Artifact critiqueIntermediateQuality, Cost & Token Economics / Measuring ROI and business impact / #21
Build the business case slide for an AI investment.
ORDER · ranking a business case slide for a supplier-risk AI, and what gets cut
Farsentry watches all 340 of Norrbeck Surgical's suppliers for financial, geopolitical, and compliance trouble, and flags the ones worth a second look before a shipment goes missing. Reyhan Ashendon owns its business case. Every fall, CFO Isambard Kroon puts any tool with no revenue line through the same test: prove it, or lose it. One week before this year's review, a dry run with a colleague nearly took the whole case down before Isambard ever got the chance.
The direct answer
Lead the slide with the model's backtested catch rate against confirmed supplier failures, not a dollar figure: 11 of 13 real disruptions caught an average of 34 days early. Back that with the small, real cost of reacting early, pulled straight from accounts-payable records, never an estimate. Cut the alert count and any extrapolated "total risk avoided" figure completely: an unprovable number is the one claim that, if a CFO manages to poke a hole in it, costs you credibility on every future ask, not just this one.
Do this, in order
Lead with the backtested catch rate, and back it with AP-verified response cost, never an estimate.Why: this pairing is what lets Isambard trust every other number on the slide.
Cut the $2.1M extrapolated figure entirely.Why: it assumes every flagged supplier would have failed without Farsentry, a guess nobody can prove, and it's the one claim that damages every future ask if it gets challenged.
Cut the 1,240-alerts count as a headline.Why: it proves Farsentry is busy, not that it caught anything real.
Backtest two years of flags against Norrbeck's own incident register before writing down a single number.Why: none of the other claims survive a real question if nobody has checked what Farsentry actually gets right.
Frame Farsentry's cost against one bad week, not against this year's documented spend alone.Why: this year's real number is honestly smaller than Farsentry's own budget, and pretending otherwise is the fastest way to lose the room.
Name the compliance-data staleness gap on the slide, in one line.Why: a disclosed gap builds more trust for next year's ask than a slide that pretends the model is perfect.
How to answer this, stage by stage
Nobody is grading whether you can name five things that could go on a slide. They're grading whether you know which one falls apart the moment someone with real numbers pushes on it.
1
Ground it in one real slide, one real reviewer
Say it like this
"Let's ground this in one case. Farsentry watches all 340 of Norrbeck Surgical's suppliers for financial, geopolitical, and compliance risk. Reyhan Ashendon owns its business case. Every fall, CFO Isambard Kroon puts any tool with no revenue line through the same test: prove it, or lose it."
Why this works
An abstract "how do you build a business case slide" answer turns into a slogan fast. One real review keeps every claim something you'd actually have to defend.
2
Say the method out loud before naming a single claim
Say it like this
"I'm going to run ORDER. Name what the slide is actually trying to win, find the claim that can't survive being challenged, say what has to be true before any number goes up, say what's cheap to check first, then rank what leads, what backs it up, and what gets cut."
Why this works
Signals a method already in motion, not five claims arriving in whatever order they occurred to her.
3
Name what the slide is actually competing to win
Say it like this
"Before I rank anything, here's the real target. This slide isn't trying to look impressive. It's trying to win one specific vote, Isambard renewing Farsentry's budget for another year. Every claim gets judged against that, not against how big the number sounds."
Why this works
Without a stated outcome, ranking five candidate claims is a gut call dressed up as a method.
4
Stress test which claim can't survive being challenged
Say it like this
"If I lead with 'Farsentry prevented two point one million dollars in risk,' and Isambard asks how I know all eleven flagged suppliers would have actually failed without us, I don't have an honest answer. That claim falls apart live, and it takes every other number on the slide down with it. The hundred and ten thousand dollars doesn't have that problem. It's sitting in our own accounts-payable records."
Why this works
This is the hardest step, and the one a rushed answer skips. A number that gets debunked in the room doesn't just fail once, it makes every future number harder to trust.
5
Say what has to be true first, and what's cheap to check
Say it like this
"None of these numbers mean anything until Farsentry's own flags get checked against what actually happened. So before I write a single figure down, I'd backtest two years of alerts against Norrbeck's own vendor incident log, and pull the accounts-payable records for the cases where we actually spent money reacting. Both already exist. Neither needs a new study."
