AI in Finance: The OCC Tells Examiners How to Judge Bank AI, Wells Fargo Makes Customers Answer for Their Agents, and Goldman Limits AI Coding Tools to Engineers

Title: AI in Finance: The OCC Tells Examiners How to Judge Bank AI, Wells Fargo Makes Customers Answer for Their Agents, and Goldman Limits AI Coding Tools to Engineers
Date: 5 November 2026
Type: Paper
Author: SAASiQ (contact@saasiq.ai)
Word count: 2762 words
Reading time: 11 min
Published: 05-11-2026
The Office of the Comptroller of the Currency, which supervises US national banks, issued a framework on 4 November for how its examiners judge banks' use of AI. Wells Fargo has published new online banking terms that treat anything a customer's AI agent does under their login as authorised by the customer. The terms take effect on 24 November. Goldman Sachs has limited Anthropic's Claude Code and GitHub's AI coding tools to its engineers, Bloomberg reported on 2 November. Fraxion, a procurement software company, released AI risk grading for supplier invoices on 29 October. This paper sets out how each use works, what is known about cost, and where it breaks.
What happened this week
This is the sixth of SAASiQ's Thursday papers on how AI is used in practice, sector by sector. Finance came first on 1 October, followed by HR, the public sector, the supply chain and customer service. This week the rotation returns to finance and starts its second round.
The week from 29 October to 4 November brought four pieces of news. A US bank regulator told its examiners how to assess AI. A large US bank changed its terms so that customers carry the risk of their own AI agents. An investment bank took AI coding tools away from most of its non-engineering staff. And a software supplier put AI on invoice checks in accounts payable.
Two of the four follow on from last week. SAASiQ's paper of 29 October covered the Axios report of 22 October that US banks were finding customers' AI agents inside online accounts. The OCC framework and the Wells Fargo terms are the first formal answers to that problem from a regulator and a bank.
Use case one: how a regulator will examine bank AI
The OCC published its framework as Bulletin 2026-54, titled "Super Intelligence: Supervisory Framework". The OCC uses the term super intelligence, or SI, where most of the industry says AI. The bulletin is a resource for examiners. The OCC says it does not set enforceable standards, prescriptive requirements or new supervisory expectations for banks.
The framework has 16 components, grouped into six pillars: purpose, risk, control, oversight, resilience and assurance. It applies to generative and agentic AI used in normal business, such as credit decisions, payments, fraud detection, customer service, operations and risk management. Older, non-generative models stay under the OCC's model risk principles in Bulletin 2026-13, according to a footnote in the framework.
Two points will matter most to banks. Examiners are told not to treat AI as a separate risk category. And banks are not required to set a separate risk appetite for it. AI risk is to be judged inside the categories examiners already use, such as credit, operational and compliance risk.
Examiners are to assess each use case by its nature and materiality, the consequences if it fails, how much decision-making authority passes to the system, and how much human oversight remains. One summary of the framework describes four tiers, with a fully autonomous, customer-facing payment agent at the top. SAASiQ has not confirmed the tier labels against the bulletin itself.
Agents the bank does not control
The framework also deals with agents that belong to someone else. In its section on operational resilience, the OCC warns that external agents run by customers, fintechs or other businesses can create risk for a bank even though the bank does not control them, Decrypt reported.
It suggests examiners look at five controls for those agents: identity verification, delegated permissions, transaction limits, monitoring, and the ability to suspend or revoke access. Decrypt reported that the OCC did not link the framework to any particular incident.
For a bank, this points to three records an examiner may ask for. The first is a list of the AI systems the bank runs and what each can do without a person. The second is a list of the outside agents that reach the bank's channels, and how the bank detects them. The third is the procedure for switching one off.
Where supervision of AI breaks
The first weak point is the inventory. A framework built on use cases assumes the bank knows its use cases. AI features now arrive inside software updates from suppliers, as SAASiQ's 1 October paper noted. A bank that counts only the systems it built will miss some of the ones it bought.
The second is autonomy creep. A tool approved to draft a payment can later be set up to send one. The framework's test of how much authority passes to the system works only if someone re-checks the use case when its permissions change. In SAASiQ's view, a change in what an AI system may do should go through the same approval as a new system.
The third is the outside agent. A bank can verify the identity of an agent only if the agent identifies itself. Most of the agents found in customer accounts in October were using the customer's own login, according to the Axios report. To the bank, they looked like the customer.
