to every step is the fastest way to lose almost half your new reps before they ever see a benefit.
At its worst: a rep abandons setup entirely because six identical "are you sure" dialogs stood between them and the first useful screen, and Callmark never gets a chance to prove itself.
The decision that mattered
Put almost no friction on anything reversible. Put real friction, a specific one, not a generic dialog, on the single decision that can reach a real client without anyone reading it first.
What I would leave alone: the risk score preview itself needs zero friction, no confirmation, no extra click. Reading a number costs nothing and can't hurt anyone; it's a place friction would only slow people down for no protection at all.
The extra click was never the expensive part. The client who got a tone-deaf, unreviewed email was.
The lesson: "how much friction" is the wrong question shape. The real question is "which one decision, out of all of them, actually needs a person to stop and think," and everything else should get out of the way.
Now here is the same thing as a story
The short version above is what you'd say defending this design in a product review. Read this one for how the choice actually got tested.
Callmark can be set to send its drafted follow-up on its own, no review, right after a call ends. Dorian turned that setting on in his second week, once a dozen drafts in a row had read exactly the way he would have written them himself.
One of these costs shows up the instant it happens. The other one waits, quietly, until a client mentions it on a call three weeks later.
Two months in, Callmark drafted a follow-up after a tense renewal call, one where the client had raised a real concern about pricing. The draft opened with a cheerful line about "exciting next steps." It went out unread, because auto-send had nothing standing between a draft and a client's inbox.
Real friction doesn't mean more clicks everywhere. It means one specific moment gets built with actual care.
The client didn't complain to Callmark. They complained to Dorian, on the next call, in a tone that made clear the tool had cost him something he couldn't easily get back.
Two of these four actions live in the corner that actually deserves a pause. The other two never did.
The old design treated every action in Callmark's settings the same way: light, fast, no friction anywhere, because friction had always been the enemy of a good onboarding number. The redesign keeps that everywhere except one toggle, which now shows the actual draft, a plain sentence naming what a bad send could cost, and asks for a typed word instead of a click.
Two green items, two orange ones. The whole design decision fits on one short list.
We built the fast, frictionless version everywhere because a good onboarding-completion number was easy to point at in a review, and friction always looks like the thing standing between a rep and their first good day. It took one client noticing before we did to see that removing friction everywhere had quietly removed it from the one place it actually mattered.
PICK, in one screenNot a dial to turn up or down. A single decision, isolated from all the ones around it.
P
Position. The pick, first.
Almost no friction on anything reversible. Real friction on the one decision that isn't: unreviewed auto-send.
The hardest step, and the one that actually answers the question.
I
Impact. Who feels each error.
Too much friction costs Dorian time and costs the company a rep who abandons setup. Too little friction, on the one risky action, costs Dorian a client relationship.
Names a person on both sides, not two abstract percentages.
C
Cost asymmetry.
An extra click is cheap and visible immediately. A bad email reaching a client is hidden until the relationship is already damaged.
The actual insight: the two errors were never the same size.
K
Kill criteria.
If reps started under-preparing because a preview-only risk score read like a guarantee, that would justify adding friction there too.
Shows the position would actually change given real evidence, not just repeated on principle.
Cost per incident: setup abandonment vs a bad auto-sent email
These aren't remotely the same size. Weighing them equally is exactly how friction ends up in the wrong place.
The recap, one line per letter: position is almost-no-friction with one real exception, impact is a rep's time against a client relationship, cost asymmetry is the visible click against the hidden email, and kill criteria is real evidence of preview-only harm.
And if you want to be sure it really works, try it somewhere elseSame four letters, a peer-to-peer car-sharing marketplace instead of a sales tool. A different domain, and the friction moves to an approval decision instead of a send decision.
Haulbright is a marketplace where people rent trucks and vans from their neighbors. An AI feature can auto-approve a booking request based on a guest's trust score, with no host review at all. Marcus Oyelaran hosts three vehicles and turned auto-approve on to stop reviewing every request by hand.
Mapped onto PICK: position is almost no friction on browsing, messaging, or reviewing a guest's profile, and real friction specifically on the auto-approve toggle for first-time guests or anyone with a damage claim on file. Impact is Marcus losing hours to manual approvals if friction is everywhere, versus Marcus handing his van to someone he'd never have approved himself if friction is nowhere. Cost asymmetry is a manual review costing a few minutes, visible immediately, against a bad approval costing a damaged vehicle, discovered days later. Kill criteria is real data showing trust scores predict damage poorly for first-time guests specifically, which would justify a review step for every first-time booking, not just borderline ones.
Three of the four branches still land on a human confirming. The friction isn't evenly spread, it's concentrated exactly where trust hasn't been earned yet.
Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "friction goes on the one irreversible decision, not everywhere," and stop.
Cost: there's no engineering time this quarter to build a custom confirmation screen. Say so, and start with a plain text warning on the toggle itself, cheaper to build, still concentrated in the right place.
