InterviewAdvancedQuality, Cost & Token Economics / Measuring ROI and business impact / #25
Make the business case for an AI investment to me as if I were the CFO.
PICK · a live boardroom pitch for an RFP-response drafting AI at an environmental contractor
Bidwright drafts the technical and compliance sections of every RFP response Ferngale Environmental's four-person bid team writes, before a person edits and signs off. Araminta Colthurst owns its business case. Every October, CFO Malachy Vranken makes anyone asking for budget defend it across the table, in dollars, on the spot. Six weeks before this year's review, a near miss on a Nevada solicitation nearly proved the case wrong instead of right.
The direct answer
Approve the $420,000, nine-month build of Bidwright Phase 2. The certain part of the math, recovering thirty four RFPs a year Ferngale already declines, pays the investment back about ten times over using today's win rate alone, before any bet on the model writing a better draft. Frame the ask around that certain number, not the win-rate hope, and set a hard compliance-accuracy bar now that would make you kill the riskiest part of it yourself if it doesn't clear it.
Do this, in order
Lead with the capacity number, not the win-rate number.Why: recovering 34 declined bids a year pays back the investment on today's win rate alone, so the win-rate lift is a bet nobody has to make yet.
Name both ways this goes wrong, in dollars, before Vranken has to ask.Why: a pitch that only shows the upside gets picked apart in the room; naming both sides first is what makes the number trustworthy.
Say which mistake shows up on a dashboard and which one never does.Why: approving and underperforming is visible and bounded at under one percent of revenue; declining and losing $7.3 million a year to a rival never appears on an invoice at all.
Set the 99.5 percent compliance-accuracy kill line before the build starts, not after a scare.Why: a kill line named in advance is what separates a business case from a pitch nobody can hold you to.
Keep every compliance and certification claim behind a person, no matter the confidence score.Why: the Nevada near miss proves a high-confidence draft can still carry a stale fact; the guardrail has to cover the whole risk class, not just low scores.
Don't answer the near miss by hiring two more writers instead.Why: $780,000 over three years buys the same four-person ceiling and never touches the compliance gap that actually caused the scare.
How to answer this, stage by stage
Nobody is grading whether you can recite a memo about AI ROI. They are watching whether you can hold a real number steady, live, while someone plays a CFO who is already reaching for a reason to say no.
1
Name the room you're actually in
Say it like this
"Happy to do this live. I'm going to talk to you the way I'd actually talk to our CFO, not describe the conversation from the outside. Ferngale Environmental drafts RFP responses for cleanup and remediation contracts. Our tool, Bidwright, already helps four writers cover most of what we're invited to bid on. I'm here to ask you for the next $420,000."
Why this works
Names the product and the ask in the first breath, and treats the interviewer as the actual audience, not a hypothetical one.
2
Say the shape of the pitch before making it
Say it like this
"Before I make the ask, here's how I'm going to make it. I'll give you my position first, no hedging. Then who actually pays if I'm wrong, in both directions. Then which mistake is the one that really costs you something. Then exactly what would make me take this ask back myself."
Why this works
Signals structure to a skeptical CFO up front, so what follows reads as a plan instead of four points arriving as they occur to you.
3
Give the position, committed, no hedge
Say it like this
"Here's my position, straight: approve the $420,000, nine-month build. Forty nine solicitations we're invited to bid on go unanswered every year, purely because four people can't get to them in time. Recovering thirty four of those, at our current win rate, pays this back about ten times over in year one. I'm not asking you to bet on the model getting smarter. I'm asking you to fund capacity we already know is worth having."
Why this works
This is the direct answer, said live, before the CFO has to dig for it. It also draws the line between the certain claim and the bet before anyone can blur them.