Use case two: customers' agents and who carries the loss
Wells Fargo's answer to that problem is contractual. Its updated Online Access Agreement takes effect on 24 November. The bank's published text says customers are responsible for AI agents "if and when we permit the use of such agents". Transactions an agent carries out under the customer's access are treated as authorised by the customer. The bank says it does not provide or control third-party agents. Tech Times reported the terms on 3 November.
The new section 15 sets out how authority ends. A customer who wants to stop an agent must revoke its authorisation properly and tell the bank. The agreement adds that the customer "must retain evidence of such revocation and must provide such evidence to us, at our request". Until the bank is told, the customer stays responsible for what the agent does, according to coverage of the terms.
The bank also keeps control of the channel. Under section 15(g) it can require agents to register, limit what they do, change how they get access or block them altogether. The bank has no duty to check whether an agent's authority is valid or what it covers, unless a law or regulation requires it.
The liability terms apply subject to applicable law or regulation. In the US, Regulation E protects consumers against unauthorised electronic transfers. How that rule applies when a customer gave an agent access and the agent then did something the customer did not intend has not been tested in the reporting so far. Commentators on the terms differ on how far the contract can shift the loss.
Where customer agent terms break
The first problem is proof. A customer can revoke an agent's access in the agent's own app. Whether that leaves evidence the bank will accept depends on the app. Most consumer AI assistants were not built to produce a revocation record for a bank.
The second is timing. An agent can act in seconds. A notice to the bank takes longer. Losses in the gap fall on the customer under these terms.
The third is the business customer. Small firms often run their banking through the owner's login. If the owner gives an AI assistant that login to pay suppliers, the firm carries the risk of every payment the assistant makes. A finance team in that position should treat the assistant as a person with payment authority, with limits and a second approval.
The fourth is that terms do not stop the agent. A contract decides who pays after a loss. It does nothing to prevent one. The controls the OCC listed, such as registration, limits and revocation, are the ones that reduce the number of losses in the first place.
Use case three: AI coding tools inside a bank
Bloomberg reported on 2 November that internal guidance, updated in October, limits Anthropic's Claude Code and GitHub's AI coding tools to Goldman's engineers. The change followed a review of productivity and risk. About 12,000 engineers keep access.
Traders and bankers who had been using the tools, for tasks such as market monitoring, must now use Goldman's internal AI assistant or ask for alternatives, according to reports based on the Bloomberg story. A Goldman spokesperson said the tools were now with the people best placed to use them over the long term. Reports differ on whether Goldman's strats, staff who combine trading, quantitative research and software engineering, keep access.
Coding tools of this kind can write and run programs on a user's behalf. In a bank, a trader with such a tool can build a small program that pulls market data, checks positions or sends alerts. The same tools also produce software that sits outside the bank's normal development controls: no code review, no testing and often no owner once the trader moves on.
Some reports say code written by non-engineers often needed significant rework. That claim comes from secondary coverage, and SAASiQ has not seen it in Bloomberg's own report. Several outlets also link the decision to the cost of the tools. The reports based directly on Bloomberg give productivity and risk as the reasons, not cost.
Where staff AI tools break
The first risk is the program nobody owns. Banks have dealt with this before in spreadsheets and desktop databases built by business users, known in the trade as end-user computing. Many banks keep an inventory of the ones that matter. An AI coding tool makes these programs faster to build, so the inventory grows faster.
The second is access. A coding tool that can run commands on a trader's machine can reach whatever that machine can reach. If the trader can see client data or send orders, so can the code the tool writes. Engineers usually work in separate development environments. Traders usually do not.
The third is the shift to workarounds. Staff who lose a tool they found useful look for another one. Goldman's answer is an internal assistant that it controls. A firm that withdraws a tool without offering anything in its place should expect personal accounts and unapproved apps to fill the gap.
Use case four: invoice risk grading in accounts payable
The fourth use is in the finance department itself. Fraxion, which sells procure-to-pay software, announced AI-assisted risk assessment for accounts payable on 29 October, in a release carried by CFO Dive and CFO.com.
The system grades each invoice for risk when it is created. Each grade comes with the reasoning behind it and a suggested action. The checks cover unusual amounts, supplier reliability, matching against the purchase order, duplicate submissions, price differences and quantity differences. Three-way matching, of invoice against purchase order against goods received, is done line by line, and differences are shown before payment. Fraxion says the AI flags issues and recommends where to look, and the accountable person still gives final approval. The features are available now. Fraxion does not publish prices.
