top of page

AI in Finance — Deployment Digest – Week of 3 November 2026

Nov 4
8 min read

Updated: Nov 12


Ten items from the monitoring run of 3 November 2026, drawn from a broad search covering roughly the seven days to 3 November. The week's hardest-edged item is a bank contract rather than a product. Two items were announced in earlier windows and surfaced in this week's searches; each is dated where it appears. Every capability below is asserted by an interested party, and every boundary is described rather than independently demonstrated in this run's sources.


Who signs for the agent

Last week this log left the bank's door in two leaves: reading let in, acting kept out. Danske Bank opened the read leaf for its business customers' agents, while JPMorgan, Schwab and Citi were reported to keep outside agents from acting inside customers' accounts at all (Weeks of 20 and 27 October).

This week a bank wrote the act leaf into its terms. Wells Fargo's updated customer agreement, effective 24 November, contemplates a customer's AI agents acting on the account and answers the question the block had left open: who signs for what the agent does. Its answer is the customer.


The other items this week give different answers to the same question, or none. They are arranged below from the customer's signature, to a signature shared out, to a signature on every action, to agents counted with no signer named. They are logged in different sections of the run and do not refer to one another.


The customer signs

Wells Fargo (updated Online Access Agreement) — Wells Fargo has published an updated Online Access Agreement, effective 24 November 2026, that adds AI-agent language under its provisions for Authorized Representatives and Delegates. "All transactions that an Authorized Representative or Delegate (including the transactions of their AI Agents) performs" are transactions authorised by the customer, and the customer is "solely responsible for those transactions." The customer is also "solely responsible for supervising all of your Authorized Representatives and Delegates (including their AI Agents)." The bank "may rely and act on the instructions of" those agents, and reserves the right to "restrict or block access from certain third-party AI tools." No approval step for an agent-initiated transaction appears in the text that could be read. The bank keeps a gate at its own perimeter, the right to block, and places responsibility for each action with the customer.


Confirmed (agreement published, "Version Effective November 24, 2026"; the clauses quoted above were read in the bank's own text). The document makes no capability claim; it allocates risk. The full section on agents could not be read directly, because the fetch truncated at about a quarter of the agreement. Two further clauses, that use of an AI agent is "AT YOUR OWN RISK" and that the bank "does not monitor, screen, or review agent activities," are reported only by Forkast and remain unverified. Forkast is an AI-written outlet covering the "agentic economy," with a stake in the story; the consumer-sentiment and agentic-order figures it cites alongside are unsourced or vendor-reported and are not used here. The agreement does not say how "solely responsible," qualified by "unless prohibited by applicable law," sits with consumer protections for unauthorised electronic transfers. It moves the access fight logged since the Citi block (Week of 20 October) and the "rogue agents" remarks (Week of 27 October) from earnings calls into a signed contract, and it addresses the line between reading from an account and acting on it, followed since Morgan Stanley's MCP opening in the June baseline (Week of 4 June), by contract rather than by mechanism.


The signature, shared out

Lead Bank and Primitive (liability frameworks) — Two firms have proposed ways to distribute responsibility for an agent's actions instead of assigning it to one party. Lead Bank's Jackie Reses describes a "waterfall" of responsibility, in which "an AI provider would be liable where failure arose from a model defect" and "you can revoke transaction credentials." Reses says the technology "to assign and automatically enforce accountability exists" and is "completely within our control… today." Primitive's Derek White offers an "Agentic Financial Services Protocol" for traceability and common guardrails. Neither framework places a person's approval on each action.


Reported (via American Banker, 13 October; announced in an earlier window and surfaced in this week's searches alongside the Wells Fargo change, so it is logged for the thread and not counted as new). Both are interested parties: Lead Bank sells programmable-banking and banking-as-a-service APIs, and Primitive is a start-up promoting its own standard. No case is cited in which liability was actually assigned under either framework. Set beside the Wells Fargo agreement, the month holds two unreconciled answers to the same question: assign it to the customer, or distribute it by protocol.


The signature on every action

Aion ("Aion Agent") — Aion released an agent for its business-finance clients that handles payments, submissions and portal work, with "its own browser, so it can handle portal work that other software often can't reach." The boundary is stated in the body of the release rather than in a caveat: "Every payment, submission, and signature it prepares waits for a person on the client's team to approve it," and "money moves only after approval inside Aion." Credentials stay with the client; "the Agent never sees them." The company describes the design as built "to do the work and let the team approve it." Two edges are left undefined: the agent can "email an account verification letter" with no approval named, and the release does not say whether approval is given item by item or in batches.


Confirmed (announced, stated as available "today to all Aion clients"; vendor release carried by CPA Practice Advisor, 2 November). Aion is an interested party, and the support offered is one chief-executive quote and one customer testimonial; no metrics, error rates, pilot data or independent validation are given. The release does not use the word "autonomous." It joins the prepare-and-approve designs logged at Schwab's Charley and in the Sibos payment agents (Week of 6 October) and at TD's Credit Assist (Week of 27 October).


A boundary drawn, criteria unpublished

BBVA ("The Frame") — BBVA described a governance framework under which each AI agent will have boundaries "specifying which actions it can perform autonomously and which require human oversight." BBVA's Casas said that "people retain control over context, decision-making and oversight." Agents are to "orchestrate and execute the development lifecycle," and a multimodal customer assistant is named "Blue." BBVA reported pilot figures of close to 90% resolution, more than 50% of conversations filtered, and around 28% developer-productivity gain. No sign-off step for agents going into production is described, and the article reporting it notes the absence of "test methodologies, evaluation thresholds or criteria for approving agents for production."


