Artifact critiqueIntermediateAI Opportunity & Model Strategy / Build vs buy vs fine-tune decisions / #19
Structure the one-page recommendation memo for a build-versus-buy decision.
PICKthe real number wasn't missing, it was on page fourteen of a deck nobody read in order
Harborlyne makes property management software. LeaseVane is the feature meant to pull renewal dates and rent-escalation clauses out of scanned leases. Ingrid Sundborg is the product manager who has to write the memo recommending whether Harborlyne builds that extraction in house or buys a vendor's document AI, and Dov Jarrahi is the ops lead whose one offhand question changed how every memo after it got written.
The direct answer
Put the recommendation, in one sentence, on line one, before any table. Then one paragraph on who feels each kind of mistake. Then the paragraph that actually carries the argument: the cost asymmetry, in the reader's terms, not a vendor's spec sheet. Close with the kill criteria, the exact evidence that would reverse the call. Every comparison table and pricing tier goes in a labeled appendix, after all of that, not before it.
Do this, in order
Put the one-sentence recommendation first, above any table.Why: readers skim. Bury it, and the memo fails at the one job it has.
State who feels each kind of error, in plain terms, not a metric.Why: a vendor accuracy percentage means nothing until it's "a lease auto-renews at the wrong rate."
Give the cost asymmetry its own paragraph, the real argument.Why: it's the one sentence the whole memo should be remembered by.
Always name kill criteria, the number that would reverse the call.Why: without it, "approved" becomes a decision nobody ever revisits.
Push every comparison table and pricing tier to a labeled appendix.Why: evidence belongs after the decision, not blocking the reader's way to it.
Keep the decision itself to one page; let the appendix run as long as it needs.Why: length is what actually determines whether anyone reads it at all.
How to answer this, stage by stage
Nobody is scoring how thorough the memo sounds. They're scoring whether a busy reader could act on it without ever opening the appendix.
Stage 1
Scope it to one real memo
Say it like this
"Let's ground this in one real memo. Harborlyne makes property management software. LeaseVane is supposed to pull renewal dates and rent-escalation clauses out of scanned leases. Ingrid has to write the one-page memo recommending whether we build that extraction ourselves or buy a vendor's document AI. That's the actual document I'd structure."
Why this works
Keeps the answer from turning into generic memo-writing advice with nothing real underneath it.
Stage 2
Say the structure out loud before any content
Say it like this
"I'll structure it as PICK. Position, my recommendation, in one sentence, first. Impact, who feels each kind of mistake. Cost asymmetry, the one paragraph the whole argument turns on. Kill criteria, what would make me reverse it."
Why this works
Signals a repeatable memo structure, not a one-off writing style that happened to work once.
Stage 3
Reframe: it isn't "how thorough," it's "what does the reader need to act"
Say it like this
"This isn't really a question of how thorough the memo should be. It's a question of what a busy portfolio manager actually needs to read to act, and what can wait in an appendix for whoever wants to check the work."
Why this works
This is where a strong answer separates from someone who just says "keep it short" with no reason why.
Stage 4
Give the one decision: the actual section order
Say it like this
"Here's the actual layout. Line one: 'Recommendation: buy the vendor API now, revisit build past 56,000 leases a year.' Then one paragraph on who feels a missed renewal date versus who feels an over-flagged clause. Then the cost asymmetry paragraph. Then the kill criteria. Every comparison table and pricing tier goes in a labeled appendix below all of that."
Why this works
This is the direct answer, stated as an actual document layout you could hand someone this afternoon.
Stage 5
Prove it with the compressed evidence
Say it like this
"Three months ago, a thirty-page comparison deck for a different feature nearly got approved on the wrong option, because the real number was on page fourteen and a reviewer, going on gut feel, argued for the pricier one out loud in the room. Dov caught it after the meeting, not during it."
Why this works
Compresses the whole case into the one real failure that justified rebuilding the format at all.
Stage 6
Name the AI-specific reasoning and the trade-off
Say it like this
"The reason this isn't just a memo-writing tip is the actual asymmetry underneath it. A vendor model that over-flags a routine clause costs an analyst two minutes. One that misses a renewal date costs Harborlyne a lease auto-renewing at last year's rate, discovered months later, when it's too late to renegotiate. So the memo has to say plainly: we're accepting more of the cheap error to avoid the expensive one, or the reader only remembers the accuracy percentage, not what it actually costs."
Why this works
This is the load-bearing, AI-specific judgment. A generic vendor decision doesn't hinge on which kind of model mistake is cheap versus catastrophic.
