ConceptIntermediateAI Opportunity & Model Strategy / Opportunity identification for AI / #13
What is the risk of picking your first AI feature based on what is easiest to build?
GUARD · the number nobody wrote down next to the one everybody did
Greetwell is Strathmore Systems' tool that builds a short, personalized welcome video for every new customer another business signs up. Iskandar Millbrooke, the product manager who owns Greetwell's roadmap, had to pick the first feature to ship after launch. Clemency Roscommon, Strathmore's VP of Product, wanted something live before a board meeting. Bertille Bouchard, who runs Bouchard Fabricare across four locations, just wanted a welcome video that knew which plan she'd actually bought.
The direct answer
Never let build time decide which AI feature ships first by default. Make every candidate feature write down its real user value case, on paper, completely apart from its build estimate, and rank by that case first. Skip this step and the feature that ships is just the one that was cheapest to demo, and the AI initiative spends its first real shot at proving itself on something almost nobody finishes.
Do this, in order
Pick the first AI feature by its real user value case, not by which one takes fewer weeks to build.Why: this is the direct answer, and every bullet below just protects it.
Make every candidate feature write its value case down, on paper, apart from its build estimate, before anyone ranks them.Why: without a written value case, build time is the only number in the room, so it decides by default.
Name who gets hurt if the easy one wins: the customers whose real situation was harder to build for, not easier.Why: a customer running four locations is invisible if the only question anyone asks is whether an AI feature shipped at all.
Check usage on the shipped feature honestly, inside the first month, broken out by customer type.Why: a flat blended number can hide two very different numbers sitting underneath it.
Leave the fast, cheap option alone when its value case is genuinely just as strong.Why: gating every quick win to look thorough just teaches the team the process is theater.
When the numbers show ease won the argument, rebuild around the real value case and say so out loud.Why: a quiet fix teaches the next roadmap review nothing at all.
How to answer this, stage by stage
Nobody is grading whether the generic video was a bad idea. They're grading whether you can turn "the easiest one might win by default" into a real, defensible fix.
1
Scope it to one product and one first-feature decision
Say it like this
"Let me make this concrete. Say a company builds an AI tool that creates a short welcome video for every new customer another business signs up. Right after launch, the team has two ideas for the first feature: a generic video, one script for everyone, and a video that actually names the specific plan the customer bought. That's the scenario I'll run, because 'the risk of picking the easiest one' means nothing until there's a real, more valuable option sitting right next to it."
Why this works
Keeps the answer from turning into generic prioritization advice with an AI label stuck on it.
2
Say your structure out loud
Say it like this
"I'll run this as GUARD. Groups, who's actually carrying the risk here. Unequal, whose situation gets worse specifically because the easy one won. Ability to contest, did anyone push back on 'easy' with a real value comparison, or did it win by default. Reduce, the actual fix to how a first feature gets picked. Detect, how you'd know, from the numbers, that this already happened."
Why this works
Two seconds of structure tells the interviewer you have a method, not just an opinion about which video is nicer.
3
Reframe what the question is actually asking
Say it like this
"This isn't really asking whether the generic video was a bad idea. A simple first feature can be exactly right. It's asking what happens when build time is the only number anyone measured before deciding, so it wins the argument even though nobody ever chose that on purpose."
Why this works
Separates this from generic scope-cutting talk and lands on the real judgment: ease quietly substituting for value.
4
Give the one decision
Say it like this
"Here's what I'd actually do. Before picking, both candidates write down their real user value case, one paragraph, completely apart from the build estimate. I rank by the value case first. Build time only breaks a tie when two value cases are genuinely close, and here they weren't close at all."
Why this works
This matches the direct answer word for word. If it doesn't, the interviewer notices before you do.
5
Prove it with the compressed failure
Say it like this
"Here's what happens without it. A team ships the generic video because it's three weeks faster and a board meeting is five weeks out. It looks fine at launch, an AI feature is live. Ten weeks later, completion is stuck at 23 percent and nobody's asked why. Then one customer tells a support rep she stopped watching at eighteen seconds because the video never once said which of her four locations it was even about. That's when someone finally splits the number by customer type and sees the customers who needed it most are the ones it served worst."
