AI in Finance — Deployment Digest – Week of 7 October 2026

Eleven items from the monitoring run of 6 October 2026, drawn from a broad search covering roughly 29 September–6 October. Every capability below is asserted by an interested party or self-reported, and every gate is described rather than independently demonstrated in this run's sources. The through-line is a change of question again: not where the gate sits, nor who answers when it fails, but whose hand is on it.

Who sets the gate

The question this log has put to every launch is where the agent stops. Last week's run asked who answers when it gets the money wrong. This run asks something that sits before both: who decides when a person must approve what the agent does. The eleven items give different answers, and they are not variations on one another. The firm: Schwab's new assistant does not trade, on Schwab's own statement. The customer: Robinhood ships approval switched on and lets the user switch it off. The bank, by its own rule: at Sibos, payment agents propose and people release, and HSBC brings people in for exceptions and higher-value transactions. The vendor, by a threshold it does not publish: WealthAi brings reviewers in "when their input or sign-off is needed." The system's own measure: Pravar AI sends output to a person when its confidence is low. And the law: twenty-six state attorneys general defend a consumer's right to human review in AI lending decisions, and the Financial Stability Board's draft recommends human approval above risk thresholds.

These are logged in different sections of the run and do not answer each other. Set side by side, they mark out who currently holds the decision that a person must be asked: the seller, the buyer, a number in the system, or a statute.

The firm: the line drawn in its own voice

Charles Schwab ("Charley") — Schwab announced an AI assistant for its retail clients, with a phased rollout beginning in October 2026. The boundary is stated by Schwab itself: Charley "does not execute trades or autonomously move money." It can "complete select actions" — enrolling in dividend reinvestment, adding beneficiaries, updating contact information — "each requiring client confirmation," and it carries a standing disclaimer that it "may provide inaccurate or incomplete results." Schwab reserves "more meaningful actions" to a person. The gate here is fixed by the firm, and the client confirms inside it. The run notes that the exclusion of trading is stated for launch and is not characterised as permanent.

Confirmed (announced; phased rollout from October 2026; via Schwab's press room, ThinkAdvisor, Crowdfund Insider and WealthTech Safari). A firm describing its own product; the confirmation step is asserted, not independently demonstrated. The boundary is firm in the present tense and open in the future tense.

The customer: the gate as a setting

Robinhood (agentic trading) — Robinhood's agents "analyze markets, build strategies and trade around the clock" and can "execute conditional strategies." The gate is a setting in the user's hands: "manual trade approval on by default," with trades able to execute "without individual confirmation if the user allows it." The stated safeguards sit around that choice: segregated agentic accounts holding only the funds the user allocates, no margin or leverage, a push notification for each trade, a one-tap disconnect. The disclaimer states that agents "can behave in unexpected ways" and that users bear "full responsibility for every trade." The product is not new this week — it has been live since May 2026. What surfaced in this window is its scale: more than 150,000 agentic accounts and roughly 30 million daily tool interactions.

Confirmed (live since May 2026; usage figures surfaced this week via WealthTech Safari). The figures are Robinhood's own. How many of the 150,000 accounts run with approval switched off is not stated: the scale figure and the autonomy figure are different measures, and only the first is given. The extension to crypto (17 August) is background, outside this window.

Schwab and Robinhood are two large US retail brokerages, and this week they stand on opposite sides of the execution line. The pairing is an outlet's, not the firms': Crowdfund Insider reported the Schwab launch as "falling short of Robinhood's agentic trading," which it describes as trading "without requiring approval per transaction." The run reads the pair as "a single line drawn twice, in opposite directions" and adds that "the boundary is now a product-positioning choice, not a technical limit." What the evidence carries is narrower: one firm sets the line for the client, the other hands the client the switch. Schwab's announcement is this week's event; Robinhood's product is five months old, and neither firm addresses the other.

Robinhood's disclaimer also supplies a fifth answer to the question last week's digest set out. There the error was answered by a named executive, a compliance gate, a refund, or no one. Here it is answered by the customer.

