Command Center

• "Walk me through what happens from a customer first hearing about you to them having paid." (map the flow — gaps announce themselves)
• "What question do you answer over and over?" → AI concierge/FAQ
• "How does someone book you at 9pm on a Sunday?" → 24/7 booking + payment
• "How do you collect money, and when does it go wrong?" → deposits, auto-charge, no-show protection
• "What happens after the sale — reviews, repeat customers?" → follow-ups/review asks
• "What task would you pay anything to never do again?" (their #1 goes on the map, whatever it is)

Pick the 3 with the clearest money math. For each, write on the map: today's cost → automated → what changes.

Always their numbers said back to them — never your estimates alone.

• Yes: "Proposal tonight — one page, fixed price. 50% starts it, live in 2–3 weeks." Get a verbal date.
• Maybe: "The map's yours either way. The $397 credit is good 60 days — I'll check in Friday." (calendar the follow-up before leaving the table)
• No: "Then the map just saved you from buying AI you don't need — that's the guarantee working. Who's the one operator you know who IS drowning in DMs?" (always leave with the referral)

Same day, never next day: email the finished Action Map PDF within 4 hours ("Your AI Action Map — [Business]"); proposal in the same email if yes/maybe; log it in Pipeline.

Yes → book the 2 hours on the spot. Maybe → "Think on it — and either way, who else should I show this to?"

Every intake call is a case discussion where the prospect is the protagonist. Five beats, ~25 minutes:

• Beat 1 — Cast the case. "Every business problem is a case study: a protagonist, a decision, and a clock. You're the protagonist — so what's the decision on your desk, and when does it have to be made?" (If there's no decision, there's no engagement — politely park them.)
• Beat 2 — Demand exhibits. "A case is only as good as its exhibits. Give me the three numbers closest to this decision — revenue it touches, hours it eats, what it costs today." No numbers = the first deliverable IS producing the exhibits (that's a teardown).
• Beat 3 — The board. Place their decision on Moat & Meter out loud ("high P&L impact, but anyone can copy it — that's Commons; here's the enclose move"). Name the failing Factory station. This is the moment they feel the method.
• Beat 4 — The cold-call reversal. "If you had to commit today, what would you do?" Their answer surfaces every objection and anchors the write-up. Never skip — people defend plans they authored.
• Beat 5 — The assignment. "Here's how we work: we write your case up properly — 60 minutes, three Factory Tickets, each with a dollar meter. That's the $397 Teardown, or it's free."

Within 24 hours: log the case in the intake engine below, write the one-line lesson, and check the Lessons Board before your NEXT call — that's the compounding loop.

HBS Case 9-625-015 (Bojinov & Lakhani, 2025) — the enterprise blueprint our Island Factory scales down. The five lessons we carry into client conversations:

• Leverage over headcount. The giants solved AI demand with reusable tooling, not hiring. → When a client says "we'd need to hire someone," show the parts bin first.
• Assemble; never marry a vendor. Own the connective tissue, rent best-of-breed components, keep every piece swappable. → Counter any monolithic "AI platform" subscription they've been pitched.
• Prove value inside the funding window. Ship imperfect, show a win, earn the next round. → Every proposal runs Prove→Run→Grow gates.
• Adoption is engineered. The giants budget entire cycles for training, champions, and honest usage metrics. → Station 4 is a line item in every build, never an afterthought.
• Fit-to-context. The blueprint says skip components your context can't use — even ones the tech giants swear by. → When asked "shouldn't we have X like Google?", run the context test before the checkbook.

🎓 Take this to class

The setup. July 2026. The protagonist owns an AI-operated e-bike rental on a small Caribbean island. The machine works — a 24/7 AI concierge quotes and closes bookings, payments collect themselves — but it's low season and cash is tight. A guest just paid $93.50 for a 12-hour ride and complained; the owner's gut says "12 hours should be $35." Eight bikes sit idle today. The decision (clock: this weekend). Cut prices across the board? Rebuild the price ladder? Spend the last marketing dollars on paid ads? Or hold and ride out the season? The exhibits.

