AI in Finance — Deployment Digest – Week of 13 October 2026

Nine items from the monitoring run of 13 October 2026, drawn from a broad search covering roughly 6–13 October. The run is the reverse of last week's: regulatory-heavy and deployment-light. Its two substantive events are regulatory, and they are the only two items that reach the log through accountable public bodies rather than interested parties — the Financial Stability Board's delivery of its finalised AI practices to the G20, and the Monetary Authority of Singapore's binding AI risk-management guidelines. Every product capability below is asserted by an interested party, and every boundary is described rather than independently demonstrated in this run's sources. This week the digest is organised around a yardstick: what the FSB recommended, and where each of the week's deployments sits against it.

The yardstick

The Financial Stability Board submitted its finalised sound practices for the responsible adoption of AI to the G20 on 12 October, ahead of the finance ministers' meeting on 15 October. On the FSB's own statement, the practices are "finalised" and incorporate "feedback from the public consultation." The Chair's covering letter, from Andrew Bailey, speaks only of "continued vigilance" on the "risks associated with artificial intelligence." The final text itself had not been published at the time of writing. What can be set out with confidence is what the consultation it finalises recommended in June, as reported at the time.

For agents, that consultation asked firms to define the actions an agent is prohibited from taking or may take only with human approval; to give each agent its own identifier and treat agents as "synthetic employees" under adapted HR controls; to limit agents' access to external environments until they can be shown to operate safely; and, where agents touch money, to require "human approval or dual authorisation above a threshold value," restrict direct access to payment systems and keep audit trails of every agent transaction. Its larger move concerned where the person sits. Conceding that "decision-by-decision review does not scale," it proposed "human-in-command" — "high-level oversight of autonomy boundaries and guardrails" — with AI-in-the-loop monitoring as agent numbers grow, and "ultimate accountability" retained by institutions and individuals. The practices are not binding and are not an international standard.

Confirmed (finalised report delivered to the G20 on 12 October; via the FSB's press release and the Chair's letter). The recommendations above are those of the June consultation, as reported by The Asian Banker and TechInformed on 11 June; any change in the final text is not yet visible and will be logged when it is. The consultation's survey figures — 52% of firms in active agentic use, 23% scaling — are carried over from the draft. The item resolves a watch item this log has carried since the consultation first appeared in June.

The yardstick also closes a loop in this log's own questions. Across recent weeks the question put to every launch moved from where the agent stops, to who answers when it gets the money wrong (Week of 29 September), to whose hand is on the gate (Week of 6 October). Last week a vendor, Mastercard's Johan Gerber, and a regulator, the Bank of England's Sarah Breeden, each said that a human on every action would not last. This week a standard-setter put that judgment into a delivered report and answered the earlier question with it: when the per-decision gate goes, a named institution or person owns the outcome. What follows sets the week's deployments against each recommendation in turn.

Define what the agent may not do, and what needs a person

Zocks ("Advisor Intelligence" plugin for Claude) — Zocks released a plugin of seven Claude skills running on "a read-only MCP connector drawing on past-meeting data." A single request — "prep the annual review" — produces a summary, a client letter, a branded deck and briefs for the client's accountant and attorney. Zocks says Claude should "finish the job," not just answer. One limit is defined: no external actions, no sending, no trading. What is not defined is the step the recommendation would point to first — whether an adviser must review the "client-ready" documents before they reach a client.

Confirmed (released 30 September, surfaced in this window; via InvestmentNews and WealthTech Safari). A vendor describing its own product. Zocks reports use at more than 5,000 firms, including seven of Barron's nine top mega-RIAs; the figure is its own. Anthropic's "Claude for Financial Advisors" launch (14 September) is the background to the plugin, not an event of this window.

OneVest and PlannerPal — OneVest, selected this week by Network Wealth, carries the "agentic account-opening" launch logged last week; the coverage gives no detail on which steps a person approves. PlannerPal raised $6m (£4.3m, led by Mercia Ventures) for what it calls a "client-intelligence layer," used by more than 250 UK advice firms with around £150bn in client assets; the human boundary is not described.

Confirmed (selection and raise; via WealthTech Safari). Vendor self-descriptions; firm and asset counts are the firms' own. In both cases the coverage defines no action as prohibited or as requiring approval.

The week's clearest answer to this recommendation comes from M1, at the close of this digest.