4
Name who feels each error, in dollars, both directions
Say it like this
"Two ways this goes wrong, and you feel both of them differently. If you approve it and it underperforms, that costs $420,000 plus about $85,000 a year to run, under one percent of revenue, and you'll know by month six because that's when the compliance checkpoint reports back. If you decline it, Ferngale keeps missing forty nine bids a year, worth about $7.3 million in contract value we're not even in the running for, and a competitor twice our size, who already runs a tool like this, picks most of it up instead."
Why this works
Turns an abstract "what's the risk" question into two real numbers the CFO can actually weigh against each other.
5
Name the asymmetry, from the CFO's own chair
Say it like this
"Here's the asymmetry, from your chair specifically. A Bidwright that underperforms is the cheap mistake. It shows up on a dashboard, it's bounded at half a million dollars, and you can kill it in month six. A Ferngale that keeps declining bids is the expensive one, and it's expensive because you'll never see it on an invoice. Nobody bills you for a bid we didn't write. It shows up two years from now as a state agency quietly not inviting us back, and by then it isn't a missed contract anymore, it's a missed relationship."
Why this works
This is the hardest move in PICK. Naming which mistake never appears on the CFO's own reporting is what makes the ask an argument, not a hedge.
6
Prove it with the failure, compressed to four sentences
Say it like this
"And I'll be straight about the risk, because you'd find it anyway. Six weeks ago, Bidwright drafted a hazardous-waste permit number into a Nevada bid, pulled from a proposal we won in a different state. That permit had actually been reissued under a new number back in 2024. Our senior writer caught it three hours before the deadline. If it had shipped, we'd have misrepresented our own credentials to a state buyer, and this $420,000 is built to make that failure structurally hard to miss, not just lucky to catch."
Why this works
A real near miss, told plainly, makes the ask sound like risk management instead of a growth pitch dressed up in AI.
7
Say what would make you walk it back, unprompted
Say it like this
"And here's what would make me take this back myself, before you even ask. If the compliance cross-check doesn't clear 99.5 percent accuracy against our live license database, across three years and thirty-plus jurisdictions, by the month-six checkpoint, I pull the lighter review queue completely. Every claim keeps getting a person, forever, and this becomes a smaller, capacity-only tool. And if a backtest against three years of our own lost bids shows no real signal that better drafts win more often, I drop the win-rate story from every report after that and hold this to the capacity math alone."
Why this works
A pitch that names its own failure conditions, unprompted, is what separates a business case from a sales pitch.
8
Close on the decision, in one breath
Say it like this
"So: approve it on the number that's already proven, capacity, not the number that's still a bet, win rate. The cheap mistake is fixable in six months. The expensive one is already costing you $7.3 million a year, and it's the one number on this page that will never show up on your desk unless I put it there myself."
Why this works
Restates the position and the asymmetry together, so the interviewer leaves with the decision, not just the pitch.
Let's learn
Every month, Ferngale Environmental's four-person proposal team looks at nine or ten new RFPs and already knows, before anyone opens a laptop, that it can only really finish six of them.
Bidwright is the tool that closes some of that gap. It reads a solicitation, drafts the technical approach, staffing plan, and compliance sections from Ferngale's own past proposals, scores each drafted section for how sure it is, and hands the whole thing back to a writer to edit and sign off.
Six steps. The fifth one, the compliance cross-check, is the step Phase 2 is actually built to finish.
Knowledge spark: what's a confidence-scored draft?
Bidwright doesn't just hand over a paragraph. It scores how sure it is about each section, against past proposals a person has already reviewed. High score, a writer skims it. Low score, a writer reads it closely. Certification and permit claims get a person's eyes every time, no matter the score.
Before any drafting tool existed, Ferngale's writers finished about 58 of the 140 RFPs they're invited to bid on each year, 41 percent, and declined the rest for lack of time. Bidwright's first version, running for 14 months now, raised that to 91 of 140, 65 percent. Forty nine solicitations still get a decline every year, mostly the biggest and newest ones, the kind where even a fast draft doesn't leave enough runway to check it properly.