• A — Monthly revenue: the fall began BEFORE the June price increase (peak-month revenue roughly halved year-over-year, with the decline underway pre-hike).
• B — Fleet utilization: ~13%. The constraint is not capacity.
• C — Traffic sources: nearly all visits "direct"; effectively zero paid or organic discovery.
• D — The funnel gauge reports 68% checkout abandonment — but the tracking only fires on one return path, so the number is structurally unreliable.
• E — A premium fleet tier exists with ZERO bookings ever; it's configured as overflow-only, invisible to buyers.

Discussion questions. (1) Is this a pricing problem — what do Exhibits A and C say? (2) Which gauge do you trust, and what's the cheapest way to find out? (3) Plot "cut all prices 40%" on Moat & Meter — where does it land and why? (4) What do you do Monday morning? 🎓 Take this to class

The setup. July 2026, the same island operator. Three weeks earlier the business quietly added a ferry-ticket concierge: tourists face a government booking app that shows "sold out" while seats still exist at the terminal window, so a real person buys the official tickets for them, pays the municipal eco-tax, and texts the tickets to their phone. The true do-it-yourself cost is $6.50 all-in — and it's public. The owner priced on instinct: $14.55 flat, then a hastily-built group ladder. Every single order ever quoted has paid, including a 10-person group and a buyer three days out. The owner: "this business is larger than I originally thought." A new SEO page just went live; traffic is coming. What's the strategic price? The decision (clock: the traffic arrives priced-as-is). Raise across the board? Hold and ride the volume? Or cut — the ferry buyer is a captive prospect for the $60/day e-bike rental, so should the ferry be bait? The exhibits.

• A — The order book: 5 paid orders in 3 weeks, 100% of quotes converted, zero price objections ever recorded. Party sizes: 1, 4, 4, 4, 10.
• B — True DIY cost: $6.50 round-trip all-in, printed on the tickets the customer eventually receives. The official portal sells only a fraction of each sailing online — "sold out" usually isn't.
• C — The alternatives: $80–150 flights, $100+ day tours, a local taxi service charging roughly a $10-per-ticket buying fee, and the government checkout itself, which charged one reviewer $22.40 in fees for 4 tickets.
• D — The quoting machine: an AI agent quotes every price in chat, and a +25% rush multiplier applies inside 5 days. The current ladder has half-dollar tiers: $14.50 × 1.25 = $18.13.
• E — On the same page: one-way $7.99, solo round-trip $22. Two one-ways = $15.98.

Discussion questions. (1) What does Exhibit A actually tell you — and is it good news? (2) Where is the fairness ceiling when the $6.50 face value is discoverable (Exhibits B, C)? (3) Which party sizes do you raise and which do you protect, and what does the e-bike funnel have to do with it? (4) What breaks in Exhibits D and E that has nothing to do with willingness-to-pay? (5) You ship a new ladder Monday — in what ORDER do the config flip, the website copy, and the AI prompt change, and why does the order matter? 🎓 Take this to class