The bank: agents that propose, people who release

Sibos 2026 (six named production deployments) — The reporters' own summary of the conference was "agents that propose and people who release." BNY Mellon described a payment-repair "digital employee": "AI does the reasoning, deterministic guardrails do the vetting, and people make the judgement calls." BNY says it handles "over 10% of BNY's payment repair issues globally." Finastra presented "Repair Recommendations," its fixes validated against Swift, Fedwire, SEPA, UPI and Nexus rules, with "final action subject to human review and approval." HSBC's "Smart Checking" keeps "people brought in for exceptions and higher-value transactions." BNP Paribas said securities trade-processing agents handle "80 to 85 per cent of the work," reached within weeks, with processing steps cut from ten to six; where a person releases is not explicitly stated. Deutsche Bank's "Ada" was presented through one example: a lending client onboarded and a loan issued "in a day, a process that would have taken a month." Citi's Stephen Randall described walking supervisors through "where the human sits in, on or out of the loop" for each use case — the gate presented as a variable, explained to the supervisor case by case.

Under these sits an architecture stated by SmartStream's chief executive, Akber Jaffer: deterministic systems for "control, auditability, and exception handling that banks can't compromise on," probabilistic AI for "judgment and pattern recognition." It is the same split the log recorded at Caddi (Week of 28 August), Altruist's Hazel (Week of 8 September) and Evergreen.ai (Week of 29 September).

Confirmed (named production deployments; via conference reporting from beri.net, The Fintech Times and Global Finance, with Finastra's own release as the one primary source). The run calls this the log's first set of named, in-production wholesale-bank deployments, as distinct from vendor launch announcements. The throughput figures — BNY's "over 10%," BNP's "80 to 85 per cent" — are single-source and not independently measured; Deutsche Bank's "in a day" is a single-client anecdote. The banks set these gates themselves, and the run notes that several are vague on who sets the escalation threshold and at what value a transaction stops being eligible for automatic handling.

The vendor: "when needed," with the threshold unstated

WealthAi (onboarding and KYC agents) — WealthAi launched agents that "orchestrate the complete client onboarding and KYC workflow," including triggering "checks covering identity verification, PEP and sanctions screening." The human model is escalation only: "human reviewers are only brought in when their input or sign-off is needed." Who defines when sign-off is needed, and at what level of risk, is not stated. The run names the open point directly: the compliance accountability for an agent-triggered PEP or sanctions clearance that a person did not see. This is also where the week's activity relevant to wealthy clients sits. With Schwab's assistant and the Deutsche Bank Source of Wealth agent logged last week, it lies in onboarding, KYC and adviser productivity — the operational and compliance edge of the relationship.

Confirmed (launch; vendor-asserted). The chief executive, Jason Nabi, states the vision; no independent verification of live performance is provided, and the "60% admin-time reduction" is the vendor's claim. "Orchestrate the complete workflow" is the marketed verb; "when needed" is the gate, and its owner is unnamed.

Jump, OneVest, Edward Jones, Arch (item 5 cluster) — Four adviser-side launches leave the gate undescribed. Jump, whose meeting-management suite is now in front of 22,000 Northwestern Mutual representatives, markets "in-meeting account opening." OneVest markets "Canada's first agentic account opening solution." Edward Jones describes a "Retirement Plan Management Center" as "end-to-end," from "prospecting through plan servicing," rolling out from 2027. Arch extended its platform into diligence, framed as decision support.

Confirmed (launches and partnerships; via WealthTech Safari's Week of 2 October), except Edward Jones, which is in development. Vendor self-descriptions. In each case the coverage does not detail where a person approves.

The system: a confidence score decides when a person looks

Pravar AI (item 5 cluster) — Pravar AI offers a continuous evaluation layer that "routes low-confidence output to human review" while maintaining audit records. It is the cluster's most explicitly demarcated design: the gate is the product. It is also a gate of a different kind from the others this week. The trigger is the system's own measure of its confidence; a person sees what the system marks as uncertain.