Why this works
This is the story below, compressed to four sentences. The long version proves it actually played out this way.
6
Say what you'd track, and what you'd leave alone
Say it like this
"I'd check completion by customer type every month, not just the average, because a flat 23 percent hid a 54 percent and a 21 percent sitting inside it. And I'd leave the fast option alone anywhere its value case is genuinely just as strong. A one-click button that resends the video is fast to build, and it's still the right call. Speed was never the problem. Speed with no value case standing next to it was."
Why this works
Shows judgment about where the real risk sits, not blanket suspicion of anything quick to build.
7
Close on something checkable
Say it like this
"You'll know it's working when a feature that tests low gets rebuilt around its real value case, out loud, with everyone told why. You'll know it's still broken when a low number just quietly gets filed as 'needs more marketing' and nobody ever reopens the original decision."
Why this works
Ends on a test the interviewer could actually go verify, not a promise that it's handled.
Let's learn
Greetwell builds a short welcome video for every new customer a business signs up. A company plugs Greetwell into its sign-up flow, and within minutes of paying, a new customer gets a video in their inbox: a friendly voice walking them through what to do first.
Four parts. One of them, account data, is the part that decides whether the video knows anything real about the customer watching it.
Before Greetwell shipped its first post-launch feature, Strathmore's roadmap review looked at two ideas side by side. A Welcome Reel: one script, one voice, the same video for every single new customer. And a Plan-Aware Reel: a video that opens by naming the actual plan the customer bought and the first thing to set up on it. The review scored both ideas the way it scored everything else on the list: by how many weeks each one would take. Welcome Reel, three weeks, one engineer. Plan-Aware Reel, nine weeks, two engineers, because it had to pull the customer's real account record and check it before saying anything specific out loud.
Four ideas on the same list. Only one of them, build effort, ever made it into the actual decision.
Knowledge spark: what happens if the account record is stale?
Plan-Aware Reel has to read a customer's real plan and location count the moment it renders. If that record hasn't synced yet, say a customer added a fifth location an hour earlier, the model can state the wrong number with total confidence, because it has no way to know the record is out of date. Greetwell's fix: check the record's last-updated time before rendering, and fall back to safe, general language if it's more than 10 minutes old, instead of guessing. Before the wider rollout, the team set a bar too: a spot check of 150 rendered videos, and at least 97 out of 100 had to state the exact right plan and location count. The first batch cleared it at 98.
Welcome Reel shipped. It worked, technically. Every new customer got a video within minutes, right on schedule for Clemency's board meeting. And for ten weeks, completion sat around 23 percent, and nobody asked why.
The problem was never that the video was wrong. It was never wrong. The problem was a number nobody had written down anywhere, weighed against a number everybody had.
At its worst, this is Bertille Bouchard, four locations of Bouchard Fabricare, watching eighteen seconds of a video that never says which location it's even about, then closing the tab. It's Applecross Mutual, Strathmore's client, whose own new customers are quietly forming an opinion about Applecross from a video Applecross didn't even write. And it's Greetwell's own roadmap, six months in, still pointing at 23 percent completion as its one shipped feature's whole track record.
The choice I would take back
Nothing in Strathmore's roadmap process required a candidate feature to write its value case down anywhere, once it left the review as a decision. That was fine the first two times the team picked the fast option, because both of those fast options happened to have a real, strong value case too. It stopped being fine the first time fast and valuable pointed in different directions, and nobody had ever built a rule for telling the two apart.
What I would leave alone
The one-click "resend this video" button never needed a written value case. It's fast to build and the need is obvious: sometimes a video lands in a spam folder. Gating something that small, just to look consistent, would only teach the team the whole process is theater.
The lesson: an AI feature picked because it's easy to build optimizes for what an engineer can finish first, not for what a customer like Bertille is actually short on. Those two things sometimes line up. The job is checking, on purpose, whether they line up this time, not assuming they always will because they did the last two times.