Stage 7
Say what wouldn't get this structure, then close
Say it like this
"I wouldn't force this same one-page structure onto an early, exploratory bake-off with only two vendor demos and no accuracy numbers yet. For LeaseVane, the structure held: recommendation first, cost asymmetry in the middle, kill criteria at the end, appendix for whoever wants to dig."
Why this works
Closes with real judgment about when the format doesn't apply, and restates the direct answer in one breath.
Let's learn
The memo is one page, printed single-sided, that Ingrid pins above her desk until the decision gets signed off.
LeaseVane reads a scanned lease and pulls out the renewal date and the rent-escalation clause, instead of an analyst reading the whole lease by hand.
Before this format existed, every AI build-versus-buy decision at Harborlyne came with a comparison deck, 28 to 34 pages, feature matrices and vendor pricing tiers, read start to finish in a 90-minute meeting almost nobody actually finished. Now Ingrid writes one page. Portfolio managers read the top third, glance at the kill criteria, and sign off in under ten minutes, appendix optional.
This is the actual path a thirty-page deck took through one real meeting, not the path anyone designed on purpose.
Here's the turn: the extra length was never making the decision safer. It was hiding the one paragraph that actually carried the argument inside pages nobody read in order, so the real recommendation could get lost, or overturned by whoever spoke loudest in the room.
We didn't make the decision more thorough. We made it easier to overturn by accident.
Missed renewal dates vs over-flagged clauses, vendor API vs in-house prototype
Vendor APIIn-house prototype
The vendor is worse at the cheap error and better at the expensive, hidden one. That's exactly the asymmetry the memo has to name, not average away.
At its worst, a wrong option gets approved in the room because the real number was on page fourteen, and nobody notices until a renewal date is missed months later.
The choice I would take back
Harborlyne's decision template never required a stated kill criteria, and never separated the recommendation from the evidence. Both lived merged in one long document. That made sense when Harborlyne shipped one AI feature a year and nobody needed a fast path. It stopped making sense once decisions started arriving monthly and nobody had time to read thirty pages in order every time.
What I would leave alone: I wouldn't force a one-page memo onto an early, exploratory bake-off with only two vendor demos and no real accuracy numbers yet. That decision genuinely needs the longer, exploratory format, because there's no cost asymmetry to state until real evidence exists.
The lesson: a memo's job isn't to prove you did the work. It's to let someone who trusts you skip the parts they don't need to read.
Now here is the same thing as a story
The short version above is what you'd hand someone this afternoon. Read this one for what it actually felt like the day a thirty-page deck almost won on the wrong page.
Every other Tuesday, Ingrid prints her memo and walks it into the portfolio review before anyone else arrives.
Ingrid Sundborg had written product decision memos for six years, and could turn a messy debate into a one-page recommendation faster than anyone else on the team. Before her one-pager format existed, though, Harborlyne wrote full comparison decks for every build-versus-buy call, and Ingrid was always the one asked to write them, because she was thorough.
The decks thinned in three beats, in the wrong direction. At first they ran 15 pages, and people read them start to finish. Then 22 pages, and people started skimming the middle. Then 30 pages, and people only read the first two and the last two, skipping the actual evidence in between.
Four tells that a memo has stopped doing its job, none of them about the quality of the evidence inside it.
In one review, Dov Jarrahi leaned over mid-meeting and said, "I only ever read your first paragraph and the price line. What's actually on page fourteen?" Nobody in the room could answer him.
Knowledge spark: what is a cost asymmetry?
Two kinds of mistake a model can make almost never cost the same. One is cheap and happens often; the other is rare and expensive. A cost asymmetry is naming which is which, on purpose, instead of averaging both into one accuracy number that hides the difference.
It turned out page fourteen held the one number that mattered: the actual missed-renewal-date rate from a different vendor bake-off. The room had just approved the wrong option, based on a louder objection instead of that number.
We didn't lose one afternoon rereading page fourteen. We nearly locked Harborlyne into the wrong vendor for a year.
One error is small and cheap and everywhere. The other is rare, expensive, and invisible until it isn't.
Ingrid never had a fixed rule for how long a memo should be. She had a switch: either the reader could find the real recommendation and the number it depended on in the first page, or they couldn't, and then the room decided by whoever spoke last.
When the decision-memo template was first built, nobody separated "the ask" from "the evidence," because early decisions were small enough that a five-page memo covered everything anyone needed, recommendation and evidence together, no harm done.
The whole memo, reduced to its four moving parts. Nothing on this page is an appendix.