Confirmed (launch; via WealthTech Safari). A vendor describing its own product; the routing is asserted, not demonstrated. The run's excerpt does not say who sets the confidence threshold.

The law: the gate as a right, and as a threshold

Twenty-six state attorneys general (letter of 24 September) — A letter led by Colorado's Phil Weiser and New York's Letitia James warns that unchecked AI "could soon threaten our financial system, critical infrastructure, and national security," citing "alarming reports of AI agents breaking containment." The specific protection it seeks is narrower than the warning: to preserve against federal preemption a consumer's right to human review in AI lending decisions, such as Colorado's disclosure-and-human-review requirement. Here the gate is neither a product feature nor a setting. It is something the consumer is entitled to by statute.

Reported (letter; a warning, not an event; dated 24 September and captured in this run; via American Banker). The letter is a primary document. The reporting notes it gives "no specific mechanism," "no estimates of financial-system exposure, and no bank-specific recommendations." It joins Apollo's "agentic bank run" warning (Week of 29 September) as a second systemic-risk framing of the removed human.

The customer therefore appears twice this week, in opposite senses. On Robinhood's platform the customer holds a switch that removes the approval step. In the Colorado rule the attorneys general defend, the customer holds a right to it. The two concern different activities — self-directed trading and a lender's decision — and nothing in the run connects them.

Financial Stability Board and Bank of England (regulatory posture) — The FSB's draft on agentic AI calls it a "distinct challenge for human oversight," since agents can "pursue unintended actions without staff awareness." It recommends "defining operational boundaries, requiring human approval above risk thresholds, and treating agents as 'synthetic employees.'" Its survey reports 52% of firms in active agentic use and 23% scaling. The Bank of England's Sarah Breeden said that current frameworks "were not built to contemplate autonomous agents" and that "relying on a human in the loop for all agent actions is unlikely to be realistic," floating circuit-breakers and "kill switches" in its place.

Reported (background, not this week's events). The FSB draft dates from June and was first logged in the Week of 11 June; Breeden's speech was given at Sintra in July. Both reached this run through September and October roundups (llrx.com, bespacific.com). The FSB's final report was expected in October 2026 and is not confirmed released as of this run — a standing watch item.

The gate in the future tense

Sibos day-three panel (agentic payments) — On the same conference's third day, panelists said "irreversible transactions" require "heavy human-in-the-loop" oversight. Mastercard's Johan Gerber forecast that the requirement will lapse: "two or three years from now, this question will no longer be relevant. It'll just be machine at machine speed." The panel named its own open problems: "who bears responsibility when agents cause overdrafts or missed payments," the need for "common infrastructure and common standards" before regulation is imposed, and steeper barriers in business payments, from corporate authority matrices to treasury integration. McKinsey's Uzayr Jeenah called adoption "the fastest pace we've seen… ever" while placing it "at the first innings of a five-day test match." The panel put agent accuracy in trade reconciliation at 99.5–99.9%, against roughly 80–85% for people.

Reported (panel discussion, not deployments; via FinTech Futures). Every speaker sells into agentic payments. Gerber's line is a forecast, not an event. The accuracy figures are panel-asserted and single-source, and they are unrelated to BNP Paribas's "80 to 85 per cent of the work" above: one is a claimed human accuracy rate, the other a share of workload.

One conference, two tenses. The deployments described at Sibos as running today keep a person on the release. The panel at the same event forecasts the person coming off. Gerber's forecast and Breeden's July remark both hold that a human on every action does not last, and they part on what follows: the vendor expects the question to disappear, the regulator wants thresholds, circuit-breakers and accountability engineered in its place. They were not said in the same week or to each other. Schwab's "does not execute trades" belongs with them in one respect only — it too is stated for now.