The protagonist. Priya Raman, Chief Financial Officer of Ardent Specialty Group (ASG) — a fictional $9.2B-revenue specialty commercial insurer whose real crown jewel is a proprietary corpus of nearly five decades of underwriting and loss data that no rival can rent. Priya is 47. Eighteen months ago she personally championed ASG's $41M-a-year enterprise-AI program, "Project Lodestar," and put her name on the business case. Her predecessor as CFO was quietly pushed out after a $200M "digital transformation" no one could ever tie to a number — and Priya knows the board remembers. Her credibility, and the program, are now the same asset. She is the institution's measurer-in-chief, now asked to produce the one number everyone treats as a routine finance exercise — the audited ROI of the AI spend — that she is finding harder to honestly produce the deeper she digs, with the CEO certain it is enormous and the audit chair certain it is near zero. The setup. Ardent Specialty Group underwrites the commercial risk standard carriers won't touch; its edge has always been judgment — 47 years of proprietary loss experience that lets it price what rivals can't read. Eighteen months ago CFO Priya Raman sponsored Project Lodestar, a $41M-a-year enterprise-AI program spanning underwriting, claims, customer service, and knowledge work, most of it running on a frontier model provider's API, and she signed the business case. By every internal account it is a hit: thousands of daily users, adoption dashboards lit green in every direction, a benefits deck claiming $118M of annual value. Two forces are now colliding over her head. CEO Ellery Boyd — an evangelist who has already told two sell-side analysts that ASG will "materially increase" its AI spend — wants FY27 doubled to $85M. Gordon Vale, chair of the board's Audit & Technology subcommittee and a lifelong numbers-hawk, read the same all-green dashboards, looked at the combined ratio sitting exactly where it sat the day Lodestar launched, and asked for one deliverable before he approves another dollar: "the audited ROI of what we've already spent." The morning she sat down to build the slide, Priya opened a competitor's press release announcing the one AI capability in her portfolio with a clean, booked P&L number — now shipped as an off-the-shelf product any carrier can buy. The subcommittee meets in eleven days. She has a benefits slide whose footnotes she cannot unsee, a rising invoice she cannot argue with, two vendor calculators that disagree threefold, one genuinely booked win a competitor just matched, and three doors — none of them clean. Eleven days to decide which number, or which argument, she is willing to put her name to. The decision (clock: eleven days). Priya must walk into the subcommittee in eleven days with a recommendation on the FY27 AI budget. Three doors, each with a real case and a real cost. DOOR 1 — Renew and expand to $85M, backing the CEO's thesis that AI is a strategic imperative and that ASG should press its advantage — the stranded intake pilot, the loss-corpus engine, and a firming market where speed-to-quote wins share — before rivals do; the cost is that she must put her name, under audit, on a blended ROI number she cannot fully defend. DOOR 2 — Cut to ~$18M, backing the audit chair: fund the one booked winner and the cheapest proven use-cases, stop the rest; disciplined and auditable, but it forecloses the corpus engine and the stranded pilot before either can prove out, and concedes the program she championed grew faster than its evidence. DOOR 3 — Decline to produce the single blended ROI figure the board asked for, tell them why it cannot be honestly produced, and re-govern the entire $41M portfolio on the two axes that CAN be measured — unit cost and durable advantage — funding some use-cases harder and killing others, without ever handing the board the one number it requested. The first two doors give the board what it asked for. The third refuses the question eleven days before the board votes on her budget and, implicitly, on her judgment — the same board that funded her on a promise of measurement. The exhibits.

