One Big Idea

Your deposits were never loyal to your bank — they were loyal to the effort it took to move them, and that effort is disappearing unevenly. Reachable is not the same as gone: the balances that leave first are the ones with no gate standing in front of them.

Last week's issue asked whether you can govern a model you cannot open. This week asks a sharper question, one that lands on your balance sheet rather than your risk committee: can you keep the deposits when your customer's own AI agent is built to optimize against you? This week's research pass came down to a 77% number and a biometric scan, and the two turned out to be the same story told in two markets — one where the fight is already over, and one where you still have time.

Across the last rate cycle, wholesale non-operational deposits passed through 77% of every rate rise, against roughly 40% for retail deposits — cumulative betas over March 2022 to March 2024, on deposits of 7 days or less, self-reported as a panel mean (US Federal Reserve Senior Financial Officer Survey, March 2024). That survey measures betas, not sweep tools, so why the wholesale beta ran so high is this letter's inference rather than a survey finding: corporate treasuries already run automated sweeps that hunt yield without asking permission. When a balance is gated, leaders should map it before they price it. Agentic deposit flight is not a future retail panic — it is live today for corporate and small-business cash, and held back for retail Vietnam by two regulatory walls that will not stand forever.

The Insight · Two camps, one missing question

Executives who hear "AI agents will drain your deposits" split into two camps, and both miss the same piece. The first camp panics, picturing every retail account emptying overnight — a silent run staged by software. The second camp shrugs, trusting an inertia that has already collapsed for the balances that actually fund the bank. Most banks fail here because they score a balance on only two questions — how hard does it chase yield, and can an agent reach it — and skip the third question that decides everything: can the money actually settle out?

A deposit only leaves when three conditions line up: it is rate-sensitive, an agent can reach it, and the authorization gate lets the instruction clear. When a balance is gated, leaders should watch the wall, not the wire. Corporate and small-business surplus cash clears through maker-checker approval, not a biometric scan, so nothing gates it — a beta of 77% is what that looks like on a balance sheet. Retail deposits in Vietnam sit in the opposite position: increasingly visible as open-banking rails standardise, but still gated, because a transfer to a different account holder above 10 million dong — or at or below that figure once the day's transfers pass 20 million — needs biometric matching combined with a one-time password no autonomous agent can produce. Reachable is not the same as gone.

Rate-sensitivity and agent-reachability tell you which balances want to leave. Only the authorization gate tells you which ones can: two axes and a gate, read as a 5-step ALCO drill, not a checklist.

Framework of the Week · The Deposit Loyalty Gate

The framework is The Deposit Loyalty Gate — two axes and a gate, scored per segment rather than as one bank-wide verdict:

  • A1 Rate-Sensitivity — the vertical axis: a balance's deposit beta, how hard it chases yield. Highest on surplus corporate and SME cash, and rising on idle retail savings the moment a spread opens.
  • A2 Agent-Reachability — the horizontal axis: how easily an agent can see a balance and move it, a function of open-API and aggregator maturity.
  • F, Friction and Authorization — the overlay, and where the signature mechanic lives: the deterministic settlement gate of biometrics, maker-checker limits, counterparty licensing and rate caps. Neither axis alone predicts flight; the gate is the thing that changes the verdict. Reachable is not the same as gone.

Score a segment on the two axes, apply the gate, and read the Vulnerability Quadrant as a 5-step ALCO drill: map the book, ring-fence surplus corporate and SME cash, move liquidity monitoring to intraday, become the agent, and watch the gate.

The full grid, with both axes, the gate and the Vulnerability Quadrant drawn out, lives in the Frameworks library.

Use Case · Which balances leave first

Score the Vulnerability Quadrant as an illustrative example, not a real bank's book, and one segment lights up before any other: surplus corporate and SME operating cash — high rate-sensitivity, high agent-reachability, and no biometric gate in its path. Retail idle savings score rising on both axes too, but the gate — the biometric threshold, and a payment-initiation rail an unlicensed agent provider cannot reach — holds them in place, for now.

The ALCO conclusion is uncomfortable and specific: the deposits leaving first are not the retail savers the headlines warn about. They are the corporate and SME balances Treasury has been pricing as sticky, cheap funding for years. Score the segment, apply the gate, ring-fence the quadrant with nothing standing in front of it.

Risk Note

The risk this week is a regulatory clock, not a runaway model. Vietnam's Circular 64/2024/TT-NHNN standardised open banking from 1 March 2025; 1 March 2027 grandfathers older connections already serving individual customers, and is not the moment the rails switch on. And it constrains as much as it opens: Article 7(3) permits payment-initiation APIs to be opened only to banks and licensed payment intermediaries — not to your customer's agent provider. The authentication wall standing beside it is Circular 50/2024/TT-NHNN as amended by Circular 77/2025/TT-NHNN, effective 1 March 2026, which is where the 10-million-dong biometric threshold sits. Price is the limb that does not reach: Circular 48/2024/TT-NHNN frames a ceiling whose numbers sit in Decision 2411/QĐ-NHNN, and it binds only demand deposits and tenors under 6 months — the 6-month-and-longer band a sweep targets is uncapped. So the moat is real, and it is not biometrics plus a price cap: it is licensing plus authentication, which is the sturdier pair. When a gate loosens, leaders should reprice before the quarter closes — widen who may hold that licence, or let the transaction classes drift, and the retail quadrant joins the corporate one with none of the multi-year lead time a bank needs to reprice its funding model.

Reachable is not the same as gone.

Latest Video

This week's briefing — AI Agents Are Coming for Your Cheapest Funding — walks the two axes and the gate end to end, and the Vulnerability Quadrant read you can run on Monday: score one segment of your book on rate-sensitivity and agent-reachability, ask whether an authorization gate actually stands in its way, and ring-fence the segment where the answer is no. Then watch the gate, because the two walls holding retail in place are dated instruments, not laws of nature.

Watch: youtu.be/-wzkVsYyjls

The full deep-dive — the three conditions, both axes scored, the gate and the Vulnerability Quadrant with its sources and caveats — is in The Deposit Loyalty Gate: Which Deposits Agentic AI Takes First — and the Gate That Holds Them.

The free five-page playbook in the Frameworks library turns it into a Vulnerability Quadrant read for your own book.

Reply and tell me which segment of your book has no gate in front of it — most teams reach for the retail savings number first, then find it was the corporate and SME surplus cash Treasury had priced as sticky all along. I read every response. Forward this to a banking executive still pricing corporate cash as cheap, sticky funding.

Was this forwarded to you? Subscribe to The AI Architect Letter — free, every Saturday.

Minh Tran · AI Business Architect · LinkedIn · Workshops & advisory: aibusinessarchitect.ai