Ninety one submitted looks like a win. Forty nine declined is the number that never makes it into a quarterly update.
Here's the turn. The forty nine declines were never the real story, because they show up on a dashboard already, easy to count and easy to justify. The real story is what happens inside the 91 that do get submitted, once writers trust Bidwright's clean score enough to stop reading every certification claim the way they used to.
We didn't buy Ferngale four more writers. We bought the habit of trusting a checkmark instead of the permit itself.
Phase 2 is two things. First, a version of Bidwright fine-tuned on Ferngale's own 340 past won and lost proposals, aimed at closing more of the capacity gap, from 91 submissions a year to about 125, ninety percent of what's offered. Second, an automatic cross-check of every compliance and certification claim against Ferngale's live licenses database, with a review queue that spends a writer's time on the handful of sections still flagged uncertain instead of every line of every draft. It costs $420,000 to build over nine months, plus about $85,000 a year to run.
Recovering 34 more submissions a year, at Ferngale's proven 22 percent win rate and $680,000 average contract value, is worth about $5.1 million a year in additional expected contract value. That number doesn't depend on the model getting any better at writing. It only depends on Ferngale showing up to bid on solicitations it's already qualified to write. Even after Phase 2, about 15 of the biggest, newest solicitations will still sometimes get declined, and that's fine. Rushing the riskiest drafts is exactly the wrong place to buy speed.
Cost, by the numbers: what each option spends against what it's actually worth
Cash spent, year oneContract value recoveredContract value missed
Declining looks free on the invoice. It is the most expensive option on the page, and the only one where the cost never appears anywhere Vranken already looks.
The choice that mattered
When Bidwright launched, the plan was that a person would read every drafted section end to end before it went out, no exceptions. Fourteen clean months later, that full read had quietly become a skim of anything scored high, including certification claims nothing had ever actually checked against a live database.
What I would leave alone: the 15 largest, newest solicitations Ferngale will still sometimes decline even after Phase 2. Some fights aren't worth rushing into. A drafting tool that pushed writers to answer everything, at any speed, would trade a bounded, honest decline for an unbounded, dishonest one.
The lesson: a business case for an AI investment doesn't get stronger by adding a bigger number. It gets stronger by separating the number you can prove today from the one you're still hoping for, and never letting the second one carry weight it hasn't earned.
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 a permit number that never shipped was closer to a real problem than any bid Ferngale ever lost.
Ferngale's proposal room empties out by six most nights, except during bid season, when Araminta Colthurst is still there at nine, comparing a fresh draft against the solicitation's own checklist line by line. She's done this for six years. She can tell a compliant technical narrative from a hopeful one before she's finished the first paragraph.
Bidwright arrived fourteen months ago, and for most of that time it was the best thing that had happened to her team's mornings. It read a solicitation overnight and handed back a scored draft by eight. For the first few months, Araminta and her three writers read every section end to end anyway, the way they always had, checking each drafted claim against the actual certificate, the actual license, the actual permit. The drafts kept holding up. Nobody found a real error for weeks. Then months.
So the full read thinned. Not on purpose, and not all at once. First it was the boilerplate sections that got skimmed instead of read. Then it was the technical narrative. By month ten, a section Bidwright scored above ninety, including a compliance claim, got a quick glance and a signature, the same as everything else that scored clean. Nobody decided this was fine. It just kept being fine, week after week, until it wasn't a decision anymore, just a habit nobody had named.
The trigger wasn't a pattern. It was luck, and Araminta knows it.
Four facts, and only one of them was luck. The other three were already true before anyone opened the draft.
Six weeks before this year's budget review, Bidwright drafted the compliance section for a hazardous-waste cleanup RFP in Nevada. It described Ferngale's transporter permit accurately in every way except one number, pattern-matched from a proposal Ferngale had won in Arizona two years earlier. The permit itself had been reissued under a new number after a 2024 regulatory update, and nothing in Bidwright's first version had ever been built to check that against anything live. The section scored a 94. It would have gone out with everything else.