Why this works
Naming the dependency, and the cheapest way to test it, is what keeps an unproven number from ever reaching the slide.
6
Give the rank, defend the top pick
Say it like this
"So, in order. Lead with the catch rate, eleven of thirteen confirmed supplier failures caught an average of thirty four days early, because that's what makes every other number worth reading. Back it with the hundred and ten thousand dollars in documented, checkable response cost. Mention the Kesserling near miss once, as the story, not the headline. Cut the alert count. Cut the two point one million dollar estimate completely, or hold it back for if someone asks directly."
Why this works
Restates the direct answer out loud, with the actual order and the reason, not a recap of the priority list.
7
Close on the honest tension, and one line
Say it like this
"One more thing I'd say straight to Isambard. This year's documented spend doesn't cover what Farsentry costs to run, and I'm not going to pretend it does. What it buys is an eighty five percent chance of thirty four days of warning on a failure that would cost us more than Farsentry's entire budget in a single bad week. That's the case."
Why this works
Closing on the uncomfortable honest number, instead of hiding it, is what makes the rest of the slide's numbers believable.
Let's learn
Farsentry watches every one of Norrbeck Surgical's 340 suppliers, financial trouble, geopolitical trouble, compliance trouble, and flags the ones worth a second look before a shipment ever goes missing.
Before Farsentry, only the biggest suppliers ever got a second look. A small one going quiet could sit unnoticed for weeks.
Farsentry blends three kinds of signal into one score from 0 to 100 for every supplier: how their payments and credit are moving, whether their region has new customs delays or sanctions trouble, and whether their certifications and audits are still clean. Cross 68 and a formal review opens, one where procurement pre-qualifies a backup before there's an actual outage.
Three signal groups feed one score. The fifth step, the threshold, is the one every claim on Reyhan's slide quietly depends on.
Knowledge spark: what counts as a confirmed disruption?
An event Norrbeck's own procurement team logged as real: a shipment more than two weeks late, a quality recall, or a supplier going under. Not a guess, and not a flag that never turned into anything. Farsentry's catch rate only gets checked against these, 13 of them across the last two years.
Farsentry has been running for two years now, and it has been busy: 1,240 risk alerts across the 340 suppliers this year alone. Checked against Norrbeck's own incident log, it caught 11 of the 13 confirmed disruptions, an average of 34 days before the trouble would have shown up through normal channels, a late-shipment notice, a missed payment run.
Two years, 13 confirmed supplier failures, checked against what Farsentry actually flagged
Caught earlyMissed, genuinely unpredictableMissed, a fixable gap
85 percent caught. One miss nothing could have flagged. One miss Farsentry should have caught, and almost did.
Here's the turn. The alerts were never the real risk to the budget. The real risk was a number nobody had checked yet.
The alerts were never the real risk to the budget. The real risk was a number nobody had checked yet.
Two confirmed catches this year cost Norrbeck real, logged money to react to, and both are sitting in accounts-payable records right now. Kesserling Titanium, Norrbeck's single-source supplier for surgical-grade titanium bar stock, showed two straight late payment terms and a lien filing; Farsentry's score crossed 68 thirty four days before the missed shipment, and Reyhan's team spent $71,400 on a bridge order and expedited freight from a backup supplier. A packaging-materials supplier in a region hit by a sudden customs backlog crossed the same threshold on its geopolitical score, and $38,600 in expedited air freight kept a shipment moving before the backlog actually reached Norrbeck's dock. That's $110,000 this year, on top of $95,000 in one confirmed case last year, all of it real, all of it checkable against the accounts-payable ledger.
Farsentry's running cost against the confirmed cost of reacting early, two years
Farsentry's own run costConfirmed cost of reacting early
Two years in, the confirmed number still sits under the run cost. That gap is real, and the slide should say so out loud instead of hiding it behind a bigger, softer number.
The decision that mattered
Farsentry's dashboard never recorded whether a closed alert had turned out to be a real, confirmed disruption. That was fine when the team was small enough to remember every case by heart. It stopped being fine the day a budget review needed proof nobody had to take on faith.
What I would leave alone: the suppliers below Norrbeck's smallest spend tier who already have an easy backup lined up. A missed flag there is a shrug, not a real loss, so it's fine that Farsentry checks them on a longer, looser cycle.
