One Big Idea
In financial services, by some estimates more than eighty percent of in-house AI builds fail — and they don't fail because the algorithms are wrong.
Here is a number that should change how your next artificial-intelligence budget is argued. More than eighty percent of in-house AI builds fail in banking, against roughly a quarter to a half for ordinary technology projects. Billions sit stranded in proof-of-concept, returning nothing on the profit-and-loss statement.
Most banks fail here because the data and the operating model underneath the model cannot carry it in production. A fraud detection model with near-perfect accuracy in a sandbox is useless if it cannot reach the core banking ledger in real time. These are not AI failures; they are infrastructure failures wearing an AI label.
The Insight
The economics, plainly. A custom build runs twelve to twenty-four months before it returns anything; a commercial platform delivers in three to six. Across three years, buying costs roughly half. And your own foundation model costs seventy-eight to one hundred million dollars to train from scratch, against a few thousand to fine-tune an existing one.
For a bank, building your own model is almost never the answer.
Stop building what you can buy. Build only what you alone can own.
Framework of the Week · The Build-Buy-Partner Grid
Place every capability on two axes — Strategic Differentiation (moat vs commodity) against Internal Readiness (can you sustain it, not just pilot it) — and route it to one of four verdicts:
- Build — only where you own the data moat: fraud detection on your own ledger, credit underwriting on your own data.
- Partner — the edge is real but you can't sustain it alone; co-develop and keep the workflow.
- Buy — a commodity function; purchase it, carry none of the maintenance.
- Assemble — fine-tune or retrieve on a bought model rather than train one.
Most banks invert it: they build the commodities and under-fund the moat.
The full grid, with both axes and all four verdicts drawn out, lives in the Frameworks library.
Use Case · The proof
HSBC partnered for AML network analysis. Morgan Stanley partnered for advisory and wrapped the bought model in its own guardrails. DBS generated close to a billion Singapore dollars by orchestrating more than fifteen hundred models — assembled, not trained from scratch. And across SR 26-2, the EU AI Act, DORA, and Vietnam's data decree, the message is identical: outsourcing the model never outsources the liability.
Risk Note
Regulated finance adds three forces to the grid. Liability stays home — buying a model does not move the regulatory responsibility, so demand audit logs and explainability. Speed beats opportunity cost — three months often beats twenty-four. Lock-in is real — architect for model agnosticism so you can swap providers as the economics shift. And buying is the easy part: the value is redesigning the operating model around the tool. The purchase order is not the strategy.
Latest Video
This week's eight-minute briefing walks the grid and the loop to run on Monday: score differentiation, score readiness, route the verdict — and for any Build, demand the data moat in writing. Govern the vendor with a concentration limit and a model-agnostic exit. Run it across the whole portfolio and the picture inverts: the capabilities you were about to build, you buy; the model you were about to train, you assemble; and scarce talent concentrates on the two or three places you genuinely own a moat.
Watch: youtu.be/6IHvqq-nVFc
The free five-page playbook in the Frameworks library turns it into a worksheet for your own AI portfolio.
Reply and tell me which capability you were about to build that you should have bought — most portfolios have at least one. I read every response. Forward this to a banking executive about to sign off on an in-house build.
Was this forwarded to you? Subscribe to The AI Architect Letter — free, every Saturday.
Minh Tran · AI Business Architect · LinkedIn · Workshops & advisory: aibusinessarchitect.ai