The family office: the AI is still in the portfolio

RBC and Campden Wealth (2026 North America Family Office Report) — Across 155 family offices — 123 in the United States, 21 in Canada, 11 elsewhere in the Americas — AI is the top investment pick for the next twelve months, named by 85%. That figure is about the portfolio. The operating figures point two ways: half say AI is "embedded and essential to their workflow," while 73% still find investment reporting too manual, a top concern for three years running. Cybersecurity and data breaches are now the top near-term operational concern, cited by 59%. Direct investments average 45% of the private-markets book, ahead of funds at 36%. The run again recorded no dedicated agentic advice or allocation launch for family offices or ultra-high-net-worth clients.

Reported (survey; published 29 September; via WealthTech Safari, with the figures here checked against the RBC release). Both publishers sell to and for family offices. "Essential to workflow" is self-reported adoption, not a measured outcome, and it sits in the same survey as the third consecutive year of reporting described as too manual. The standing observation the log has carried since the June baseline — advice-rich, operations-poor — is neither confirmed nor overturned by a figure of this kind. Last week's distinction holds for the 85%: the AI priority is an investment theme.

The substrate below the agents

Several items carry no agent claim; the log files each as the layer the agents will run on.

DTCC and iCapital — A strategic collaboration, with an undisclosed investment, integrating DTCC's post-trade infrastructure into iCapital's distribution platform, which serves 3,900 wealth firms and $1.2 trillion in assets. Confirmed (partnership). Infrastructure; no agent claim. Platform figures are party framing.

BNY Pershing ("Wove") — BNY Pershing is winding down Wove, the multi-custodial platform it launched in 2023, and absorbing its components. Confirmed (wind-down). Logged by the run as a reminder that not every wealth-technology bet survives.

Visa and Mastercard (agentic-payments standards) — Visa's Intelligent Commerce and Mastercard's Agent Pay, with its "agentic tokens" and "verifiable intent," continue to compete to set the standard for agents that pay on a user's behalf. Reported (ongoing; interested parties). The rail beneath money-moving agents, continuing the thread from Mastercard's AP4M (Week of 11 June) and Ant International's "Account for Agent" (Week of 29 September).

US Internal Revenue Service (Notice 2026-62 and Revenue Ruling 2026-20) — The IRS narrowed Section 351 ETF conversions, treating the ETF as a conduit, with possible retroactive application; comments are due 28 October. Confirmed (per item; via WealthTech Safari). No AI dimension. Logged as a reminder that the regulatory surface agents operate on is itself moving.

What the run did not carry

No cost-ceiling item: the thread that opened with Oasis Group's per-seat-pricing analysis (Week of 22 September) and gained a second data point in Deloitte's cost figures (Week of 29 September) is absent this week. No dedicated tokenisation or settlement strand: after JPMorgan's tokenised private equity, the ECB programme and Pontes last week, the nearest item is the DTCC and iCapital integration. No concentration of one AI provider across the adviser stack: the build substrate behind the item-5 tools is not named in the coverage, a second quiet week after the six-item count of the Week of 22 September. No dedicated family-office or ultra-high-net-worth agent for advice, allocation or capital execution. No confirmed release of the FSB's final report.

Recurring, not new: the regulated settings again kept the consequential step gated — Schwab, the Sibos banks, WealthAi's escalation model and Pravar continue the thread traced through Deutsche Bank (Week of 29 September), Accenture and Ritholtz/Hadrius (Week of 22 September), FinTurk (Week of 15 September), FE fundinfo and Altruist Hazel (Week of 8 September), and Caddi and Scalable (Weeks of 28 August and 1 September): a seventh consecutive run. The downward movement of autonomy toward money continued in Robinhood's trading agents, the Sibos payment agents and the card networks' standards. Role-holder naming recurred — BNY's "digital employee," Deutsche Bank's "Ada" — continuing the register the log has tracked through Citi Sky's "member of the team" and Dynasty's "Virtual Chief of Staff."

Assembled from a structured weekly monitoring run of 6 October 2026.

SMA Crown Confidential

Digital Confidantes: Bespoke AI Intelligence for Private Decision-Makers