Now here is the same thing as a story
Stage five above compresses this into four sentences. Here's the ten weeks underneath, the part a stand-up answer skips.
Iskandar Millbrooke has owned Greetwell's roadmap for a year and a half. Ask him what makes a good first feature and he'll tell you: something small enough to ship in one sprint and honest enough that nobody's embarrassed by it six months later. He believed that. He still does.
Twice before, Strathmore had picked the fast option for a first feature, and twice it had turned out fine. A one-line subject line personalization, two days to build, real value: open rates climbed nine points. A branded intro card before the video started, four days to build, customers actually liked it. Fast and valuable had lined up twice in a row, and nobody had ever had to ask whether they always would.
Then came the review for the video feature. Clemency Roscommon, Strathmore's VP of Product, had a board meeting in five weeks and wanted something live to point to. Nazneen Featherstonehaugh, the engineer who'd build it, laid out both options: Welcome Reel, three weeks, and Plan-Aware Reel, nine weeks, because it meant pulling each customer's real account data and checking it before the model said anything specific. Clemency did the math out loud: "Three weeks gets us there before the board meeting. Nine weeks doesn't." Nobody in the room asked what either option was actually worth to the person receiving it. The meeting had a number already, three against nine, and that was the only number anyone had.
Ten weeks of a flat, unremarkable number, sitting there the whole time nobody split it by who was actually watching.
Welcome Reel shipped on schedule. Iskandar watched the first batch of videos go out and felt fine about it: clean render, warm voice, no complaints. For ten weeks, that stayed true. Completion sat around 23 percent, which nobody flagged, because nobody had a number to compare it to. An AI feature was live. That was the headline everyone had wanted.
Then Bertille Bouchard called Applecross Mutual's support line, not angry, just puzzled. She runs Bouchard Fabricare, four dry-cleaning locations, and had just signed Applecross's multi-location commercial policy. Adaobi Ravensbourne, who handles Greetwell's customer success line, took the call. Bertille's whole complaint fit in one sentence: "I stopped watching after about eighteen seconds. It never said which of my locations it was even about."
Adaobi wrote it up and, mostly out of habit, pulled the completion numbers by customer type before sending the ticket along. Single-location Applecross customers, the ones whose situation the generic script fit well enough, were finishing the video 54 percent of the time. Multi-location customers, the ones with something specific and real to be told, were finishing it 21 percent of the time. The customers who needed the video most were the ones it served worst, and the number had been sitting there the whole ten weeks. Nobody had ever looked at it by segment.
We didn't lose eighteen seconds of Bertille's attention. We lost the one chance a new Applecross customer gets to trust Greetwell before deciding it isn't worth watching.
I keep coming back to that roadmap review. "Three weeks gets us there before the board meeting. Nine weeks doesn't." Clemency wasn't wrong about the math. She was answering a question nobody had actually asked her to weigh against anything. The room had one number in it. It should have had two.
Here's the redo, run properly. Same two options, but this time each one had to write down its value case in one paragraph before Clemency saw a single week estimate. Plan-Aware Reel's case: multi-location customers make up 31 percent of Applecross's book and are worth roughly double the annual premium of a single-location one, and they're exactly the segment a generic video serves worst. Ranked by that, Plan-Aware Reel won, nine weeks and all. It shipped in week twenty. By week twenty-two, completion for multi-location customers was at 51 percent. By week twenty-six, it was at 64 percent.
One design let three weeks answer a question about value. The other made value answer for itself, in writing, before anyone asked how long it would take.
What I'd tell myself, sitting in that first review: two quick wins in a row isn't proof that fast and valuable always agree. It's proof you got lucky twice and never checked.
GUARD, run against a video Bertille stopped watching at eighteen seconds
This was never really about whether the generic video was a bad idea. GUARD is for naming who actually pays when build time gets to decide by default, and turning "weigh value against effort" into a rule specific enough to check.
GGroups. Who's actually carrying the risk of the easy option winning.