Leocadia Solmundsson, Ferngale's senior proposal writer, was the one still reading closely that week, because a client call had left her with an afternoon free and an old habit she hadn't fully lost. She caught the mismatch three hours before the deadline, called the state licensing board to confirm, and swapped the number before submission. Nobody outside the proposal room ever knew it happened.
A wrong number in a state bid isn't a typo. It's Ferngale telling a government buyer it holds a permit it doesn't.
Araminta didn't wait for it to happen twice. She spent the next two weeks building the case for Phase 2: a version of Bidwright trained on Ferngale's own history instead of generic templates, and a compliance cross-check that would have caught the Nevada permit automatically, the moment the draft was generated, not three lucky hours before a deadline. Her team also priced the alternative everyone reaches for first, hiring two more writers, $130,000 fully loaded each, $780,000 over three years. She rejected it. Two more people reading drafts by hand doesn't fix the actual gap. It just adds two more people who might also, eventually, trust a clean score a little more than they should.
The decision that opened the door traced back to Bidwright's original launch meeting, fourteen months earlier. Someone asked whether the tool needed a live cross-check against the licenses database from day one. The answer was no, not yet, because Bidwright was only ever meant to produce a draft a person would read in full before it went anywhere. That was true, and it was also the exact assumption that quietly stopped being true somewhere around month ten, without anyone deciding to change it.
The whole answer to this question, in one picture. One number gets watched because it's on a dashboard. The other one has to be put there on purpose, or it never gets watched at all.
Run the budget review again, with Phase 2 already built. Bidwright drafts the same Nevada section. The cross-check flags the permit number against the live database automatically, in seconds, and routes it to a person before the draft ever reaches a writer's desk. Leocadia never needs to be the one with a free afternoon. And Araminta walks into Vranken's office with a number she can defend before he asks a single question, not one she's still hoping holds up.
What Araminta would tell herself, back in that first launch meeting: skipping the live cross-check wasn't careless. It was the sensible call for a tool everyone still planned to read in full. It just wasn't a call anyone had agreed to revisit once the reading stopped.
PICK: the case that has to survive Vranken's first question
Not a way to dress up "trust the model" in four letters. PICK is what forces a real commitment about which number this ask actually stands on, and which mistake would cost Vranken something he'd never see coming.
PPosition. Your pick, in one sentence, before any reasoning.
Approve the $420,000, nine-month build of Bidwright Phase 2. Recovering 34 of the 49 RFPs Ferngale currently declines, at today's proven 22 percent win rate, pays the investment back about ten times over in year one, without assuming the model gets any better at writing.
Say the pick first, addressed straight to the person playing the CFO. An answer that waits through the reasoning to reveal the ask has already lost the room.
IImpact. Who feels each kind of error, in real dollars.
If Vranken approves and Phase 2 underperforms, Ferngale is out $420,000 plus about $85,000 a year, under one percent of revenue, visible at the month-six checkpoint. If he declines, Ferngale keeps missing 49 bids a year, about $7.3 million in contract value it isn't even eligible to compete for, and a rival roughly twice Ferngale's size, already running a similar drafting tool, wins most of it instead.
Naming both people, the CFO who approves and the rival who benefits from a decline, keeps this from turning into a one-sided pitch.
CCost asymmetry. The heart of it.
Approving and underperforming is the cheap mistake: bounded at roughly half a million dollars, visible on a dashboard, and killable at the month-six checkpoint already built into the plan. Declining is the expensive one, and it's expensive specifically because nothing on Ferngale's own reporting ever shows it. No invoice arrives for a bid that never got written. It shows up two or three years later as an agency quietly dropping Ferngale from its invitation list, and by then it's a lost relationship, not a lost bid.
This is the step that earns the pick. Anyone can say an investment is worth it. Naming which failure is invisible to the CFO's own numbers is what makes the case survive a follow-up.
