AI in Finance — Deployment Digest – Week of 27 October 2026
Updated: Oct 30

Ten items from the monitoring run of 27 October 2026, drawn from a broad search covering roughly 20–27 October. The run's centre of gravity is corporate and commercial banking, the reverse of last week's Money20/20 weighting toward consumers and payments. Every capability below is asserted by an interested party, and every boundary is described rather than independently demonstrated in this run's sources.
The door in two leaves
Last week this log set two things side by side: on the Money20/20 floor, the rails for an outside agent to act on a customer's behalf; at the banks that hold the accounts, the door kept shut, with Citi reported to bar third-party AI assistants and JPMorgan and Schwab said to do the same (Week of 20 October). The block was blanket. It did not distinguish between an agent that looks and an agent that acts.
This week that distinction arrived from two directions. The access fight moved from earnings calls into mainstream coverage and the White House, and the resolution floated there was to split the door: let agents read, keep them from acting. In the same window, a bank opened the first leaf itself. Danske Bank released a connection that lets business customers' AI agents read their accounts, and the capability it describes stops at reading.
The items below are arranged by which side of that line they sit on. They are logged in different sections of the run and do not refer to one another.
The act leaf, kept shut
The "rogue agents" access fight (Chase, Schwab, Morgan Stanley) — Major US banks are reported to be detecting unauthorised activity from third-party AI agents operating inside customers' online accounts. JPMorgan's chief executive, Jamie Dimon, said on the bank's third-quarter call: "We're not going to allow just any outside agent to come inside the company… you don't want rogue agents — and there could be thousands of them." Schwab's chief executive said it has been blocking unauthorised agents. Morgan Stanley said it had not observed unauthorised agent activity on its platforms. Chase told customers they "should not give their username and passwords to anyone," pointing them instead to customer-authorised data sharing through open banking providers. The posture is a perimeter block, not an approval workflow: the banks refuse to let outside agents act at all, rather than defining who approves an agent's action.
Axios, reporting the activity, described agents that were not hacking accounts or draining funds but were "largely following user instructions." The White House's David Sacks, writing on X, separated "user-authorized agents" from "true unauthorized/rogue agents." He said that sharing credentials may violate bank terms but that these are "in no way cyberattacks," and suggested that banks limit agents to read-only access: able to view, not to act.
Reported (via Axios, 22 October, Madison Mills, carried by CBS MoneyWatch, 23 October; Banking Dive and CIO Dive, 15 October, for the earnings-call remarks; Sacks's comments on X). The primaries are the banks' own statements and the Axios report. This continues the restriction logged last week from Financial Times reporting on Citi, now with the earnings-call remarks at quote level and the fight carried into mainstream and White House commentary. "Rogue" is Dimon's word; the reported behaviour is the Axios description. The read-only split is a suggestion, not a policy any bank has announced.
The read leaf, opened
Danske Bank (MCP server for business customers) — Danske released a Model Context Protocol server that connects its business customers' AI agents to their accounts. It presents itself as "the first Nordic bank to let business customers bank through AI agents." The capability described is reading and querying: balances, transactions, cash-flow views and combined API queries. No payment instruction or execution is described, and so no approval step is described either: on the stated scope, the agent takes no action for a person to approve. It is the read-only access the item above floats as a resolution, shipped by a bank in the same week.
Confirmed (released 21 October; via Ghalbouni Consulting summarising Danske's release). The bank is an interested party, and "first Nordic bank" is its own positioning; no performance or usage data is given. The headline verb, "bank through AI agents," is wider than the stated capability, which is data-pull, not transaction. Danske stands on the same line this log has followed since Morgan Stanley's MCP opening in the June baseline (Week of 4 June), between reading from an account and acting on it, now in corporate banking.
An agent that pays
Visa, BILL and Billtrust (Trusted Agent Protocol pilot) — Visa reported a live pilot in which a BILL Payment Agent, registered under Visa's Trusted Agent Protocol and Agentic Directory, pays supplier invoices on Billtrust portals using tokenised credentials. The buyer approves the payment first; the agent "executes only payments the buyer has already approved and does not make payment decisions." Visa says the agent's authentication can replace CAPTCHA-style checks, and calls the pilot the "first live pilot of trusted AI agent payments between buyer and supplier platforms."
Reported (live pilot with named participants, 20 October; not a network-wide rollout; via Ghalbouni Consulting summarising Visa's release). Visa is an interested party; the "first" and CAPTCHA-replacement claims are its own, and no volume, error or outcome data is published. It is the week's most concrete deployment, and the agentic-payments architecture this log tracked through the Money20/20 protocols last week (Week of 20 October) now running in business accounts payable rather than as a protocol or demonstration. It continues the authorise-once design logged from eToro's agents (Week of 18 August) onward: the person approves up front, the agent carries the payment out.
Inside the institution: agents that prepare
TD Bank ("Credit Assist") — TD unveiled a tool for its commercial credit teams in which agents draft credit memos and people review and refine it. The boundary is stated by TD: "credit professionals remain responsible for lending recommendations and final decisions." TD says more than 300 credit professionals use it. It is also the one item this week carrying a quality figure: in a self-evaluation on 220 curated deals assessed by more than 70 domain experts, about 95% of evaluated outputs "passed subject-matter expert review."
Confirmed (unveiled 22 October at the Evident AI Symposium; via Ghalbouni Consulting summarising TD). The user count is TD's own. The 95% figure is unverified: TD ran the evaluation, on deals TD curated, against its own experts' standards; no independent verification or post-deployment outcome data is published. Most items this log tracks assert with no metric at all, so the existence of an evaluation is itself unusual; its basis is the bank's. The agent prepares and a person decides, as in the Deutsche Bank Source-of-Wealth agent (Weeks of 29 September and 6 October) and the payment agents described at Sibos (Week of 6 October).