Reported (framework announced at BBVA's Strategic Talk, 6 October; reported by QA Financial, 3 November; rollouts planned for 2027). BBVA is an interested party, and the pilot figures are its own; QA Financial flags the 28% as "a bank-reported productivity measure." QA Financial runs testing and assurance events and so has a commercial interest in the "put to the test" framing under which it reported the framework, though its coverage is critical in places. The phrase "refer decisions for human oversight when required" is QA Financial's paraphrase, not BBVA's wording.


Counted, with no signer named

UBS (third-quarter disclosure) — In its third-quarter reporting UBS cited "more than 12,700 personal AI agents," "608 live AI use cases in production, an 81% increase year-over-year," and an internal system, STAAT, that it says "generated more than 30 million insights." The disclosure says nothing about approval, sign-off or the level of autonomy for any of the agents or use cases.


Reported (self-reported counts; carried by PYMNTS, 29 October, and AI Weekly, 30 October; UBS's own investor-relations release was not read directly). The two aggregators cross-promote their own adoption research. The figures are counts of agents and use cases, not measures of outcome or accuracy, and are not independently verified. AI Weekly's own editor's note calls the agents "employee-built" and asks "who is auditing" them. The same shape of capability named, human unplaced, this log has flagged at vendor scale (WealthAi, Week of 8 September; Tandems' Hamilton, Week of 27 October), here as an institution's headline number.


The loss no one can yet assign

UOB ("Agentic Finance" report) — UOB published a forecast in which agents operate under "bounded autonomy," only where "mandates can be tightly defined, and accountability is clear." Humans set the mandates; no review of individual actions is described. The report dates "autonomous credit underwriting by 2031," and in the same document states that "financial institutions remain cautious about delegating loss-bearing decisions to autonomous systems" and that "current technology cannot resolve whether losses caused by an autonomous transaction fall on the user."


Reported (forecast; via Fintech Singapore and AF.net, 27–28 October; the UOB report was not read directly). The findings are drawn from eight respondents in digital assets, payments and infrastructure, several with a commercial stake in agentic payments. A trajectory forecast, not a deployment. It is unusual among interested parties in naming the liability problem it leaves unresolved.


The regulator's questions

CFTC (Chairman Selig's remarks) — The chairman of the US Commodity Futures Trading Commission, Selig, said the agency's staff are "leaning heavily into using AI" to police markets, and raised unresolved questions about "computer agents in trading" amid "increasingly autonomous trading systems." The remarks name no specific system, control, accountability mechanism or timeline.


Reported (speech; via CoinDesk, 28–29 October, surfaced through the LLRX roundup of 31 October). Regulator remarks, not a deployment and not a published instrument. Last week's regulator item was a document, the SEC's examination priorities; this week's is a speech.


Sold to the adviser

The adviser-tooling cluster (WealthTech Safari, Week of 30 October) — The aggregator's edition carried Nevis, Raymond James's "Raimond," Zeplyn and GatewayGO among its items. The verbs are soft — "building," "extend," "pulls into one view" — with no strong autonomy language, and no item states where a person reviews or approves the output. The week's motion in the adviser tier is capital-raising and adviser "operating-system" framing.


Confirmed for the launches and funding rounds, reported for the survey items (newsletter roundup; each vendor an interested party for its own claim; the publisher sells M&A and capital-raising services and covers deals it reports).


What the run did not carry

No dedicated family-office or ultra-high-net-worth agent for advice, allocation or execution. The two names that surfaced fall outside the window: XSpot Wealth's "XSpot 360," pre-trailed for the Dubai FinTech Summit 2026 with no date placeable in this window, and SagePoint, which keeps its AI options open with no agentic launch. They are logged so the absence is not read as a gap in the search. FIS's Agentic Insights Suite, which "recommends action" and is described as "driving FIS workflows" with no review step stated, was announced on 13 October and surfaced only in a 31 October roundup; it is noted here, not counted. No new wholesale-bank CFO-office or treasury production deployment of the Sibos kind (BNY Mellon, HSBC, Deutsche Bank, Finastra), for a fourth consecutive run. No conference set-piece. The Financial Stability Board's and MAS's human-in-command thread does not recur as its own item. Searched and set aside: a BIS bulletin on the financing of AI infrastructure (27 October), which concerns lending to AI rather than AI in the finance function, and a Deloitte CFO survey that could not be dated or read in full.


Recurring, not new: the human moved from the action to the boundary, in UOB's mandates and in the Wells Fargo agreement's supervision clause, after the Financial Stability Board (Week of 13 October) and the Money20/20 delegation rails (Week of 20 October). The Wells Fargo agreement makes the customer the signatory in the same month that the agentic-payments rails logged in October, Visa with BILL, Mastercard and World, are built to present a delegated agent at the bank's door as authorised in advance. Self-reported figures recurred at BBVA, after TD's 95% (Week of 27 October). Role-naming of agents continued in "Aion Agent," BBVA's "Blue" and UBS's "personal AI agents," after Citi Sky's "member of the team," 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 3 November 2026, with the sourcing composition noted at the top.


Bespoke AI Intelligence for Private Decision-Makers

Comments


bottom of page