One mistake gets fixed at a checkpoint that already exists. The other one gets fixed by an agency's silence, years later, if it ever gets noticed at all.
KKill criteria. What evidence flips the pick.
A small-scale prototype of the compliance cross-check, tested against five batches of past RFPs so far, is already at 99.1 percent match accuracy and climbing toward the 99.5 percent bar the full build needs to clear, across three years and thirty-plus jurisdictions, by the month-six checkpoint. If it stalls short of that bar, the lighter review queue gets cut completely and every claim keeps a human forever, shrinking this to a capacity-only tool. If a backtest of three years of lost bids shows no real signal that better drafts win more often, the win-rate framing drops from every future report, and the case stands on capacity math alone.
A pick with no kill criteria is a hope you're defending forever. Naming the exact bar, before the build even starts, is what makes this a live decision instead of a promise.
One question, decided before the build starts, so nobody has to argue about it under pressure six months from now.
The kill line, charted: compliance-matrix accuracy across five backtest batches
Prototype accuracy, five backtest batchesKill line, 99.5% required
The prototype is climbing, not guessing. It still needs the full build, and the wider jurisdiction coverage, to close the last half point before anyone bets the review queue on it.
And if you want to be sure it really works, try it somewhere else
Same four letters, a court-interpretation agency instead of an environmental contractor, and this time the near miss wasn't a permit number. It was a language pair nobody was certified in anymore.
Elderglen Language Access bids on state judicial-branch interpretation contracts. Answerforge, its RFP-drafting tool, writes the staffing plans, certified-interpreter rosters, and quality-assurance narratives that go into every bid. Calanthe Yevorah, Director of Bid Operations, is pitching CFO Guerric Ondracek on Answerforge's own next phase: $265,000 over seven months, aimed at closing the gap between the 40 of 64 eligible RFPs Elderglen currently submits and the 58 it could reach with a live interpreter-certification cross-check.
Different bid team, different language pair, the same shape of near miss and the same shape of ask.
Elderglen's near miss: Answerforge drafted a staffing plan listing an interpreter as actively certified in Hmong-to-English court proceedings, pattern-matched from a proposal Elderglen won two years earlier. That interpreter had let the certification lapse eight months prior. A reviewer caught it during a routine client reference call, not inside the proposal process at all, three days before it would have mattered.
The decision Calanthe would take back
Answerforge launched with the same assumption Bidwright did: a person reads every draft in full, so a live certification check could wait. Elderglen's reviewers stopped reading every roster line just as quietly, once six clean months in a row made a full read feel like paperwork nobody needed.
Same rank, different lever: at Ferngale the fragile fact was a permit number. At Elderglen it's a certification date, but the case runs the same way. Recovering 18 more submissions a year, at Elderglen's 19 percent win rate and $310,000 average contract, is worth about $1.06 million a year, the certain math the investment stands on. Declining and staying at 40 submissions costs about $1.41 million a year in contract value Elderglen never competes for. Same asymmetry: one number is bounded and visible, the other is invisible until a state court administrator quietly stops inviting Elderglen to bid.
Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the pick: fund the capacity that's already proven, name both ways it goes wrong, say which one you'd never see coming, name the kill line.
Cost: no budget this quarter for the full nine-month build. Ship the cross-check alone first, on the existing draft volume, and hold the capacity expansion for next year's ask once the accuracy bar is proven at scale.
The model got better, for real: say Bidwright's draft quality jumps and the win-rate bet actually clears 29 percent on its own. The pick doesn't change. You still lead with the certain capacity math, because a better model raises the size of the bet you were never asking Vranken to take in the first place.
Where people run it wrong.
They lead with the win-rate hope because it's the bigger, more exciting number, and let a skeptical CFO poke a hole in the one claim that was never certain.
They treat a near miss as proof the whole tool is broken, instead of proof of exactly which guardrail was missing.