The model gets better, for real: even if Callmark's draft quality keeps improving, the one unreviewed send still deserves friction, since a rare miss on a genuinely better model is still a miss nobody caught.
Where people run it wrong.
They apply the same confirmation dialog to every action, so the one that matters gets clicked through out of habit.
They remove friction everywhere to chase a clean onboarding-completion number, without checking which action removed friction from.
They treat "friction" as a single dial instead of a decision to be made action by action.
How to use it live. When someone asks about the right amount of friction, ask yourself first: which one action here can't be undone? Put your answer there, and take it off everything else.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework fits "what's the right amount of friction in AI onboarding"?
Tap to flip
ANSWER
PICK: position, impact, cost asymmetry, kill criteria. It's a tradeoff question, so it needs a committed pick before any reasoning.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Dorian Kessler, an account executive who turned on Callmark's auto-send after a run of good drafts.
3 · THE POSITION
What's the actual pick, stated as one sentence?
Tap to flip
ANSWER
Almost no friction on anything reversible. Real, specific friction on the one decision that can't be undone: sending something to a client unread.
4 · THE ASYMMETRY
What's the cost asymmetry in this answer?
Tap to flip
ANSWER
An extra click is cheap and felt immediately. A bad auto-sent email is hidden until a client brings it up, by which point the damage is already done.
5 · THE REJECTED OPTION
What alternative design was considered and rejected?
Tap to flip
ANSWER
A single uniform confirmation dialog on every action. Rejected because people learn to click through a warning that shows up constantly, so it stops protecting anything.
6 · THE NUMBER
Fill in the blank: setup completion fell to ___ percent once onboarding had six confirmation steps.
Tap to flip
ANSWER
46 percent, down from 91 percent with just one confirmation step.
7 · THE KILL CRITERIA
What evidence would flip this pick?
Tap to flip
ANSWER
Real data showing reps under-prepare for calls because a preview-only risk score reads like a guarantee. That would justify adding friction to the preview too.
8 · CROSS PRODUCT TRANSFER
Section 4 answers this again for a different product. Which product, and what's the one irreversible action there?
Tap to flip
ANSWER
Haulbright, a truck-sharing marketplace. There, the one action needing real friction is auto-approving a first-time or high-risk guest with no host review.
Check yourself Score: 0 / 0
Multiple choice
1. Why does this answer put friction on auto-send but not on previewing a risk score?
A. Because risk scores are always accurate.
B. Because previewing a score takes longer than sending an email.
C. Because auto-send is irreversible and can reach a real client, while previewing a score can't hurt anyone.
D. Because auto-send is a newer feature.
Show hint
Look at the quadrant sorting actions by stakes and reversibility.
Show answer
C. Friction goes where an action is both high stakes and hard to undo, not where an action merely feels important.
True or false
2. True or false: this answer recommends removing friction from every part of onboarding.
True
False
Show hint
Look at "the decision that mattered."
Show answer
False. Friction stays real and deliberate on the one irreversible action. It's only removed from the reversible ones.
Fill in the blank
3. Fill in the blank: an abandoned setup was estimated to cost about $40, versus roughly $___ for a bad auto-sent email's renewal risk.
Show hint
Look at the grouped bar chart comparing the two costs.
Show answer
12,000. The two costs aren't remotely comparable, which is exactly the asymmetry the whole answer turns on.
Short answer, name the position
4. State this answer's position in one sentence, the way you'd say it in the first ten seconds of an interview.
Show hint
Look at the direct answer at the top of the page.
Show answer
Model answer: Almost no friction on reversible actions, and real, specific friction on the one action that can't be undone.
Short answer, where it wouldn't matter
5. Name one action in Callmark where adding friction would be pure cost with no protective value.
Show hint
Look at "what I would leave alone."
Show answer
Model answer: Previewing the risk score. Reading a number can't hurt anyone, so a confirmation step there only slows people down.
Short answer, apply it yourself
6. Pick an app you use yourself. Which one setting in it deserves real friction, and which settings around it don't?
Show hint
Think about a setting that can send, delete, or share something on your behalf without a review step.
Show answer
Model answer: Many photo apps let you turn on auto-sharing to a public album. That one toggle deserves a real pause; browsing or editing a photo beforehand doesn't need any friction at all.
Before you close the answer
Why this works
Tests whether you'll commit to an uneven answer instead of a comfortable "it depends," and whether you can find the one decision that actually deserves friction instead of treating friction as a single dial to turn.
Follow-up traps
"Couldn't a rep just turn off auto-send if they don't trust it?" Response: yes, but the failure mode here is a rep who trusts it too much after a good run of drafts, which is exactly when the one hard confirmation matters most.
"Isn't concentrating friction on one action risky if you picked the wrong action?" Response: that's what the kill criteria step is for; if evidence showed a different action was actually the risky one, the friction moves, it doesn't stay stuck on principle.
If pressed
Callmark's real confirmation requires typing the word "send" rather than clicking a button, specifically because a typed confirmation measurably reduces the click-through-without-reading rate compared to any button, even a red one.
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.