• Ex. 1 — Project Lodestar: The Portfolio and the Ask. FY26 AI program spend: $41M. Composition — frontier-model + inference ('the Meter') $16M; vendor platform licenses & seats $11M; internal AI platform + 28 ML/eng FTEs $9M; change management, training & prompt-ops $5M. Seven initiatives: (1) underwriting risk copilot on a general model; (2) claims-triage assistant; (3) 'Ardent Assist' customer-service copilot; (4) policy-document extraction; (5) submission-intake pilot; (6) knowledge/search for actuarial-legal-finance; (7) a loss-corpus risk engine trained on ASG's own 47-year book, still in shadow mode. CEO's FY27 proposal: $85M (+107%). Audit chair's counter: $18M ('fund what's proven, stop the rest'). Priya's mandate before the vote: 'the audited ROI of monies already deployed.' Note for the reader: nowhere in the $41M is there a line item for a control group, a counterfactual, or an independent measurement of benefit — every dollar sits on the COST side of the ledger. Note two: initiative (5) hit 7 of 7 pilot KPIs eleven months ago and is still a pilot — no P&L owner ever funded the $2.3M mainframe integration to scale it.
• Ex. 2 — The Meter: Monthly Model + Inference Invoice, FY26. The one number in the deck no one disputes, because it arrives as an invoice. Monthly frontier-model + inference cost: Jan $0.85M · Mar $1.10M · Jun $1.45M · Sep $1.80M · Dec $2.25M. Full-year $16M; December exit run-rate ≈ $27M annualized (+165% over the January pace). Driver is usage: every new seat, every longer prompt, every 'let's also have it summarize the file' adds tokens — the cost rises with success. Under the hood: 84% of calls route to the frontier ('flagship') model, including the routine ~80% a model one-tenth the price handles identically; under 3% of calls are cached; no workload carries a spend cap, so the first anyone sees of a month is the bill. This is the ONLY figure in the entire benefits-and-cost package a financial auditor could sign without a single behavioral assumption — and the only one that grows every month whether or not the business improves.
• Ex. 3 — The Benefits Case: '$118M of Value on $41M Invested'. The slide the program office is proud of. Headline: '$2.9 returned for every $1 — 188% ROI.' Build: Underwriting — 920 underwriters × 3.8 hrs/wk saved × 45 wks × $92/hr loaded = $14.5M. Claims — 1,400 handlers × 2.6 hrs × 45 × $68 = $11.1M. Customer service — 2,400 agents × 1.9 hrs × 45 × $54 = $11.1M. Other knowledge work $9.0M. 'Loss-ratio improvement from better risk selection' $52.0M. 'Cross-sell / retention lift' $12.0M. 'Capacity we avoided hiring' $8.0M. Total $117.7M. The footnotes (8-pt, back page): survey response rate 38%, respondents self-selected; headcount is FLAT to +2% YoY — no role eliminated, 'hours saved' never converted to a dollar out; the $52M loss-ratio line has no control cohort and coincides with a firming rate environment and a benign catastrophe year (confounds larger than the claimed effect); 'capacity avoided' $8M overlaps the hours-saved lines (double-count). Strip the four unauditable lines and the checkable, redeployed-to-cash benefit is approximately $0. But 'unfindable' is not the same as 'zero': the program's one auditable, booked, cash-releasing return lives on a different slide entirely (Exhibit 6 — the extraction win, $7.9M actually removed from the P&L when the outsourced contract was cancelled). Priya's defensible FLOOR is therefore real but small — on the order of $7.9M of booked cash against $41M spent — a number she could survive an audit with, and the ammunition for anyone who would fund proven winners over the blend. Quotable: '$118M of value, none of which the CFO can find in her own ledger.'
• Ex. 4 — Two Vendor ROI Calculators, One Use-Case. Claims-triage assistant (annual run cost $4.5M). Priya asked two platform vendors to model it; both dropped ASG into their standard ROI template. Vendor A (frontier-platform reseller): baseline handle time 42 min → 31 min (−26%), adoption 75%, wage $68 loaded ⇒ savings $19.6M/yr, net ROI +336%. Vendor B (rival platform, same underlying model): baseline 42 → 36 min (−14%), adoption 55%, wage $61, less a 9% 'rework/reversal' factor for AI errors ⇒ savings $6.4M/yr, net ROI +42%. Same use-case, same model, a 3.1x spread in claimed savings — driven entirely by unobservable inputs: the true baseline delta and real adoption. Then Priya's FP&A team read the fine print. BOTH templates default to 2,100,000 claims/yr — a personal-lines volume. ASG's specialty book runs ~60,000 claims/yr (see Exhibit 6: 890K claims across 47 years, ~$92K average severity). Correction 1 — re-run Vendor A's OWN optimistic assumptions on ASG's actual volume and savings collapse to $0.56M — net ROI −88% on the $4.5M it costs to run. Correction 2 — even granting the vendors their 2.1M template, hold adoption and delta at plausible-conservative (adoption 45%, delta −8% ⇒ 42→38.6 min, wage $61) and savings fall to $3.2M — below run cost — net ROI −28%. Same use-case, four honest recomputations, a swing from +336% to −88%. FP&A's memo: 'We can audit the cost to the dollar. We cannot audit a single one of the four inputs that move the answer — book volume, baseline delta, adoption, error rate. There was never a control group; no one ever ran the claims desk without the tool. (Net ROI throughout = (savings − run cost) / run cost.)'
• Ex. 5 — The All-Green Gauge vs. the Flat Needle (two panels). PANEL A — 'Ardent Assist' customer-service copilot, the dashboard the CEO screenshots for the board: seats provisioned 2,400/2,450 (98%), weekly-active 91%, AI suggestions accepted 2.1M/mo, CSAT on AI-touched chats 4.6/5.0 (baseline 4.4). GREEN on every tile, eighteen months running. The business metric it was funded to move — cost-to-serve per claim — went $214 (FY25) → $211 (FY26): −1.4%, inside the noise, against a business case that promised −12%. First-contact resolution 71% → 71%. The CSAT 'lift' is measured only on completed chats; abandoned sessions are never surveyed. PANEL B — the underwriting copilot was credited in the deck with a −14% quote-cycle-time improvement. But ASG's Mountain West region ran two quarters with the copilot switched OFF (a systems freeze) and posted the same −14% — because a new rules engine shipped to every region in the same window. The gain credited to AI appeared in the one region AI never touched. Vale's handwriting in the margin of his printout: 'Green everywhere it doesn't cost us anything.'
• Ex. 6 — The Washout and the Starving Fortress. THE WASHOUT (initiative 4). AI policy-document extraction cut per-submission processing cost from $18.40 to $6.10 across 640,000 submissions/yr ⇒ measured, auditable savings $7.9M/yr — and the outsourced BPO contract was actually cancelled, so the money is booked. The cleanest, most honest before/after in the whole deck. Then, five months later, two rival carriers announced the identical capability — same document-intelligence vendor, same frontier model, now a standard product SKU any carrier can buy. The speed edge that let ASG quote mid-market accounts faster evaporated as the market repriced to the new normal; the $7.9M cost-avoidance survives on ASG's P&L but the competitive advantage did not. Advantage half-life: ~2 quarters, decaying. THE STARVING FORTRESS (initiative 7). The loss-corpus risk engine trained on ASG's proprietary 47-year book — 2.7M bound policies, 890K claims — data a rival can never rent even while renting the identical model. It runs in shadow mode, not yet authorized to bind, so today's measured P&L impact reads ~$0; its $1.9M budget (5% of the program) is first on Vale's chopping block for 'no measurable ROI.' Back-test signal: it would have re-priced or declined a $1.8B sub-segment that ran a 112% loss ratio last year — a modeled 0.3–0.6 point improvement to ASG's overall loss ratio, on the order of $30–55M, none of it booked because the engine can't yet bind. The single most MEASURABLE return in the portfolio is a Commons every competitor rents from the same shelf; the single most DURABLE opportunity is the one with ~$0 measured ROI — and the deck credits an unprovable $52M (Exhibit 3) to the general copilot while starving the engine that could actually earn it.

