Sources
Show notes: my writeups of the talks, interviews, and videos behind the essays. Not essays themselves, just the source material, distilled and weighted.
GTM data architecture: the warehouse is the hub
Notes on the gtm-data-architecture skill at gtmskills.com, a field manual for revenue operators who are not data engineers: why warehouse-native won on cost, AI, and speed; the reference architecture with reverse ETL as the activation path back into the selling tools; honest thresholds for when not to build it; and four named anti-patterns ending in the sentence every agent program should frame, agents amplify bad data. Also notable: the page itself ships as an installable skill for AI agents.
Bain: posture, domains, and the price of waiting
The strategy altitude of Bain's proprietary-intelligence series, Decisions 1 and 2 plus the closing chapter, in one note: why wait-and-see feels prudent and rarely is, the two-decades capability hole that money cannot quickly fill, the three tests that pick three to five domain bets, the in-year-ROI trap that funds the wrong activities, and the closing checklist, personal CEO commitment, fewer bigger bets, and a learning system, with a window that will not stay open.
Data contracts: an API, but for data
Notes on datacontract.com, the interactive guide and editor for the Open Data Contract Standard: the definition worth keeping, a data contract covers ownership, structure, semantics, quality, and terms of use between a producer and its consumers; the rare standards story, the site deprecated its own specification in favor of the Linux Foundation's ODCS 3.1; the seven-section anatomy of a contract; and why machine-testable quality rules are the piece that matters for AI agents.
Bain: your governance is almost perfectly wrong for AI
Decision 7 in Bain's proprietary-intelligence series says the quiet part about why transformations stall: existing governance is good at what it was designed for and almost perfectly wrong for AI, too slow, too backward-looking, too sequential, too control-minded. My notes on the two-motions answer, run the business and change the business in parallel, token costs as the new unit economics, portfolio funding over project ROI, and the bluntest line in the series: the CEO must own AI risk, not monitor it.
Bain: proprietary data is a moat that leaks
Decision 3 in Bain's proprietary-intelligence series makes the data case with an honest twist: your data cannot be synthesized by a competitor, yet its value decays as conditions evolve, and the live signal your operations generate daily is what keeps it fresh. My notes on the decaying-moat framing, the instruction to do only the data work the domain bets need, and why running agents through governance built for human timescales is, per Bain, the single most common reason transformations stall.
Bain: design the operating model for speed
Decision 5 in Bain's proprietary-intelligence series argues strategy and architecture mean nothing without an operating model built to deliver at pace: four requirements for speed, workflow and workforce changed in parallel rather than in sequence, a tinkerer network as the bottom-up engine, and a blunt test for the top, CEOs who delegate the hands-on experience never close the gap between what they say about AI and what they understand.
Bain: design the transformation to compound
Decision 6 in Bain's proprietary-intelligence series is the one the whole thesis rests on: the operating model delivers today, the learning system defines proprietary intelligence over time. My notes on the three layers where learning can happen and why the harness beats model selection, the 400-experiments case for machine-scale search, the memory layer that lets the tenth agent start ahead of the third, and the closing inversion, when the system absorbs the automatable work, the bottleneck becomes human judgment.
Bain: "Proprietary intelligence" is how to win with AI
Bain's June 2026 piece argues most CEOs are managing pilot portfolios while believing they lead AI transformations: 82 percent report at best partial results, and the barriers are organizational, not technical. My notes on the report, its definition of proprietary intelligence as unique data plus encoded workflows plus a learning architecture, the seven board-level decisions it says cannot be delegated, and why it claims AI advantage compounds from day one in a way earlier technology waves did not.
Bain: rent the model, own the orchestration layer
Decision 4 in Bain's proprietary-intelligence series argues the most important technology decision in an AI transformation is not which model to use: models go obsolete in months, so advantage concentrates in the orchestration layer you build around them, encoded know-how, access boundaries, and governance. My notes on the rent-versus-own line, why agents break the assumptions traditional software governance was built on, the 20/60/20 split that makes evals the real work, and the obsolescence rebuttal.
Bain: what an "AI-native enterprise" actually is
Bain's companion explainer to its proprietary-intelligence thesis defines the AI-native enterprise and draws the contrast that matters: conventional companies run pilot collections on fragmented data with one governance motion; AI-native ones run compounding intelligence on a shared semantic layer with two. My notes on the definition, the six-dimension comparison, the 20/200/2,000 leadership cascade where transformations actually break, and the three traits of the organizations furthest along.
Show notes: AI is not killing vertical SaaS
Deepak Sindwani of Wavecrest Growth on why software plus AI plus data beats AI-instead-of-software, and how growth equity actually works for $5 to $20M founders.
McKinsey: "AI data readiness is the key to scaling impact"
McKinsey's June 2026 report argues that data, not the model, is the constraint on scaling AI: only 7 percent of companies have fully scaled it, and more than two-thirds of top performers call data their main obstacle. My notes on the report, its definition of AI data readiness as a governed, reusable foundation, the four shifts stalling scaling, the six data disciplines it says to evolve, and its call for chief data officers to own the standards and control plane rather than the pipelines.