Rerun with the new structure: recommendation and kill criteria fit on page one, cost asymmetry stated in one paragraph, appendix runs twelve pages for whoever wants it. The same review now takes eleven minutes, and the actual missed-renewal-date number gets read by everyone in the room, not just whoever reached page fourteen.
One version buries the real number on page fourteen of thirty. The other prints it in the first paragraph, in a sentence a portfolio manager reads standing up.
What I'd tell myself, watching that room nearly approve the wrong vendor: length was never proof of rigor. It was just a place the truth could hide.
PICK, the four questions this memo has to answer, in orderNot a template for making memos longer. PICK is what keeps a short memo from being a thin one.
P
Position. The recommendation, in one sentence, before any reasoning.
"Buy the vendor API now for LeaseVane. Revisit build only past a stated volume." Stated on line one, before any table.
A memo that opens with context instead of a position is asking the reader to find the ask themselves.
I
Impact. Who feels each kind of mistake, in what units.
A missed renewal date is felt by the portfolio manager, months later, as a bad rate locked in. An over-flagged clause is felt by an analyst, the same afternoon, as two extra minutes.
Naming both sides is what stops the memo from sounding like it only cares about one kind of error.
C
Cost asymmetry. The paragraph the whole memo turns on.
Missing a renewal date is rare, expensive, and hides for months. Over-flagging a routine clause is common, cheap, and gets absorbed the same day. Optimize the pick against the one that's expensive and hidden.
This is the hardest paragraph to write and the one the reader should still remember a month later.
K
Kill criteria. What evidence would flip the pick.
Revisit build once lease volume crosses about 56,000 a year, the point where vendor per-page fees exceed two engineers' salaries, or if the vendor's miss rate doesn't improve after two quarters of feedback.
Without this, "approved" quietly becomes "approved forever," whether or not the facts still support it.
The kill line: annual cost by lease volume, vendor fees vs in-house build
Vendor API, per-page feesIn-house build, two engineers
Past about 56,000 leases a year, the vendor's per-page fee costs more than two engineers would. That crossing point is the kill criteria's actual number, not a round figure picked for comfort.
The recap, one line per letter: position is the one-sentence recommendation on line one, impact names who feels a missed renewal date versus who feels an over-flagged clause, cost asymmetry is the paragraph naming which error is expensive and hidden, and kill criteria is the stated number, 56,000 leases a year, that would make the whole memo worth reopening.
And if you want to be sure it really works, try it somewhere elseSame four letters, a textile mill instead of a leasing office. This time the pick lands on build, not buy.
Hollis Rieper runs quality assurance at Wovengate Textiles, deciding whether to buy a vendor's fabric-defect vision model or build a custom one for spotting weave flaws on the line. Mapped onto PICK: position is build, not buy, stated up front. Impact: a missed weave flaw ships a flawed bolt to a retailer, discovered weeks later in a return; a false alarm halts the line for a good bolt, costing a few minutes of production time. Cost asymmetry runs the same direction as Harborlyne's, optimize against the rare, hidden, expensive error, but it points the pick the other way here, because the vendor's miss rate on weave-specific flaws never dropped under 2 percent across a full season, while an in-house model trained on Wovengate's own weave patterns held under 0.6 percent. Kill criteria: revisit buying again if the vendor's rate improves past 0.5 percent within two seasons, or if in-house maintenance costs exceed the vendor's quote.
Two different industries, the same shape of decision: the error worth optimizing against always sits in the same corner.
Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "recommendation first, cost asymmetry in the middle, kill criteria at the end, everything else in an appendix," and stop.
Cost: no time to gather real accuracy numbers before the memo is due. Say so honestly, and state the cost asymmetry as your best current estimate, flagged as such, rather than skipping it.
The vendor got better, for real: if the vendor's accuracy improves past your kill criteria on its own, that's not a reason to skip the memo, it's exactly the trigger the kill criteria was built to catch.
Where people run it wrong.
They write a longer memo when they're unsure, instead of a shorter one with an honestly stated uncertainty.
They put the comparison table before the recommendation, so the reader builds their own opinion before reading the ask.
They skip kill criteria because the decision feels final, which is exactly when a decision quietly becomes impossible to revisit.
How to use it live. The moment an interviewer asks you to structure a recommendation memo, ask yourself: what's the one sentence a reader needs if they only read one sentence? Write that first, out loud, before describing any section that comes after it.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework fits structuring a build-vs-buy recommendation memo, and why?
Tap to flip
ANSWER
PICK: position, impact, cost asymmetry, kill criteria. It forces a stated recommendation before any evidence, which is exactly what a busy reader needs first.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Ingrid Sundborg, product manager at Harborlyne, six years writing decision memos, and Dov Jarrahi, the ops lead whose one question changed the whole format.