Where agents touch money: approval above a threshold

JPMorgan ("Smart Cash") and Rivo (deposit-sweep agents) — PYMNTS reported a set of products that put agents on idle deposits, under the headline "AI Agents Wake Up $1.6 Trillion in Sleepy Bank Deposits." JPMorgan's chief executive, Jamie Dimon, said Smart Cash would "move money automatically between checking and brokerage products." Rivo "watches cash flow, moves idle balances into Treasury bills, and pulls them back before bills are due." These are the week's money-moving agents, the exact case for which the recommendation asks for human approval or dual authorisation above a threshold. The coverage names no threshold, no approval step and no dual authorisation for either. That is a statement about the telling, not the product: the coverage does not say whether users approve individual moves.

Reported (products live or announced; via PYMNTS sector reporting, 8 October, aggregating bank and vendor products — interested parties throughout). Smart Cash rests on a Dimon quote. The $1.6 trillion is the article's figure for idle deposits across JPMorgan, Bank of America and Wells Fargo.

Meta ("Muse"), Raisin, Robinhood, Rocket Money — The same report places other products lower on the scale. Muse connects to accounts through Plaid, whose own note describes "read access to balances/transactions/holdings, not money movement." Raisin is a deposit marketplace placing funds across more than 250 partner banks, the money staying inside the banking system. Robinhood and Rocket Money are named for trading and for bill negotiation and savings automation.

Reported (same source). Muse's marketed account connection and Plaid's stated read-only scope are not the same thing, and only Plaid's is specific.

SEC (self-custody proposal) — The SEC proposed letting registered investment advisers self-custody client crypto, on conditions that include "two-person transaction authorisation." It is the dual-authorisation shape of the FSB's money-moving recommendation, here written into a custody rule and applied to the adviser rather than the agent. The two are separate instruments and the run does not connect them beyond the shape. The proposal follows the CLARITY Act's failure in Congress last month; the Investment Adviser Association welcomed it, and Better Markets warned of a "very high risk of loss."

Proposed (comment period of 60 days after Federal Register publication; via WealthTech Safari reporting the SEC). No AI dimension; logged for the control it writes into rule.

Limit what the agent can reach

Zocks and Muse, again — The week's two named access limits are both on reading. Zocks's connector is read-only; Muse's Plaid connection, by Plaid's account, reads and does not move. Both fit the recommendation's instruction to limit access to external environments. Neither settles what the agent produces from what it reads: Zocks's connector bounds the input, while the output is a client letter and a deck. No product this week used the FSB's "synthetic employee" framing or described an identifier for its agents.

Confirmed for Zocks, Reported for Muse (as above). The access limits are the parties' own descriptions.

From each decision to command, and a name on the outcome

Monetary Authority of Singapore (AI Risk Management Guidelines) — MAS issued binding AI risk-management guidelines for every firm it supervises. Core governance and oversight (Sections 3–4) take effect on 7 October 2027; lifecycle and third-party controls (Sections 5–6) on 7 October 2028. MAS states that agentic systems "can cause harm faster than any conventional supervisory intervention can stop it" and may act "without human approval at each step," and it expects firms to be able to "limit, suspend, or replace" them. It does not yet set a human-in-the-loop rule for agents: it names the gap and will consult the industry in 2027. What it fixes now is accountability: "engaging a third-party AI provider does not reduce a financial institution's accountability."

Confirmed (issued 7 October; via Techtimes reporting MAS, with MAS primary). Binding on MAS-supervised firms. The agent-specific rule is deferred; the four pillars of the accompanying SAFR white paper (policy-bound execution, real-time validation, auditability, interoperability) are voluntary proposals, not rules.

Set side by side, the two regulators make the same move in the same week. The FSB recommends relocating the person from approving each decision to commanding the boundaries, with ultimate accountability kept on an institution or individual. MAS defers the agent rule and fixes accountability now, including where the AI comes from a third party. That clause bears in shape on the week's products built on third-party models and connectors — Zocks on Claude, Muse on Plaid — though none of them is MAS-supervised. Both regulators accept the premise Gerber and Breeden voiced last week. They part from Gerber's forecast that "this question will no longer be relevant" by holding a line elsewhere: not at each action, but at a name. Neither the FSB consultation nor MAS states yet how that accountability is enforced once per-decision review is conceded.