Discussion questions. (1) Gordon Vale wants 'the audited ROI of the AI program' as a single number. Can Priya give him an honest one? If yes, derive it from the exhibits and then defend every assumption an auditor will attack. If no, what does she put on the board's screen instead — and does 'we can't measure it,' spoken by the CFO herself, survive contact with a board that just spent $41M and is watching the person who championed it? (2) The CEO reads Exhibit 5 as a triumph (all green); Vale reads the identical slide as a failure (flat needle). Who is right? State exactly what you would have to believe for each reading to hold — and name the single number that would settle the disagreement. (3) Exhibit 6's extraction win is the cleanest, best-measured return in the case: a real, booked $7.9M — and a competitor matched it in two quarters. Was it a good investment or a bad one? Would your answer change if ASG had never measured it at all? What does that reveal about what 'measurable' is actually worth? (4) Make the strongest possible case for DOOR 1 — expanding to $85M — using only evidence in these exhibits: the intake pilot that hit 7 of 7 KPIs and was never scaled, the loss-corpus engine's back-tested $30–55M, the genuinely booked extraction win, and a firming market where speed-to-quote wins share. Is 'press the advantage before rivals do' a strategy or a rationalization — and what single fact in the case most threatens the expansion argument? (5) The loss-corpus engine (Exhibit 6) has ~$0 measurable ROI and is first on the chopping block; the code/knowledge assistant carries a confident self-reported productivity claim and is safe. On the evidence in this case, defend the budget decision that CUTS the engine and keeps the assistant. Then argue the exact reverse. Which is right — and what does the difficulty of choosing tell you about ROI as an allocation tool? (6) Priya takes Door 3 and tells the board the blended ROI number is unmeasurable by construction. Vale replies: 'Every capital request I've approved in thirty years came with a return. You're telling me AI gets a pass?' What is her one-sentence answer — and is it strategy, or an excuse that ends her tenure the way it ended her predecessor's? 🎓 Take this to class