The lesson: a business case for a risk tool doesn't get stronger by adding more numbers. It gets stronger by cutting every number you can't hand someone the receipts for.
Now here is the same thing as a story
Read the short version above when you're in the room. Read this one when you want to feel why the biggest number on Reyhan's first draft was the one that could have sunk her.
Reyhan Ashendon ran Norrbeck Surgical's supplier reviews by hand for four years before Farsentry existed. Once a quarter, she'd pull the twenty or thirty biggest accounts and read through payment histories and audit letters herself. She was good at it. She could smell a supplier in trouble from a single late invoice, months before anyone else noticed.
Farsentry arrived and, for the first year, it was the best thing that had happened to her job. It watched all 340 suppliers, not just the big ones, all the time. When a flag came in, she'd open it, check it against what she already knew, and almost always find the model had noticed something real. She trusted it more with every quarter that went by without a surprise.
So she stopped double-checking whether a closed flag had actually meant anything. Not all at once. First she stopped logging the outcome of the small ones. Then the medium ones. By the second year, "closed" just meant closed, whether the supplier had genuinely been fine or the team had simply moved on to the next alert.
Nobody decided this on purpose. It just got easier not to write it down.
The trigger wasn't a crisis. It was smaller than that. A week before this year's budget review, Reyhan walked her counterpart in finance through the slide as a dry run. He didn't push on the story. He pushed on the total. "Two point one million in risk avoided," he read aloud. "How do you know all eleven of those would have actually failed without us?"
She didn't have an honest answer. That was the whole problem.
One of these Reyhan could hand someone the receipts for on the spot. The other only survives if nobody asks how it was built.
What that number really was, once she looked hard at it: eleven confirmed catches, times an assumed average cost of a full production stoppage, as if every single one would have shut down a line without Farsentry there. Some would have. Some would have been absorbed by ordinary buffer stock without anyone noticing. Nobody had ever separated the two.
She didn't build the two point one million dollar number to mislead anyone. She built it the way you build a number nobody has ever asked you to defend.
She spent the next three days doing the thing that should have been a twenty-minute report: pulling two years of Farsentry's flags, cross-checking them one by one against Norrbeck's own vendor incident log, and pulling the accounts-payable entries for every case where the team had actually spent money reacting. Eleven of thirteen confirmed disruptions, caught. An average of thirty four days of lead time. Two real, logged responses this year, $110,000 total, sitting in the ledger where anyone could check them.
The decision that had opened the door traced back to the very first week Farsentry's dashboard got built. Someone asked, in passing, whether a closed alert should carry a field marking it confirmed real, confirmed false, or unclear. The answer was no. The team was three people who trusted each other's judgment completely, and a confirmation field felt like paperwork nobody needed.
Run that meeting again, with the field added from day one. Same two years pass. Same dry run, one week before the review. This time, when her counterpart asks how she knows the number is real, Reyhan doesn't spend three days finding out. She pulls a report that already exists: eleven of thirteen, thirty four days average, $110,000 confirmed. Twenty minutes, not three days.
One design let a slide's biggest number be whatever sounded most impressive. The other let it be whatever Reyhan could already prove.
The whole answer to this question, in one picture. A guess dressed as a number and a number you could hand an auditor are not the same thing wearing different clothes.
What Reyhan would tell herself, back in that first week: skipping the confirmation field wasn't careless. It was the sensible call for three people who trusted each other by feel. It just wasn't a call that was going to survive being asked to prove anything, two years and one budget review later.
ORDER, or what actually earns a place on the slide
Not a way to dress up "trust me, it works" in five letters. ORDER is what forces you to say, out loud, which of your best claims would actually survive a CFO leaning on it.
OOutcome. What is the slide actually competing to win?
Not "look impressive." One specific vote: Isambard Kroon renewing Farsentry's budget for another year. Every candidate claim, the alert count, the near miss story, the catch rate, the two dollar figures, gets judged against whether it moves that one vote, not against how big or how busy it sounds.
Without a stated outcome, ranking five candidate claims is a gut call wearing a method's clothes.
RReversibility. Which claim can't survive being challenged in the room?