The aggregate question

The FSB's remit is the stability of the system, not the conduct of a single product. The deposit-sweep agents above are the first deployments this log has recorded on the thread opened by Apollo's chief economist, Torsten Sløk, whose "agentic bank run" warning (Week of 29 September) described deposits leaving "at machine speed" because no person is on the sweep. PYMNTS frames the extension as a conditional — "if every household used AI agents to optimize cash returns" — not as observed outflows. Citi's Andrew Coombs expects trust, reliability and regulation to limit third-party agents, and banks to build the capability into their own apps; Citi expects any shift to be gradual.

Reported (sector reporting and analyst comment). The sweep products are shipped or announced; the run is a hypothetical. The two describe the same capability from opposite ends, and nothing in the coverage connects them beyond that.

Beside the yardstick

NextWealth (UK managed-portfolio survey) — On-platform discretionary model-portfolio assets reached £208bn, up 32% year on year; tailored models hold around 14% of those assets, nearly three-quarters of discretionary managers offer them, and average costs have roughly halved since 2021. Reported (survey; via WealthTech Safari). No AI dimension; logged as the market the adviser tools serve. Altruist added a donor-advised-fund feature; it is not an AI product and is logged for completeness.

The adviser with no human behind it

M1 (M1 Adviser) — M1 launched an AI adviser under M1 Advisory Services, a reactivated SEC-registered investment adviser with $1.85bn in assets. It is fiduciary, non-discretionary and free to M1 clients through 2027. In the publication's own voice: "There are no human advisors behind it." And: "The AI can't act on its own"; "the client decides to take action."

Against the first recommendation, M1 is the week's most explicit answer. Nothing it recommends executes without the client. Against the larger model, it is the case the framework does not quite anticipate. The FSB's human-in-command assumes an institution whose people command its agents. M1 keeps the institution and the fiduciary duty, removes the human adviser, and puts the gate on the client, at the money. It is the mirror image of Schwab's Charley (Week of 6 October), which keeps the machine off the money and has the client confirm each action it can take, and of the "agents propose, people release" architecture the banks described at Sibos: there the machine is kept off the release; here it is put on the advice. The run calls M1 "the opposite of everything the log has tracked," where the operational tier automates and the judgment tier stays human.

Confirmed (announced; live to M1 clients; via WealthManagement.com). The firm is an interested party, but the no-action boundary is stated plainly and in the publication's own voice. Not stated: whether the Plaid connection is read-only; how the fiduciary duty, which attaches to the advice, is carried when no human adviser stands behind it; and whether "non-discretionary" is a permanent design or a launch posture — the same "firm now, open later" shape this log noted for Charley. M1 is a retail product, at the opposite end of the market from the family office.

What the run did not carry

No new production deployment from a bank or CFO office: the window falls between Sibos (28 September–1 October) and Money20/20 USA (18–21 October), and the run logs the quiet as a datum — launches appear to track the conference calendar, with this week's regulatory responses arriving in the gap. Money20/20 is flagged as next week's set-piece, where vendors are already pre-announcing agentic banking showcases; its launches can be read against the finalised practices when they land. No dedicated family-office or ultra-high-net-worth agent for advice, allocation or capital execution: the wealthy client's agentic surface remains onboarding, KYC and adviser intelligence, with M1 the one new data point on the advice tier, and at the retail end. No cost-ceiling item, for a second consecutive run, after Oasis Group (Week of 22 September) and Deloitte (Week of 29 September). No demand-side AI-adoption survey. No retail-brokerage matched pair to follow last week's Schwab and Robinhood.

Recurring, not new: the downward movement of autonomy toward money continued in the deposit-sweep agents, after Scalable (Week of 1 September), Arca (Week of 15 September), Collective (Week of 22 September), Ant International (Week of 29 September) and Robinhood (Week of 6 October). The "synthetic employees" coinage recurred in the FSB's practices; role-language recurred in Zocks's "finish the job" and Anthropic's "symphony conductor." The MCP connective-layer thread continued through Zocks. The regulatory surface the agents operate on kept moving, in the SEC proposal after the IRS notice last week.

Assembled from a structured weekly monitoring run of 13 October 2026.

SMA Crown Confidential

Bespoke AI Intelligence for Private Decision-Makers