Bertille Bouchard and every other Applecross customer who received a video chosen for how fast it was to build, not for what they actually needed to hear. And Greetwell's own roadmap, six months into proving an AI initiative is worth funding, spending its first real shot on a feature almost a quarter of its most valuable customers barely finish.
Neither group did anything wrong. Bertille signed up in good faith. The roadmap needed a real first win. Naming both before reacting to the completion number is the whole first move.
The team held the only number that made it into the room. The customer just got whatever that number picked.
UUnequal. Where the harm actually lands.
The harm doesn't spread evenly. It lands specifically on multi-location customers like Bertille, the ones whose real situation was more valuable to serve well and harder to build for, exactly the group a generic script was least likely to satisfy. Anyone checking Greetwell's overall numbers saw a shipped AI feature and a real, if unremarkable, completion rate. Nobody checking only the average would ever have found the 21 percent hiding inside the 23.
The harm concentrates on the customers a value case would have flagged first, and disappears completely the moment nobody splits the average by who's actually watching.
Welcome Reel completion rate, by Applecross customer type, week 8
Simple situation, generic video mostly fitsComplex situation, generic video misses
The blended average, 23 percent, hid both of these numbers for ten weeks. Nobody had split it by customer type until Bertille's call.
AAbility to contest. Could anyone push back on "easy" with a real value comparison.
At the roadmap review, the room had exactly one number in it: three weeks against nine. Nobody put a real user value case on the table next to it, and nobody was asked to. Clemency's own math wasn't wrong. It just never had to compete with anything, because ease alone was the only argument that showed up.
This is GUARD's sharpest question for this exact scenario: not whether the easy feature was a bad idea, but whether anyone in the room had a real number to weigh against build time, or whether ease won just because it was the only thing anyone measured.
Four of these five boxes happened exactly as designed. The one in the middle, a real value case, was never built into the path at all.
RReduce. The actual fix, not a policy memo.
Every candidate first feature now writes its real user value case, in one paragraph, before anyone sees a build estimate. Ranking happens on the value case first. Build time only breaks a tie between two value cases that are genuinely close, and Welcome Reel and Plan-Aware Reel's cases were never close.
The alternative worth naming and rejecting: score every candidate automatically, value divided by build weeks, and rank by the number that comes out. That removes the argument entirely, which sounds efficient. It also means nobody ever has to say the value case out loud, so nobody can push back on it either, and a bad guess hidden inside a formula is just as capable of steering the roadmap wrong as a bad habit is.
Four lines. None of them slow the team down for long. They just make sure ease has to argue for itself.
DDetect. How you'd know it's already happened.
Track completion by customer segment on every newly shipped feature, every month, not just the blended average. Before the fix, only the blended number ever reached a dashboard. After it, a segment split is required within the first 30 days of any feature's launch.
The failure worth naming plainly: if a low number just gets quietly filed as "needs a better subject line" or "needs more marketing," with the original decision never reopened, that's not detection. That's a low number nobody's allowed to blame on the choice that actually caused it.
Multi-location completion rate, by week, weeks 1 through 26
Multi-location completion, weekly
Twenty weeks flat around 20 percent, including the ten weeks before anyone knew to look. The fix didn't nudge the number. It tripled it.
The trade-off, said out loud
Plan-Aware Reel costs about 38 cents to render per video, against about 4 cents for a Welcome Reel that's pre-rendered once and reused for everyone, and it takes about 90 seconds to generate instead of being instant. Strathmore accepted that cost on purpose, for the segment where a generic video demonstrably failed, instead of either serving every customer the same cheap video forever or hand-personalizing every single one, which would have erased the whole reason an AI tool was worth paying for in the first place.
And if you want to be sure it really works, try it somewhere else
Same five letters, a veterinary scheduling tool instead of a welcome video, and this time the two options are a reminder text and an after-hours symptom check.
Same fork Strathmore faced, a different kind of business entirely. The honest branch is still the one a real value case points to.
