The AI Architect Letter

The archive.

Every issue of the weekly banking-AI briefing — one framework on real AI value, written for the boardroom, not the hype cycle.

The Vendor Concentration Heatmap: You Are Not Diversified

Issue #018 — The Vendor Concentration Heatmap: your AI estate has three vendors and may have one dependency. Two supervisors named substantially the same short list, and the six layers where concentration actually sits are not the layer a supplier count can see. By Minh Tran.

The AI Unit-Economics Ledger: Your AI Got Cheaper and Your Bill Did Not

Issue #017 — The AI Unit-Economics Ledger: frontier prices collapsed and bank AI bills did not follow, because a bank buys completed work rather than tokens. Four lines, four gates, and why this is a forecasting problem rather than a cost one. By Minh Tran.

The Payments Defense Grid: A Control That Passes Is Not a Bank That Is Protected

Issue #016 — The Payments Defense Grid: two failures that share no signal, six cells with an owner each, and the half of a payment defence most banks have never built. Liability moved behaviour faster than capability. By Minh Tran.

The Grounded-Knowledge Architecture: A Citation Is Not an Answer

Issue #015 — The Grounded-Knowledge Architecture: three properties a citation cannot prove, five failures with five owners, and six capabilities with one accountability wrapper. A faithfulness check cannot catch a stale source. By Minh Tran.

The Agent Attack-Surface Map: Five Surfaces, One Control Layer

Issue #014 — The Agent Attack-Surface Map: five places an AI agent actually gets attacked inside a bank, one control layer that bounds all five, and why the tool's own manual is the surface with no owner. The model is not the vulnerability. The permissions are. By Minh Tran.

The Adverse-Action Explainability Ladder: R3 Is the Floor, R4 Is the Trap

Issue #013 — The Adverse-Action Explainability Ladder: five rungs of what a bank owes the applicant it rejects, and why the fourth one is a trap. Complexity is not an excuse. By Minh Tran.

The Deposit Loyalty Gate: Reachable Is Not the Same as Gone

Issue #012 — The Deposit Loyalty Gate: two axes, an authorization gate, and which balances leave first. Reachable is not the same as gone. By Minh Tran.

Your Model-Risk Framework Has a Hole

Issue #011 — The GenAI Assurance Grid: five control rings, tier on entitlements, and the ring that binds. Assure the system, not the weights. By Minh Tran.

A More-Digital Bank Is Not a Smarter Bank

Issue #010 — The AI-Augmented Bank: five pillars in dependency order, AI proposes / policy disposes, and the pillar that binds. Digital is not smarter. By Minh Tran.

Read the Map Before You Fund the Model

Issue #009 — The 5-Layer Data-Readiness Map: the Readiness Read, the layer that binds, and Dual-Track. Readiness is necessary but never sufficient. By Minh Tran.

Stop Funding AI Pilots — Run the Portfolio Test

Issue #008 — The AI Use-Case Portfolio: two axes, four verdicts, the Portfolio Test, and the capstone of the series. By Minh Tran.

One Human, Thirty Agents

Issue #007 — The Agent Army Operating Model: four pillars, speed-of-control over span-of-control, and the Span Audit. By Minh Tran.

Stop Building Your Own Banking AI

Issue #006 — The Build-Buy-Partner Grid: two axes, four verdicts, and the three forces regulated finance adds. By Minh Tran.

Same AI, Different Return

Issue #005 — The model is commoditizing; the operating model is the multiplier. The 3 levers, the 70/20/10 Law, and the Multiplier Audit. By Minh Tran.

Most banking AI is deployed where the win is easiest to see — not where the money is.

Issue #004 — Most banking AI is deployed where the win is easiest to see, not where the money is. The Visibility Trap, the 3 Value Zones, and where the return actually lands. By Minh Tran.

Governance is not the brake on banking AI. Manual governance is.

Issue #003 — Governance is not the brake on banking AI. Manual governance is. Four guardians turn control into speed — the Gate, the Guardrail, the Ledger, the Owner. By Minh Tran.

The Agent Maturity Ladder

Issue #002 — The Agent Maturity Ladder. A boardroom framework for the question every bank now faces — how much autonomy to give an AI agent, and how to defend it. By Minh Tran.

Banking AI's largest peacetime experiment has gone backwards

Issue #001 — Banking AI's largest peacetime experiment has gone backwards. The Boardroom Equation, Vietnamese benchmarks, and a risk note. By Minh Tran.