They answer a compliance scare by adding headcount, which looks responsible and touches nothing that actually caused the scare.
How to use it live. Before naming a dollar figure, ask yourself out loud: "which of these two numbers is already proven, and which one am I still hoping for?" That question alone buys real thinking time, and it's usually the exact distinction the interviewer is listening for.
Three things worth stating directly, since this is where the real judgment sits. The alternative Araminta's team considered, and rejected, was hiring two more proposal writers for $780,000 over three years instead of building Phase 2. It lost because headcount scales linearly and arrives after a three-to-six month ramp, while the actual failure the near miss exposed wasn't a shortage of eyes, it was a shortage of anything checking a claim against a live source instead of a person's memory of the last clean month. The AI-specific failure worth naming is a confidently wrong compliance claim: Bidwright's draft scored a 94, high confidence, while quietly carrying a permit number that had been reissued two years earlier, because a model pattern-matching across past wins has no way to know a fact went stale unless something checks it against a live source. The guardrail is a risk-class rule, not just a score cutoff: certification and licensing claims never auto-include, at any confidence score, full stop, only narrative sections do. And the trade-off being accepted on purpose is cost and speed against document size: a forty-page addendum costs Bidwright about $18 to draft end to end, while an eight-hundred-page state solicitation costs upward of $310, because the model has to ingest and cross-reference the whole document, not just draft against it. Ferngale accepts that the biggest, highest-value solicitations get routed to the slowest, most heavily reviewed, most expensive path, on purpose, because those are exactly the bids where a wrong compliance claim would cost the most.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework is this, and what's its one job?
Tap to flip
ANSWER
PICK: commit to a position, then show which of two mistakes actually costs someone something, and name what would change your mind. Built for tradeoff and investment questions, especially live "convince me" pitches.
2 · THE PEOPLE
Who is this answer about?
Tap to flip
ANSWER
Araminta Colthurst, Director of Proposal Technology at Ferngale Environmental, who owns Bidwright's business case, and Malachy Vranken, the CFO who tests every budget ask across the table.
3 · THE POSITION
What's the P step here, in one line?
Tap to flip
ANSWER
Approve the $420,000, nine-month build of Bidwright Phase 2. Recovering 34 declined RFPs a year, at today's proven win rate, pays it back about ten times over without betting on the model getting better.
4 · THE COST ASYMMETRY
Which mistake is cheap and visible, and which is hidden and expensive?
Tap to flip
ANSWER
Approving and underperforming is cheap and visible: bounded near $500,000, shows up on a dashboard by month six. Declining is hidden and expensive: about $7.3 million a year in missed contracts that never appears on any invoice, only in agencies quietly dropping Ferngale years later.
5 · THE OLD DECISION
What decision would Araminta take back?
Tap to flip
ANSWER
Launching Bidwright without a live compliance cross-check, on the assumption a person would always read every draft in full. That assumption quietly stopped holding around month ten, and nobody had agreed to revisit it.
6 · THE NUMBER
Fill in the blank: recovering 34 more submissions a year, at Ferngale's 22 percent win rate and $680,000 average contract value, is worth about $___ a year.
Tap to flip
ANSWER
About $5.1 million a year. That's the certain math the whole ask stands on, built without assuming any win-rate lift.
7 · THE REPLAY
Same board review, Phase 2 already built, what changes?
Tap to flip
ANSWER
The Nevada permit number gets caught automatically by the live cross-check in seconds, before a writer ever sees the draft, instead of by Leocadia having a free afternoon.
8 · CROSS-PRODUCT TRANSFER
Section 4 answers this same question again for a different product. Which product, and what's the equivalent near miss?
Tap to flip
ANSWER
Answerforge, Elderglen Language Access's RFP-drafting tool for court-interpretation bids. The equivalent near miss is a staffing plan listing an interpreter as certified in a language pair whose credential had actually lapsed eight months earlier.