Fetlock Assist, built by Fetlock AI, helps a multi-clinic veterinary chain manage scheduling and pet-owner messages. Its team faced the identical fork right after their first funding round closed. Easy: AI-written appointment reminder texts, friendly filler swapped onto an existing reminder system, a two-week build. Valuable but harder: an AI symptom check for after-hours calls, telling a worried pet owner whether their dog's limp needs an emergency vet tonight or can wait for a scheduled visit, which meant training and checking a real triage model plus a legal review of the escalation wording, a ten-week build.
Same rank, mapped onto Fetlock: size the value case first. A quick look back at past after-hours calls found 40 percent of pet owners who'd called an emergency line at 2am could have safely waited until morning, and 12 percent who hadn't called were pets that needed same-night care and didn't get it. Reminder texts, by contrast, only ever addressed a no-show problem worth about 3 percent of missed appointments. Fetlock built the triage tool first. It costs more, a licensed vet tech reviews every "wait until morning" recommendation before it's sent, but it's the feature actually worth the funding round.
Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: never let one candidate feature's build estimate reach a decision with no value case sitting next to it. A number with nothing to compare against always wins by default.
Cost: no time to write a full value case this sprint. Write three sentences instead of a paragraph: who's underserved, by how much, and how you'd check. A rough value case beats no value case.
The model got better, for real: say a cheaper model drops Plan-Aware Reel's per-video cost to 6 cents next year. That's a reason to widen who gets the personalized version, on purpose, with a new value check behind it, not a reason the cheap option should have won by default the first time.
Where people run it wrong.
They treat "ship the easy one first" as automatically fine, since something AI-shaped went live either way.
They let whichever number reached the room first, here, build weeks, become the only number anyone ever measures against.
They wait for a customer complaint to reveal the gap, instead of splitting usage by segment the moment a feature ships.
How to use it live. Before answering, ask out loud: "was there ever a real value case written down for the harder option, or did build time win because it was the only number in the room?" Naming that split buys a few seconds of thinking time, and it's most of the real answer.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
Which framework fits "what's the risk of picking your first AI feature based on what's easiest to build"?
Tap to flip
ANSWER
GUARD: groups, unequal, ability to contest, reduce, detect. It fits because the real test isn't whether the easy feature was bad, it's whether anyone weighed it against a real value case before it won by default.
2 · THE PEOPLE
Who are the people this answer names?
Tap to flip
ANSWER
Iskandar Millbrooke, Greetwell's product manager. Clemency Roscommon, Strathmore's VP of Product. Nazneen Featherstonehaugh, the engineer. Adaobi Ravensbourne, customer success. Bertille Bouchard, who owns Bouchard Fabricare and is an Applecross Mutual customer.
3 · THE OLD DEFAULT
What did Strathmore's roadmap review default to before there was a real process?
Tap to flip
ANSWER
Ranking every candidate feature by how many weeks it would take to build, because two earlier quick wins happened to also be the right call, so nobody had ever built a check for when fast and valuable wouldn't line up.
4 · THE TWO-WAY TRAP
What's the two-way trap this question is actually testing?
Tap to flip
ANSWER
Ship the easy feature with no value case, and the AI initiative may burn its first real shot on something almost nobody needed. Refuse to ship anything until every value case is airtight, and nothing ships at all, including the fast wins that really were the right call.
5 · THE OLD DECISION
What decision would you take back?
Tap to flip
ANSWER
Nothing in Strathmore's roadmap process required a candidate feature to write its value case down anywhere. Build time was the only number that ever reached the room, so it decided by default.
6 · THE NUMBER
Fill in the blank: Welcome Reel's completion rate sat at ___ percent overall, but split by customer type it was ___ percent for single-location customers and ___ percent for multi-location customers like Bertille's.
Tap to flip
ANSWER
23 percent overall. 54 percent single-location. 21 percent multi-location. That 33-point gap between segments is what a single blended number hid for ten weeks.
7 · THE REPLAY
Same launch, new process. What changes?
Tap to flip
ANSWER
Both candidates write a value case before anyone sees a build estimate. Plan-Aware Reel wins on value despite costing six more weeks, ships in week 20, and multi-location completion climbs from 21 percent to 64 percent by week 26.