The $2.1M extrapolated figure. It sounds like the strongest number on the slide, since it's the biggest one. But it assumes all eleven caught flags would have become full stoppages without Farsentry, a counterfactual nobody can prove, and a CFO who asks "how do you know" gets an honest answer that unravels the claim on the spot. The $110,000 confirmed response cost doesn't carry that risk: it's sitting in the accounts-payable ledger, checkable by anyone who wants to look.
This is the hardest step, and the one a rushed answer skips. A claim that gets debunked live doesn't just fail once, it costs every future ask the benefit of the doubt too.
DDependency. What has to be true before any number goes on the slide?
None of the four candidate claims mean anything until Farsentry's own flags have been checked against real, confirmed cases, not just counted. That backtest is the one thing every other number quietly depends on: without it, the catch rate is a guess, the response cost is unverifiable, and the extrapolated total is worse than either.
Naming the dependency is what stops an unproven number from ever reaching a budget meeting in the first place.
The order a spreadsheet alone won't give you: check the model against real cases before any number built on top of it gets said out loud.
EEvidence. What could you check cheaply before the meeting?
Two things already sitting in existing systems: backtest two years of flags against the vendor incident register, and pull the accounts-payable entries for the cases where the team actually spent money reacting. Neither needs a new study. Both are ready before the next budget cycle, not the one after that.
Cheap evidence beats a claim that's only ever been tested on how it sounds out loud.
RRank. State the order, defend the top pick.
Lead with the catch rate, 11 of 13 confirmed disruptions, an average of 34 days early. Back it with the $110,000 confirmed response cost from the ledger. Mention the Kesserling near miss once, as the story, not the headline. Close with the tail-risk comparison, not the alert count. Cut the alert count and the extrapolated total entirely. Catch rate leads because it's the only claim that doesn't ask Isambard to trust anyone's account of what happened.
If the order would look the same with a different outcome named in step one, it was ranked by gut and the outcome got written afterward.
Three things worth stating directly, since this is where the real judgement sits. The alternative Norrbeck considered, and rejected, was building the case around coverage instead: leading with "Farsentry now scores all 340 suppliers in real time," positioning it as a visibility tool rather than a risk tool. It lost, because watching everything proves nothing about whether the watching actually catches real trouble, and a tool that's busy but unproven is a more expensive way to feel informed, not a business case. The AI-specific failure worth naming by name is input staleness dressed up as good news: the compliance-signal feed for one supplier only updates every three months, so between updates Farsentry silently treated "no new bad news" as "still fine," when that supplier's real certification had already lapsed six weeks earlier. The guardrail is a standing rule: any Tier 1 supplier whose compliance-signal timestamp is older than 30 days gets marked "unverified" on the dashboard, in a visibly different state, instead of quietly scoring as low risk by default of stale, clean-looking data. And the trade-off being accepted plainly: the review threshold sits at 68, not lower, because a shadow run at 55 roughly doubled the number of dual-source reviews procurement had to open, most of them for suppliers who were never actually at risk. Norrbeck is trading a slightly higher chance of missing a slow-moving compliance signal, like this one, for keeping the review queue at a size procurement can actually work through instead of rubber-stamp.
And if you want to be sure it really works, try it somewhere else
Same five letters, a truck fleet instead of a supplier book, and this time the fragile number wasn't extrapolated from nothing, it came from a formula nobody had rechecked in three years.
Axlewatch is Kettleworth Sanitation District's predictive-maintenance tool. It watches fault-code frequency, brake-wear trends, and engine-temperature patterns across 96 garbage and recycling trucks in three yards, and scores each one for how likely it is to break down mid-route. Cross 74 and a truck gets pulled for a priority inspection before it fails on a collection route, where a breakdown means an emergency tow, an overtime relief crew, and a missed-stop penalty in the city's contract.
The decision Kettleworth's team would take back
Axlewatch's first budget slide, eighteen months ago, led with a template calculation: flagged trucks times the average cost of a major repair, about $296,000 a year. Nobody had ever checked whether the flagged trucks would actually have broken down on a route, or would just have needed a routine service anyway. That formula was fine when the fleet was 12 trucks in one yard and everyone could eyeball whether the number felt right. It stopped being fine at 96 trucks across three yards, where nobody actually checks the formula's assumption anymore.
