This Week in Fintech: Banks restrict outside AI agents as platforms build controlled ones

U.S. banks are tightening the rules for AI agents that try to access customer accounts, while BILL showed what controlled AI access can look like inside a finance platform: logged actions, on/off switches, and a governed channel for outside assistants. A Google Cloud outage exposed how much customer access depends on shared infrastructure. For fintech teams, the focus is shifting to permissions, oversight, and recovery. An $18.6 billion advisor breakaway from J.P. Morgan adds a separate challenge for wealth platforms: keeping client data intact as relationships move between firms.

Citi and Bank of America join JPMorgan in blocking third-party AI agents from accounts

Citigroup and Bank of America have restricted third-party AI agents from logging into customer accounts or moving money, joining JPMorgan Chase and Charles Schwab, the Financial Times reported. Citi said it does not allow third-party AI agents or assistants to log in or transact directly within its digital banking platforms, while still permitting limited read-only access through certain open-banking services. Bank of America is limiting access while it works out who is liable for payments authorized by AI, with CEO Brian Moynihan saying a human should be in the loop to approve money leaving an account. Banks are also concerned that agents could shift idle cash into higher-yielding products, and regulators have privately backed the restrictions.

Source: Financial Times; Retail Banker International

Why it matters

Open banking has expanded authorized access to financial data, but agent-initiated transactions raise unresolved questions about permissions, approval, and liability. Banks are answering them on their own terms for now: read-only data through authorized channels stays available, while credential-based logins and agent-initiated money movement are restricted. Agent access to money is becoming a permissioned layer controlled by the account holder's bank.

The deposit angle explains why the line matters to banks. An agent that sweeps idle checking balances to the best available yield erodes the low-cost deposits banks lend against, so access policy is now a funding decision as well as a security one. For fintechs, delegated authority becomes infrastructure work: agent identity, scoped permissions, limits, approval for outgoing money, revocation, and a liability model banks will accept.

What teams should do

This applies to agentic finance and PFM apps, neobanks, payment apps, open-banking aggregators, and banks designing their own agent access policies.

• Replace credential-based agent access with bank-authorized integrations where available, and verify which accounts and actions each integration permits.
• Design agent permissions as scoped grants covering which accounts, which actions, what limits, and for how long, visible to and revocable by both the customer and the bank.
• Expect bank partners to require human approval for money leaving an account, and build a clear approval step that records who approved each movement.
• Agree with each banking partner how your agent will authenticate, which permissions it can use, and what approval is required for each action.
• Define liability for agent-initiated payments in your terms and partner agreements before launch, because unclear liability is the stated reason access is restricted.
• Plan your roadmap for a period where read access stays open and write access stays restricted: decide which features still work and which need a direct bank partnership.

BILL adds AI controls that let finance teams log, limit, and switch off AI actions

BILL, which serves nearly 500,000 businesses, launched a suite of AI capabilities on October 19 built around customer control. A new hub lets finance teams monitor the actions BILL's AI has taken, switch AI tools on or off, and set how much the AI does. A risk intelligence feature checks invoices against a vendor's billing history before payments are scheduled; BILL says it flagged about $1.5 billion in mistaken or unusual payments in fiscal 2026. BILL also opened an external Model Context Protocol (MCP) connection so customers can query their BILL data from outside AI assistants. The company cites its own research: 86% of SMB leaders use AI in financial operations, while only 19% trust it completely.

Source: BILL

Why it matters

The same week banks restricted outside agents, BILL showed the other side of the design: AI access that the platform governs, records, and lets the customer turn off. Outside assistants connect through a defined MCP channel to ask questions about the data, while AI actions inside the platform are logged and can be switched off per capability. Trust becomes a product feature built from controls, and BILL's own survey shows why it is needed: most SMB finance teams use AI, few trust it fully.

The risk intelligence piece shows where AI earns its place in payments: before the money moves. Checking an invoice against a vendor's history at the approval stage catches errors while they are still cheap to fix, and BILL ties the model to more than $2 trillion in payment volume, data no single customer could assemble. For teams building finance automation, the architectural lesson is to put AI checks in the approval path, log every AI action, and give customers a clear off switch.

What teams should do

This applies to AP/AR and spend management platforms, B2B payment products, accounting software, SMB banking, and any product adding AI to money workflows.

• Build a central log of every action your AI takes on a customer's behalf, visible to the customer, so finance teams can review what happened without opening a support ticket.
• Give customers per-capability controls to switch AI features on or off and set how much they automate, since control is the condition SMB finance teams attach to trust.
• Run AI anomaly checks at the approval stage, before a payment is scheduled, so errors are caught while they are still cheap to fix.
• Explain why each invoice was flagged, comparing it with the vendor's own billing history, so approvers can act on the alert instead of dismissing it.
• Scope any MCP or external AI connector explicitly, defining which data outside assistants can read and which actions, if any, they can trigger.
• Track flagged-payment accuracy and false positives over time, so the risk feature earns trust with measured results.