8 · CROSS-PRODUCT TRANSFER
Section 4 runs GUARD again on a different product. Which one, and who plays the equivalent roles?
Tap to flip
ANSWER
Fetlock Assist, Fetlock AI's veterinary scheduling and messaging tool. AI reminder texts play the easy role, after-hours symptom triage plays the harder, more valuable one: 12 percent of pets needing same-night care weren't getting it, the value case that should have decided, not the two-week build estimate.
Check yourself Score: 0 / 0
Multiple choice
1. Why is picking the easiest-to-build AI feature risky, according to this answer?
A. Because easy features are always lower quality than hard ones.
B. Because ease quietly becomes the deciding factor even when nobody chose that on purpose, so the feature that ships isn't the one anyone actually needed.
C. Because customers always prefer more complicated features.
D. Because engineers refuse to build easy features carefully.
Show hint
Check the direct answer at the top of the page.
Show answer
B. The risk isn't that easy is bad. It's that ease substitutes for a real value comparison nobody ever made on purpose.
True or false, with why
2. True or false: the Welcome Reel gave customers factually wrong information.
True
False
Show hint
Check the block-highlight sentence right after "How it goes now" in Let's learn.
Show answer
False. The video was never wrong. It was accurate and generic. The problem was that it never said anything specific to a customer like Bertille, not that it lied.
Fill in the blank
3. Welcome Reel's overall completion rate sat at ___ percent for ten weeks, hiding a ___ percent completion rate among multi-location customers underneath it.
Show hint
Check flashcard 6, and the bar chart in the GUARD recap.
Show answer
23 percent, and 21 percent. The blended number looked unremarkable. The segment underneath it was the real story.
Short answer, name the rejected alternative
4. Strathmore considered one other fix besides requiring a written value case. What was it, and why was it rejected?
Show hint
Check the Reduce step in the GUARD recap.
Show answer
Model answer: An automatic formula scoring value divided by build weeks. Rejected because it hides the judgment call inside a number, so nobody has to say the value case out loud, and nobody can push back on a formula the way they can push back on a written case.
Short answer, apply it yourself
5. Think of an app you use with an AI feature in it. Name one harder, more valuable feature it could have built first instead, and what real user need would have made the case for it.
Show hint
Look for a place where an available model made a generic version easy, while the version that actually used your own data would have taken real integration work.
Show answer
Model answer: A budgeting app choosing to auto-label transactions with a generic category (easy) instead of building AI alerts tied to a user's specific upcoming bills and real balance (harder, but the actual reason most people open a budgeting app).
Fill in the blank, work the number
6. Multi-location completion went from 21 percent before the fix to 64 percent by week 26 after Plan-Aware Reel shipped. Roughly how many percentage points did it gain, and is 64 closer to double, triple, or quadruple the starting 21 percent?
Show hint
Subtract first, then divide 64 by 21.
Show answer
About 43 percentage points, and roughly triple. 21 times 3 is 63, almost exactly the new number. The fix didn't nudge the segment. It tripled it.
Before you close the answer
Why this works
Tests whether you can catch a decision that never felt like a decision at all, three weeks beating nine, because nobody put anything next to it. Most candidates only notice the risk once a customer complains.
Follow-up traps
"What if you genuinely can't tell which feature has more value ahead of time?" Response: write the value case as your honest best guess and name what you don't know yet. The honest case for Plan-Aware Reel didn't need a crystal ball, it needed someone to check what share of Applecross's book was multi-location before the review, not after.
"Isn't a written value case just a way to justify whatever the PM already wanted?" Response: only if nobody else in the room can challenge it. The fix works because the value case gets written where Clemency, Nazneen, and anyone else in the review can see it and argue with the actual numbers, not because one person's opinion counts more.
If pressed
The rejected auto-formula idea, value divided by build weeks, was actually run backward against Strathmore's two earlier quick wins. It would have picked correctly both times, which is exactly why it felt safe to trust. It only would have gotten this one wrong, because both inputs going into it are still guesses, and a formula just makes a guessy input look precise once it lands on the other side of a division sign.
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.