Backbase ("Delivery Factory") — Backbase launched a system that assigns AI agents to roles in building and maintaining banking software: a Business Analyst Agent, a Solution Architect Agent, backend, mobile and web agents, a Migration Agent and an Upgrade Agent. Humans are named across most of the pipeline. Engineers "execute approved plans and verify code changes before work moves forward," architects' proposals go "for review," and delivery leads keep "scope, decisions, and schedules." The one point left undefined is the Upgrade Agent, which "applies changes": the coverage does not say whether its changes need approval before they are applied.
Confirmed (launched; via FinTech Global, 27 October). The productivity figures — teams that "worked twice as fast," an "80% reduction in requirements and specification effort" — were supplied by the company, with no independent verification or methodology. Bank-operations infrastructure, not a CFO-office or treasury tool. It sits beside the coding-agent deployments logged at Goldman Sachs and Citi with Devin; Citi's chief executive, Jane Fraser, cited 14,000 engineers using Devin on the bank's third-quarter call.
Sold to the adviser: capability named, human unplaced
Tandems ("Hamilton") — Tandems launched what it calls an "AI-native wealth operating system," with an assistant named Hamilton that it says can "handle up to 80% of a financial advisor's daily tasks." The list includes financial planning, portfolio rebalancing, client reviews, meeting preparation, proactive client communication, account opening, asset transfers and paperwork, with "high reliability and accuracy." The release describes controls rather than a boundary: a firm can switch off a single skill, a model or all AI in real time; every AI-assisted interaction is logged; prohibited use cases are detected and blocked. It does not say which of the listed tasks run without an adviser's involvement or which need review — including rebalancing, account opening and asset transfers, which move money or change accounts.
Confirmed (launched; via a Business Wire release carried by GlobalFinTechSeries, 27 October). Every performance claim is vendor-asserted: the 80% figure and "high reliability and accuracy" carry no metrics, benchmarks, case studies or independent verification. The named customers — Wells Fargo, Scotiabank, Citizens and New York Life — and the scale cited, $80bn in assets and 2.7 million accounts, are Tandems' own. The same shape this log flagged for WealthAi (Week of 8 September). The switch-off control recurs from the regulators' side of the month: the Bank of England's Sarah Breeden floated "kill switches" in place of a human on every action (logged Week of 6 October), and MAS expects firms to be able to "limit, suspend, or replace" agentic systems (Week of 13 October).
The adviser-tooling cluster (WealthTech Safari, Week of 23 October) — Eight vendors surfaced in one aggregator edition: Docupace, Healthspan, DeepVest, Modera, FinTurk, Mili with Egnyte, Finloop and SS&C. The work is adviser productivity and back-office tooling. DeepVest's "PlanLab" is headlined as an "agentic platform," with its autonomy and the human's role not stated. Across the cluster, no item describes an agent acting autonomously or executing actions, and no item specifies where a person reviews or approves the output.
Reported (trade roundup, 24 October; each vendor an interested party for its own claim; no performance data; items carry no individual dates beyond the week of 23 October).
The regulator's own boundary
SEC (examination priorities, fiscal 2027) — The SEC published its priorities for fiscal 2027. For the SEC's own examiners, the report says AI tools are "supplementing, but not replacing, staff judgment." For registrants, the coverage frames oversight around cybersecurity, vendor and embedded-AI risk; no requirement for a human in the loop on a firm's AI is stated. The document moves its own terminology to "super intelligence," tied to a presidential order, in place of its earlier "AI" framing.
Confirmed (published; via WealthManagement.com, 23 October, corroborated by Decrypt, 401kSpecialist, FourWeekMBA and Analytics Insight). A regulator document, not an interested party. The human-judgment boundary is stated for the regulator's own staff; the coverage records none set for the firms it examines.
What the run did not carry
No new dedicated family-office agent for advice, allocation or execution in the 20–27 October window. Two substantial build-outs surfaced and both predate it: Mercer Advisors' "Aspen" (announced 30 July 2026), a unified knowledge graph said to carry "hundreds of thousands of actions a year," with no statement of which actions run without an adviser; and Compound Planning's "CompoundAI" (same 30 July InvestmentNews coverage), where "the advisor reviews and approves." They are logged so the absence is not read as a gap in the search. The family office's AI story remains an investment theme rather than internal autonomy: UBS's Global Family Office Report 2026 (28 May) has 65% invested across the AI value chain, a figure about where capital goes, not what agents do. No dedicated ultra-high-net-worth agent for portfolio construction, allocation or capital execution; the wealth-tier activity was adviser productivity, back-office tooling and one broad assistant. No new wholesale-bank CFO-office or treasury production deployment of the Sibos kind (BNY Mellon, HSBC, Deutsche Bank, Finastra), for a third consecutive run. No conference set-piece: the only named event is the Evident AI Symposium, where TD presented. The Financial Stability Board thread, logged in the Weeks of 13 and 20 October, does not recur.
Recurring, not new: the regulated settings again named the person where the agent prepares — TD's credit decision, Backbase's reviewing engineers — after Sibos, Deutsche Bank and Ballerine earlier in the month. "Capability named, human unplaced" recurred at roundup scale in the WealthTech cluster and in one wide claim at Tandems. "First" positioning recurred at Danske and Visa. Role-naming of agents continued in "Credit Assist," Backbase's role-titled agents and "Hamilton," after BNY's "digital employee," Deutsche Bank's "Ada" and the FSB's "synthetic employees." The wealthy client's agentic surface remains operational, as it has been since the June baseline.
Assembled from a structured weekly monitoring run of 27 October 2026.
Bespoke AI Intelligence for Private Decision-Makers
.jpeg)