Show notes: The hottest AI job is an old job, renamed
Jean Lee, twenty years in tech, on whether forward deployed engineer is really the highest-paid job in AI or a familiar pattern: a hot new title that is mostly an old customer-facing role repackaged, with the word engineer chosen to pull builders into work they would otherwise refuse, and a durable skill hiding under a temporary name.
Show notes: Do things that don't scale, at scale
Bob McGrew, who helped invent the forward deployed engineer model at Palantir and later ran research at OpenAI, gives the operator's playbook: why the strategy is a last resort not a default, the gravel road the engineer lays and the superhighway the product team paves, the two team roles, the margin curve that proves it isn't consulting, why AI agents have no incumbent to copy, how to price an outcome, and why the real opportunity is the gap between what AI can do and what anyone has adopted.
Show notes: One title, four companies, four different jobs
A South Park Commons panel with forward deployed leaders from Ramp, Nominal, Dataland, and OpenAI shows the same title meaning four different jobs: Ramp's defensive sword and shield, Nominal and Dataland's bespoke agent work, OpenAI's frontier push. Where four very different companies independently agree is the real signal, and where they disagree, on whether AI collapses the role into one person or splits it into specialists, is the live question.
Show notes: The title means nothing, and that is the point
Natalie Meurer of Sierra, a Palantir forward deployed engineer from the origin years, gives the dirty secret: the label has been stretched to cover so many jobs that it means nothing, the real work began as 2 a.m. DevOps babysitting not helicopter heroics, and the jobs to be done only ever accumulated. What survives when the title dissolves is a single invariant, accountability to the customer, and as code gets cheap that invariant is spreading until nearly everyone becomes a forward deployed engineer.
Show notes: A consultant who can actually build
Soleyman, an independent tech consultant, gives the plainest working definition of the forward deployed engineer: a consultant who can actually engineer and build, sitting in the gap between a slick AI demo and a working system inside a messy real business. The value here is the how-to: the six-step job from understanding the real problem to feeding the learning back into the product, the skill stack it demands, and the career advice to build the capability, not chase the title.
Show notes: The hard part is not the build, it is the trust
Colin Jarvis, who runs forward deployed engineering at OpenAI, gives the operator's field report from a frontier lab: trust, not capability, is the binding constraint, the discipline is determinism-first with evals as the starting artifact, one customer's fix becomes an open-source primitive and then a shipped product, and the whole function is run as a research feeder that can refuse services money because OpenAI monetizes the model elsewhere.
Show notes: Forward Deployed Engineering 101
Kevin Bai (Anthropic, ex-Palantir, founding Rippling FDE) on the forward-deployed model: sell the outcome not the software, the one quadrant that actually needs it, a design partnership scaled to the enterprise, the platform that keeps it from becoming a dev shop, and why every agentic product in 2026 quietly puts its customers in Palantir's old shoes.
Show notes: Who's afraid of Chinese models?
Ben Thompson's case that the Kimi and Qwen panic is mostly an economics misread. Marginal costs are back, intelligence is the commodity not tokens, and today's prices are a compute-shortage umbrella. The one real fear is cybersecurity.
Show notes: Agents are the new SaaS
Greg Isenberg's playbook for building an agent business: sell the job done, not the tool. Find a workflow with a paycheck, shadow the human, ship the smallest useful agent, prove it with evals, then productize the pilot.
Show notes: Why AI won't make you rich in 2026
Alex Hormozi on why AI-maxing businesses are not making more money: AI raised capacity, not judgment, and it did not cancel any other form of leverage.
Show notes: What SaaS buyers actually want in 2026
Rob Walling and Einar Vollset on the five moats private equity now requires before they will even look at a SaaS company.
Hormozi's AI test: are you making more money?
Alex Hormozi spent two hours on The Diary of a CEO arguing that founders are misallocating the AI boom: wrapper companies the models will eat, years of an $11,000 bill spent automating a non-constraint, and 'token maxing' that never touches the P&L. His one-question test, where he says value survives when intelligence is cheap, and the deliberately unglamorous program he prescribes instead, with his frameworks drawn out.
Reinforcing clip: Hormozi, AI does dumb things faster
A 3-minute MoreMozi clip that restates the two-post Hormozi thesis in sharper form. Not its own post: woven into the existing pair as independent corroboration. Where each line landed.
Show notes: Hormozi on DOAC, the AI misallocation episode
Business and AI themes only from Alex Hormozi on The Diary of a CEO (July 2026): the stop-chasing-AI thesis, reality as the moat, the value equation, pricing, talent-as-funnel, and the numbers he cites, all labeled as his claims. Personal segments excluded by design.
Doing the wrong things, faster
Alex Hormozi asked his wealthiest friends how much AI they actually use. Almost none, it turned out, and their businesses keep compounding anyway. The gap between token bills and profits has an old explanation: AI raised everyone's capacity to work, and capacity without judgment flows straight into the backlog of things that never deserved doing.