Last year, a council member whose brother is a mechanic asked the finance committee, in public, how many of the flagged trucks would really have failed on-route versus just needed a scheduled service anyway. Endellion Thackway, Kettleworth's fleet maintenance manager, couldn't answer. The funding item got tabled pending "real numbers."
This year, Endellion rebuilds the case the same way Reyhan did. Eighteen months of repair orders and roadside-breakdown logs, cross-checked against Axlewatch's own flag timestamps: 7 of 9 confirmed near-failures caught, an average of 9 days before they would have failed on a route. One miss was collision damage, nothing predicts that. The other was a slow oil-pressure drift Axlewatch wasn't reading yet at the time, a real, fixable gap, not a mystery. The confirmed, repair-order-verified cost of pulling those 7 trucks in for scheduled repair instead of an emergency tow: $58,200, smaller than Axlewatch's own $145,000 annual cost, and Endellion says so plainly instead of hiding it.
Different fleet, same corner of the chart wins. The claim built on a record nobody has to take on faith is the one that survives a challenge.
Same rank, different lever: validating against real records still comes first here too, but the fragile claim points somewhere new. At Norrbeck, the risky claim was extrapolated from nothing. At Kettleworth, it was an old formula that had simply never been rechecked since the fleet was an eighth of its current size, the kind of number that felt safe because it had always been on the slide, not because anyone had tested it. Endellion was tempted to add one more projection, "if we prevent one breakdown every six weeks, that's about $35,000 a year," and cut it too, for the exact same reason the $2.1M got cut: it sounds right and nobody has checked it yet.
Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the rank: lead with the confirmed catch rate, back it with the AP or repair-order-verified figure, cut the count and the extrapolated total.
Cost: there's no budget this quarter to run the full backtest. Pull the smallest cheap slice first, the last 90 days of records, since that's already sitting in the system, and say plainly the catch rate is provisional until the full window gets checked.
The model got better, for real: say Axlewatch's catch rate climbs from 7 of 9 to 9 of 9. The rank barely moves. You still lead with the catch rate over the dollar figure, still back it with verified records, not the template estimate. A better model raises the number in the lead slot, it doesn't change which slot leads.
Where people run it wrong.
They lead with whichever number is biggest, not whichever one survives a follow-up question.
They let an old formula ship once and never recheck it as the real thing it estimates changes shape.
They treat one dramatic save as the whole case instead of the thing that makes the boring, checkable number credible.
How to use it live. Ask, before naming any number: "which of these claims could I hand you the receipts for right now, and which one only works if you don't ask a follow-up." That question alone tells you the rank.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework is this, and what's its one job?
Tap to flip
ANSWER
ORDER: rank candidate claims by which one is hardest to walk back if it gets challenged. Built for prioritization questions, including what goes on a business case slide and in what order.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Reyhan Ashendon, who owns Farsentry's business case at Norrbeck Surgical, a supplier-risk tool that watches all 340 of Norrbeck's suppliers.
3 · THE OUTCOME
What is the slide actually competing to win?
Tap to flip
ANSWER
One specific vote: CFO Isambard Kroon renewing Farsentry's budget for another year. Not looking impressive, winning that one decision.
4 · THE FRAGILE CLAIM
Which claim on the slide is hardest to walk back if it gets challenged?
Tap to flip
ANSWER
The extrapolated $2.1M "risk avoided" figure. It assumes all eleven caught flags would have become full production stoppages without Farsentry, a counterfactual nobody can prove.
5 · THE OLD DECISION
What decision would Reyhan take back?
Tap to flip
ANSWER
Building Farsentry's dashboard without ever recording whether a closed alert had turned out to be a real, confirmed disruption. It made sense when the team was small enough to remember every case by heart.
6 · THE NUMBER
Fill in the blank: Farsentry's backtest caught ___ of ___ confirmed supplier disruptions, an average of ___ days early.
Tap to flip
ANSWER
11 of 13, an average of 34 days early. About 85 percent.
7 · THE REPLAY
Same dry run, the confirmation field added from day one, what changes?
Tap to flip
ANSWER
Reyhan pulls a report that already exists instead of scrambling for three days: the catch rate and the $110,000 confirmed response cost, both checkable, ready in twenty minutes instead of the night before the meeting.
8 · CROSS-PRODUCT TRANSFER
Section 4 answers this same question again for a different product. Which product, and what's the different fragile claim?