3 · THE HABIT
What habit did the review room fall into, that the new memo format broke?
Tap to flip
ANSWER
Skimming only the first two and last two pages of a thirty-page deck, skipping the actual evidence in between, including the one number that mattered.
4 · THE ASYMMETRY
What's the cost asymmetry in this memo?
Tap to flip
ANSWER
An over-flagged clause costs an analyst two minutes, the same day. A missed renewal date costs Harborlyne a bad lease rate, discovered months later. Optimize against the missed date.
5 · THE OLD DECISION
What decision would you take back?
Tap to flip
ANSWER
Never separating the recommendation from the evidence in the memo template, so both lived merged in one long document nobody read start to finish.
6 · THE NUMBER
Fill in the blank: revisit the build decision once lease volume crosses about ___ leases a year.
Tap to flip
ANSWER
56,000 leases a year, the point where vendor per-page fees exceed the cost of two engineers.
7 · THE REPLAY
Same review, new memo structure, what changes?
Tap to flip
ANSWER
The real missed-renewal-date number moves from page fourteen to the first paragraph, and the room signs off in eleven minutes instead of nearly approving the wrong vendor.
8 · CROSS PRODUCT TRANSFER
Section 4 answers this again for a different product. Which one, and how does the pick differ?
Tap to flip
ANSWER
Wovengate Textiles' fabric-defect vision model. There the pick lands on build, not buy, because the vendor's miss rate on the rare, hidden error never dropped low enough.
Check yourself Score: 0 / 0
True or false
1. True or false: the thirty-page deck failed because it didn't have enough evidence in it.
True
False
Show hint
Look at where the real missed-renewal-date number actually was.
Show answer
False. It had plenty of evidence. The real number was on page fourteen. It failed because the structure buried the recommendation behind pages nobody read in order.
Multiple choice
2. Why does the memo optimize against missed renewal dates instead of over-flagged clauses?
A. Over-flagged clauses never actually happen in practice.
B. A missed renewal date is rarer but far more expensive and stays hidden for months, while an over-flagged clause is cheap and caught the same day.
C. The vendor charges an extra fee whenever a date is missed.
D. Analysts refuse to review any flagged clause at all.
Show hint
Look at the cost asymmetry step in the PICK recap.
Show answer
B. Rare and expensive and hidden beats common and cheap and visible, every time you're choosing which error to optimize against.
Fill in the blank
3. Fill in the blank: revisit the build decision once lease volume crosses about ___ leases a year.
Show hint
Look at the kill criteria step, and the line chart.
Show answer
56,000 leases a year. That's the point where vendor per-page fees exceed the cost of two engineers, the actual line charted below.
Short answer, where it wouldn't matter
4. Name a decision at Harborlyne where this one-page memo structure would NOT be the right format yet, and say why not.
Show hint
Look at "what I would leave alone."
Show answer
Model answer: An early, exploratory bake-off with only two vendor demos and no accuracy data yet. There's no real cost asymmetry to state, so the longer, exploratory format is still the honest one.
Short answer, apply it yourself
5. Think of a recommendation you've had to make at work or school. What's the one sentence that should have come first, and what got read instead?
Show hint
Think about the difference between what you wanted read and what actually got read.
Show answer
Model answer: Recommending a group project split into two workstreams. The sentence that should have come first was "split by deadline risk, not by topic," but what got read first was a long list of who liked which topic.
Short answer, work the number
6. If the vendor's per-page fee dropped by half, would the 56,000-lease kill line move, and which direction?
Show hint
Look at how the line chart's vendor cost line is built from a per-page fee.
Show answer
Model answer: It would move higher. A cheaper per-page fee means vendor cost climbs more slowly with volume, so in-house build only becomes the cheaper option at a much larger number of leases a year.
Before you close the answer
Why this works
Tests whether you can turn a real, quantified trade-off into a document someone will actually use to decide, instead of one they nod at and forget by the next meeting.
Follow-up traps
"Isn't a one-page memo just hiding the real complexity?" Response: the complexity is still there, in a labeled appendix. The one page states the position and the number it depends on, so a reader can go deeper by choice, not by force.
"What if leadership wants to see every vendor option compared?" Response: that comparison lives in the appendix, available to anyone who wants it, but it isn't what determines whether the memo gets read or acted on.
If pressed
The real kill criteria used two conditions, not one: lease volume crossing 56,000 a year, or the vendor's renewal-date miss rate failing to improve after two quarters, whichever came first, since either signal alone could have been noise.
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.