Check yourself Score: 0 / 0
Multiple choice
1. Which number should Araminta lead the pitch with?
A. The projected win-rate lift, 22 to 29 percent.
B. The capacity recovered from bids Ferngale already declines, at today's proven win rate.
C. The total number of RFPs Ferngale has ever submitted.
D. The size of the competitors already using similar tools.
Show hint
Check the P step in the PICK recap, and the direct answer at the top of the page.
Show answer
B. It's the only claim that doesn't depend on the model getting any better, so it's the one that survives a skeptical CFO's first question.
True or false
2. True or false: the Nevada permit number slipped through because Bidwright's draft accuracy had gotten worse.
True
False
Show hint
Check "what I would leave alone" and the story's opening paragraphs about the full read thinning.
Show answer
False. Bidwright's own drafting never changed. What thinned was the writers' habit of reading every clean-scored section in full, once fourteen months of good results made a full read feel unnecessary.
Fill in the blank
3. Closing the capacity gap from 91 to 125 submissions a year, using Ferngale's proven 22 percent win rate and $680,000 average contract value, is worth about $___ a year.
Show hint
Look at "Let's learn," right after the Phase 2 description, or flashcard 6.
Show answer
About $5.1 million a year. Thirty four more submissions, times a 22 percent win rate, times a $680,000 average contract value.
Short answer, name the rejected alternative
4. What alternative did Araminta consider instead of building Bidwright Phase 2, and why did she reject it?
Show hint
Look at the "three things worth stating directly" paragraph near the end of Section 4.
Show answer
Model answer: Hiring two more proposal writers, $780,000 over three years. Rejected because headcount scales linearly, arrives after a three-to-six month ramp, and doesn't fix the actual gap the near miss exposed: nothing was checking a claim against a live source instead of a person's memory of the last clean month.
Short answer, apply it yourself
5. Think of an investment case you've seen pitched, AI or not, that only showed the upside. What hidden, harder-to-see cost was probably missing from it?
Show hint
Think about which cost would never show up on a normal report, not which cost is biggest.
Show answer
Model answer: A retailer pitched a self-checkout rollout purely on labor savings. The hidden cost was shoplifting loss creeping up over the following year, a number that never appeared on the same report as the labor savings, so nobody compared them until a full audit forced the comparison.
Short answer, work the judgement
6. If Vranken's own finance team believed only half the $7.3 million opportunity-cost estimate was realistic, would the case for approving Phase 2 still hold?
Show hint
Check whether the P step's math depends on the opportunity-cost number at all.
Show answer
Yes. The $5.1 million certain math never depends on the opportunity-cost estimate. It's built entirely from bids Ferngale is already qualified to write. Halving the hidden-cost figure changes how urgent declining looks, not whether approving pays for itself.
Before you close the answer
Why this works
Tests whether you can hold a real number steady under live pushback without folding into a vaguer, safer claim, and whether you volunteer your own kill criteria before being asked. A candidate who never considers being wrong hasn't made a business case, they've made a pitch.
Follow-up traps
"Isn't $7.3 million just a guess about what a competitor would do?" Response: it's an estimate, and the case never depends on it holding exactly. The $5.1 million certain math, built only from bids Ferngale is already qualified to write, justifies the investment on its own.
"Why fund the riskier compliance feature at all, why not just the capacity piece?" Response: because the near miss proves capacity alone isn't the real risk. A faster drafting tool with no live compliance check just means more unchecked claims move through, faster.
If pressed
Bidwright's routing isn't one global confidence threshold. Certification and licensing claims never auto-include, at any score, full stop, only narrative sections do. And cost scales with the document itself: a forty-page addendum costs about $18 to draft end to end, an eight-hundred-page state solicitation costs upward of $310, because the model has to ingest and cross-reference the whole thing, not just draft against it. The biggest, priciest RFPs get routed to the slowest, most heavily reviewed path, on purpose.
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.