Tap to flip
ANSWER
Axlewatch, Kettleworth Sanitation District's predictive-maintenance tool for its truck fleet. The fragile claim there is a $296,000 template estimate nobody had rechecked since the fleet was an eighth of its current size.
Check yourself Score: 0 / 0
True or false
1. True or false: this year's documented, AP-verified response cost on its own is enough to cover what Farsentry costs Norrbeck to run.
True
False
Show hint
Check the paragraph comparing $110,000 to $340,000 in "Let's learn."
Show answer
False. $110,000 is smaller than Farsentry's $340,000 annual cost. The case doesn't rest on this year's documented number covering the bill by itself, it rests on the catch rate plus the tail-risk comparison.
Fill in the blank
2. Farsentry's backtest found it caught ___ of ___ confirmed supplier disruptions, an average of ___ days before the disruption would have shown up through normal channels.
Show hint
Look at the bar chart in "Let's learn," or flashcard 6.
Show answer
11 of 13, an average of 34 days. That's the number every other claim on the slide depends on being trusted.
Multiple choice
3. Which claim would do the most damage to Farsentry's credibility next year if Isambard Kroon found a real hole in it?
A. The 1,240 alerts issued this year.
B. The $2.1M extrapolated "risk avoided" figure.
C. The Kesserling near miss story.
D. The compliance-data staleness gap.
Show hint
Look at the R step, reversibility, in the framework recap.
Show answer
B. It's the one claim built on an unproven counterfactual, the kind that damages every future ask if it gets debunked in the room.
Short answer, name the old decision
4. What old decision would Reyhan take back, and why did it make sense when it was made?
Show hint
Look at the key point box titled "The decision that mattered," right after the second chart.
Show answer
Model answer: Building Farsentry's dashboard without a field recording whether a closed alert turned out to be a real, confirmed disruption. It made sense at the time because the team was small enough to remember every real case by heart, so a confirmation field felt like paperwork nobody needed.
Short answer, apply it yourself
5. Think of a business case or pitch you've seen that leads with an impressive-sounding number. What's one way you'd check whether that number would survive being challenged?
Show hint
Think about whether the number is measured against everyone involved, or just against the people who make it look best.
Show answer
Model answer: A fitness app's ad claims "members lose 12 pounds in 30 days on average." Before trusting it as a headline, check whether that's measured against everyone who signed up, including people who quit after a week, or only against people who finished the program, since those are very different, very checkable numbers.
Short answer, work the judgement
6. If Norrbeck lowered Farsentry's review threshold from 68 to 55, would you expect the catch rate to go up or down, and why might that not be the right call even if it does?
Show hint
Check the trade-off paragraph right after the R-rank step in the framework recap.
Show answer
Up. A lower threshold flags more suppliers, including slower-moving cases like the compliance staleness miss. But a shadow run at 55 roughly doubled the number of dual-source reviews procurement had to open, most for suppliers never actually at risk, so the real trade is a small gain in catch rate against a queue procurement can no longer actually work through.
Before you close the answer
Why this works
Tests whether you treat picking what goes on a slide as a real ranking problem with stakes, not a data-dump instinct, and whether you know a risk tool's honest case is calibrated reliability plus a small, real cost, not a promise that the dollars already cover the bill.
Follow-up traps
"Why not show the $2.1M as a range with a caveat, instead of cutting it completely?" Response: a caveated range still gets remembered and repeated as "$2.1M" past the room. If Reyhan can't defend it line by line, it doesn't go on the slide, it goes in her back pocket for if pressed.
"If the documented cost is smaller than what Farsentry costs to run, isn't it actually losing money?" Response: only if you score it as a savings tool instead of a risk tool. $340,000 a year is smaller than the cost of one uncaught single-source failure, and the 85 percent catch rate is what makes that framing believable instead of a hand wave.
If pressed
The 68 threshold wasn't picked in the abstract. A shadow run at 55 roughly doubled the number of reviews procurement had to open, for a small gain on slow-moving signals like the compliance one. 68 was the point where the queue stayed a size procurement could actually work through, and a single uncaught single-source failure runs $370,000 to $550,000 in lost output and contract penalties, more than Farsentry's entire annual budget in one bad week.
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.