Revolut's money stayed safe, but its cards stopped working

For several hours on October 20, Revolut's cards, transfers, and app stopped working, while customer funds stayed secure. Just after 7 a.m. UK time, a power problem shut down part of a Google Cloud data center near London. Revolut relies on Google Cloud for core infrastructure, so cards were declined at checkout, ATM withdrawals failed, and many customers were locked out of the app. The disruption reached customers in the UK, Ireland, Spain, Japan, and other countries. Revolut blamed the Google Cloud outage, said the issue was fixed by the evening, and confirmed that customer money and data were safe. Several other UK banks were also affected.

Source: Euronews; The Irish Times

Why it matters

Customers usually treat two things as one: whether their money is safe and whether they can use it. On October 20 the money was safe all day, yet for several hours the app and cards stopped working, and that is what customers remember. Cloud providers split each region into separate zones so a failure in one stays contained. That protection works only if every service a payment needs can keep running when one zone goes offline.

Card payments show why this is hard. Approving a single card payment depends on several services at once: fraud checks, balance checks, card tokens, and login sessions. If any one of them runs in only one zone, the whole payment can fail. Public details leave Revolut's own setup unclear, but every fintech team can check which of its critical services would keep working if one cloud zone went down. In the UK, regulators already expect firms to keep important services such as payments running within set limits during disruptions.

What teams should do

This applies to neobanks, card issuers, payment apps, and any fintech whose logins, transfers, or authorizations run on a single cloud provider.

• Map every service in your card authorization and login paths with the zone and region each runs in, so single-zone dependencies are visible before an outage.
• Run zone-failure drills in a production-like environment, shutting down one zone on purpose and measuring whether authorizations keep flowing.
• Work with your processor to define stand-in authorization rules for when your core is unreachable, within the limits your scheme rules, available balance data, and risk appetite allow.
• Host your customer status channel outside your primary cloud, so you can still explain what is happening when your main stack is down.
• Set and test recovery time objectives for each important business service, and compare them against regulatory impact tolerances where those apply.
• Decide which critical components justify multi-region or multi-provider redundancy, and budget for it explicitly rather than assuming the provider covers it.

J.P. Morgan advisors launch $18.6B independent wealth firm

Phil Scott and Jim Atwood, who managed about $18.6 billion in client assets at J.P. Morgan Advisors, launched Latigo Peak Partners, an independent RIA serving ultra-high-net-worth families. The firm starts with 48 employees and a presence in Florida, Washington, Massachusetts, and California, and will custody assets with Charles Schwab and BNY Pershing, with Fusion Financial Partners advising on the launch. The founders said they turned down private equity partnerships because of the exit timelines that come with that capital. The launch places Latigo Peak among the largest independent wealth firms in the country.

Source: WealthManagement.com; InvestmentNews

Why it matters

A launch at this scale shows how far independent wealth infrastructure has come. A team overseeing about $18.6 billion can stand up its own firm with two custodians, a launch partner, and a family-office-style service model, which means the operating platform a large team needs is available outside a wirehouse. For bank-owned wealth businesses, that puts more weight on what the platform itself contributes to the client relationship.

For the new firm, a critical operational task is moving client accounts with every record intact. Each transfer requires client authorization, and cost basis, beneficiary designations, and household and entity structures all have to arrive clean, because a record that breaks in transit becomes a client-facing error in the firm's first weeks. For wealthtech providers, firms in transition need migration tooling, multi-custodian reconciliation, and day-one reporting.

What teams should do

This applies to wealth platforms, RIA technology providers, custodians, and banks with advisor-led wealth businesses.

• Build authorized client data transfer into your migration flow, with a consent record for each account moved, so transitions follow client instructions and leave an audit trail.
• Build migration tooling for breakaway transitions, covering bulk transfer tracking, cost-basis carryover, and household mapping, so a new firm goes live with complete records.
• Support multi-custodian setups from day one, with one consolidated view of positions and cash across custodians.
• Reconcile positions and cost basis after each transfer batch against the receiving custodian's records, and flag breaks before clients see their first statement.
• Model ultra-high-net-worth households as a first-class data structure, including trusts, entities, and philanthropic vehicles, since family-office-style firms run on those relationships.
• Track transition time and error rates for every migration you support, so you can show the next breakaway team proof of speed and accuracy.

Closing insight

Financial products depend on permissions, infrastructure, and external providers that can change or fail independently of the product itself. Start with one critical customer workflow, such as a login, a card authorization, an AI-assisted payment approval, or an account transfer. Identify its external dependencies, who or what is allowed to act on it, and what happens when one of those services becomes unavailable.

That review usually surfaces the same needs: scoped permissions for people and AI agents, a log of every automated action with a way to switch it off, tested recovery procedures, and reliable data migration. At Itexus, we help fintech teams design these controls and build resilient financial workflows. Contact our team to discuss your product's critical dependencies and the changes to prioritize.