From Models to Money (Davood Shamsi and Robert Luenberger)（Z-Library）

# From Models to Money: A Leader's Guide to AI Unit Economics and the Total Cost of Intelligence ## 【One-Line Pitch】 A practical field guide for CTOs, engineering leaders, and technical product managers who want to stop treating AI as a technology experiment and start governing it as a capital allocation problem—with real frameworks for moving from pilot to production profitably. ## 【Book Arc】 - **Opening (~0%–9%)**: Diagnoses the "pilot paradox"—why 88% of AI proofs-of-concept never reach production and why successful demos routinely produce failed programs. Establishes the core reframe: technical progress ≠ economic progress. - **Early (~9%–28%)**: Introduces the three-circle Venn diagram (business problem, organizational fit, scalability/reliability) and the four failure areas (A–D), illustrated with case studies like Amazon Just Walk Out, Zillow's iBuying collapse, and Google Flu Trends. Builds the "pyramid of evidence" framework for climbing from demo to production. - **Early (~28%–38%)**: Covers the hidden output of pilots—institutional learning—and introduces the Assumption Register as a living document for tracking tested vs. unknown assumptions. Addresses data realism, legal constraints, and the ownership question ("who gets the call at 3 a.m.?"). - **Middle (~38%–53%)**: Distinguishes between efficiency pilots (bounded, measurable, short-cycle) and new offering pilots (transformative, longer evidence periods, deeper organizational analysis). Uses PayPal fraud detection and Morgan Stanley's AI assistant as contrasting success cases, with UnitedHealth's nH Predict as a cautionary tale. - **Middle (~53%–end)**: Explores portfolio management—must-win vs. innovation bets—and tackles the measurement problem for generative AI systems where outcomes are open-ended and success criteria are partly subjective. The excerpts suggest a closing emphasis on building evidence with internal users before approaching product teams. ## 【Key Takeaways】 - **The demo is real; the inference from it is false** (Opening): A controlled pilot proves technical possibility, not business viability. The gap between a rehearsed performance and a production system is where most AI programs die—treating a technical proof as a business case is the root failure. - **Three independent questions define pilot feasibility** (Early): Business problem, organizational fit, and scalability/reliability form a Venn diagram. Pilots fail when teams assume all three are satisfied without testing each dimension explicitly. - **Area B failures look like success** (Early): Amazon Just Walk Out appeared autonomous but required ~700 manual reviews per 1,000 transactions by ~1,000 contractors—a permanent, invisible operating cost. The pilot didn't lie; the organization extrapolated without evidence. - **Area C failures are quiet capital sinks** (Early): Zillow's $569 million write-down came from repurposing a display-level valuation model as a capital-at-risk pricing engine. Systems that align with managerial incentives but change no real decisions absorb talent and capital while producing zero ROI. - **Skipping evidence layers doesn't accelerate production—it accelerates failure** (Early): Google Flu Trends skipped validation that the capability could change real health decisions and overestimated peak flu cases by 2x. The pyramid of evidence (technical demo → business alignment → scalability → production) must be climbed in order. - **Legal review is a scalability test, not a compliance formality** (Early): The Trade Desk's Mark Davenport calls legal "the number one killer" on the path to production. Contractual data rights can prohibit commingling data that demos treat as fungible—involve legal before building, not as an end gate. - **Ownership must be named before scaling** (Middle): "Who gets the call at 3 a.m.?" separates demos from production systems. GE Predix failed partly because no single team owned end-to-end outcomes across silos. A named individual with authority over deployment, monitoring, and rollback is non-negotiable. - **Efficiency and new offering pilots need different management** (Middle): PayPal's fraud detection (efficiency) had a clear baseline and measurable counterfactual; Morgan Stanley's AI assistant (new offering) required an eval framework built with advisor input before deployment. Managing both with the same tools and timelines is a mistake. ## 【Reading Tips】 - **Deep-read the Opening (~0%–9%)** for the failure statistics and the core reframe—this is the book's thesis and will anchor everything else. The IDC, MIT, and McKinsey data are worth remembering for internal conversations. - **Skim the case studies in Early (~9%–28%)** but focus on the framework they illustrate: the four failure areas (A–D) and the pyramid of evidence. These are the book's most reusable mental models. - **Pay special attention to the Assumption Register (~28%–38%)**—this is the most immediately actionable tool. The categories ("tested," "accepted risk with price," "unknown") and the ownership requirement are directly implementable. - **The Middle section (~38%–53%)** on pilot types is where the book moves from diagnosis to prescription. The efficiency vs. new offering distinction will change how you resource and time-box pilots. - **Watch for the generative AI measurement problem (~53%+)**—the excerpts cut off mid-discussion, but this is likely the book's most forward-looking contribution. If the treatment feels incomplete, supplement with the authors' discussion of eval frameworks from the Morgan Stanley case. ## 【Coverage Limits】 The excerpts cover roughly the first half of the book (through ~53%), with the generative AI measurement discussion cut off mid-sentence. Later chapters on real options for learning systems, risk-adjusted value, and minimal credible evidence are not covered in this guide. ##