The demand behind this kind of tool is payment fraud. Yooz, which also sells accounts payable software, published a 2026 Payment Fraud Readiness Report based on a July survey of 750 US finance, accounting and accounts payable staff. In it, 70 per cent said their organisation had faced a payment fraud attempt in the past two years or could not rule one out. And 48 per cent named an AI-powered threat as their top emerging fraud concern, but only 21 per cent said finance teams were extremely prepared for one.
Of those surveyed, 49 per cent said they had let, or almost let, an unusual payment request go ahead without extra checks because it seemed to come from a trusted source. Teams with mostly manual processes lost money in 42 per cent of known fraud attempts, against 30 per cent for highly automated teams. Teams using AI spotted fraud attempts at 63 per cent, against 30 per cent for teams without it. Yooz sells the kind of software the survey favours, and the figures are its own.
Where invoice risk grading breaks
The first weak point is the data. A risk grade built on supplier reliability needs a clean supplier master file. Duplicate supplier records, old bank details and missing purchase orders are common in accounts payable. A grade calculated from bad records will look confident and be wrong.
The second is the change of bank details. A common payment fraud against finance teams asks for a supplier's bank account to be changed. The invoice itself can look normal: right amount, right purchase order, right supplier. A check focused on the invoice can pass it. The control belongs on the change of bank details, with a call to a number already on file, as SAASiQ's 1 October paper set out after the Fideuram case.
The third is alert fatigue. If most invoices come back amber, staff learn to approve amber. The measure to watch is how many flagged invoices turn out to be real problems. If that share is low, the thresholds need tuning before staff stop reading the reasons.
The fourth is the approver. Fraxion keeps a person on the final decision. That person now sees a grade and a suggestion before forming a view. In SAASiQ's view, a finance team should sample approved invoices that the system graded low risk, to check the approver is still checking.
What it costs and who pays
Few prices were published this week. Fraxion asks buyers for a demonstration. Goldman did not say what it spends on AI coding tools, and the reports that link its decision to cost do not give figures. The OCC framework adds no fees, but it adds work: inventories, use case assessments and records of human oversight all take staff time.
The Wells Fargo terms show where some of the cost is moving. When a customer's agent makes a payment the customer did not intend, the bank's terms put the loss with the customer, subject to the law. The assistant's maker will have its own terms. For a business customer, that makes AI agents with bank access an insurance and controls question as well as a software one.
Inside a finance team, the costs that sit outside the licence are the same as in earlier papers in this series. Someone keeps the supplier records clean. Someone tunes the risk thresholds. Someone reviews a sample of what the AI cleared. A business case that counts the hours saved on invoice checks but not those roles will understate the cost.
A framework for finance teams
The steps below apply whichever bank and whichever software suppliers a finance team uses.
List every AI tool that touches money. Include features that arrived in updates, assistants staff use on their own machines and any tool that holds a bank login. For each, record what it can read, what it can do without a person and whose credentials it uses. This is the inventory a bank examiner will now ask a bank for, and a finance team needs the same list.
Read your bank's terms on agents. Find out whether your bank allows AI agents, whether they must register, and who carries the loss if one acts wrongly. If the terms resemble Wells Fargo's, write down how the team would revoke an agent's access and how it would prove it did.
Give agents their own access. Where a bank or supplier offers a separate connection for software, use it instead of a person's login. Set limits on what it can pay and to whom. Keep a second approval for any new payee or changed bank details.
Decide who may build software with AI. If finance staff use AI coding tools to build reports or automations, treat what they build as end-user computing. Record it, give it an owner and review the ones that feed accounts or payments.
Measure the AI checks. For invoice risk grading, track how many flagged invoices are real problems and how many low-risk invoices turn out not to be. Review a sample of both every month.
Keep the call-back rule. For any request to change a supplier's bank details, call a number already held on file. No AI grade should replace that call.
What happens next
Wells Fargo's new Online Access Agreement takes effect on 24 November. The OCC's examiners can use the framework from now on. In the UK, the FCA's second AI Live Testing cohort, which includes agentic payments, runs to the end of 2026, with the FCA's evaluation due in the first quarter of 2027.
SAASiQ - Intelligent Solutions for SaaS ©


