One Big Idea
Pilot purgatory is a capital-allocation failure, not a technology one. Stop funding pilots — start managing the portfolio.
Here is the number that should end the debate. Last year, a reported forty-two percent of organizations abandoned the majority of their AI initiatives before production — up from seventeen percent the year before. The rate more than doubled in twelve months, and Gartner projects more than forty percent of agentic projects cancelled by 2027.
Most banks fail here not because the models are weak, but because they fund AI as dozens of disconnected pilots with no shared thesis and no rule for when to stop. Capital gets trapped in proofs-of-concept that never ship. A bank would never fund a loan book this way — yet that is how most fund AI.
The Insight
The discipline already exists inside every bank — it is just applied to the wrong asset. A loan book is underwritten: each loan carries a thesis, and the book carries a rule for writing off the ones that fail. An AI book, in most banks, carries neither. Nothing states which line of the profit-and-loss statement a use case will move, and nothing fires when it does not.
That is why the abandonment rate is a capital-allocation number rather than a technology one. The models did not fail; the funding discipline was never there to begin with.
Stop funding pilots. Start managing the portfolio. A profit-and-loss thesis is the entry ticket; a mechanical kill rule is the exit.
Framework of the Week · The AI Use-Case Portfolio
Score every use case on two axes — Value (only what reaches the P&L, read by value zone, no vanity metrics) against Feasibility (data, operating model, sourcing) — and route it to one of four verdicts:
- Scale — high value, high feasibility; fund without constraint.
- Gated Pilot — high value, low feasibility; capital released in tranches.
- Quick Win — low value, high feasibility; deploy fast on a vendor tool.
- Kill / Park — low value, low feasibility; decommission before it bleeds.
Then run the book on a rhythm — the Portfolio Test. A P&L thesis is the entry ticket: no use case is funded without naming the line it will move. Then four gates — feasibility, control (shadow mode), launch, and value realization. If gate four fails, a mechanical kill rule fires automatically, so stopping is a default rather than a decision someone has to champion. And the book is repriced quarterly, like credit risk.
The full portfolio, with both axes and all four verdicts drawn out, lives in the Frameworks library.
Use Case · The proof
The banks that got out of pilot purgatory did it by managing a book, not a queue. JPMorgan reportedly runs more than four hundred fifty use cases in production. DBS unified its data first, then scaled more than fifteen hundred models, cutting time-to-market from fifteen months to under three (reported).
The gates are not theatre either. One Gated Pilot ran in shadow mode to about ninety-four percent agreement over about one hundred twenty-seven thousand transactions before it was allowed to go live. And the Kill verdict has a cautionary twin: a firm outside banking that spent three years on an initiative before decommissioning it, for want of a Gate Four.
Risk Note
The regulation is already a portfolio gate, whether or not you built one. The revised United States model-risk guidance traces data lineage; the EU AI Act makes credit scoring high-risk; Vietnam restricts data routing. Read together, they say the same thing the portfolio says: a use case has to be able to account for itself before it reaches a customer — and the ones that cannot should never have been funded past a gate.
Where the Series Comes Together
This is the capstone, and the two axes are where the whole series lands.
- Value is the Boardroom Equation and the Value Zones — what actually reaches the P&L, and where it lands.
- Feasibility is the Operating-Model Multiplier and the Build-Buy-Partner Grid — whether you can carry it in production, and whether you should build it at all.
- A Scale verdict then deploys through the Agent Army operating model, governed by the four Guardians.
The portfolio is the engine that runs all of it.
Stop funding pilots. Start managing the portfolio.
Latest Video
This week's briefing — Stop Funding AI Like a Science Fair — walks the two axes, the four verdicts, and the Portfolio Test you can run on Monday: name the P&L line for every use case, score value and feasibility, route the verdict, and let the kill rule fire on its own when gate four fails. Then reprice the book quarterly, the way you already reprice credit risk.
Watch: youtu.be/wH_ICWsaqvY
The full deep-dive — pilot purgatory, the two-by-two, the four gates and the quarterly rebalance — is in Stop Funding AI Pilots — Run the Portfolio Test.
The free five-page playbook in the Frameworks library turns it into a worksheet for your own AI portfolio.
Reply and tell me which use case in your book has no P&L thesis and no kill rule — most portfolios have several, and they are the ones quietly consuming the budget. I read every response. Forward this to a banking executive about to fund another disconnected pilot.
Was this forwarded to you? Subscribe to The AI Architect Letter — free, every Saturday.
Minh Tran · AI Business Architect · LinkedIn · Workshops & advisory: aibusinessarchitect.ai