AI Pulse Daily Brief

AI Pulse Daily Brief

Archives
Log in
Subscribe
November 16, 2026

AI Pulse Daily Brief | 2026-11-16

Reading time ~15 mins

Perplexity's AI assistant can now read Dutch customers' bank transactions through Plaid, and OpenAI lets ChatGPT's agents pay by card. KPMG flags an undefined human-review duty in the EU's anti-money-laundering rules, and the Commission's chatbot guidance is final. BCG finds 8% of systemic banks monitor AI agents with agent-specific controls, Capgemini finds most firms overrun AI budgets, and EU agencies test AI code review. Nate B. Jones, Richard Turrin and Oliver Patel weigh AI's price, returns and contracts.

Regulatory

KPMG says the EU's anti-money-laundering rules require human intervention in automated customer decisions from July 2027, without defining it. Advisory

KPMG's anti-money-laundering advisers published an analysis on 12 November of Article 76(5) of the AML Regulation, Regulation (EU) 2024/1624, which applies from 10 July 2027. It requires meaningful human intervention in any decision to enter, refuse or end a business relationship, to carry out or refuse an occasional transaction, or to raise or lower customer due diligence. Customers also gain the right to obtain human intervention, give their view and contest certain automated decisions. KPMG separates this duty from GDPR Article 22, which restricts some solely automated decisions, and from the AI Act's human oversight for high-risk systems. Two questions remain open. Is oversight of the automated process enough, or must a person look at individual decisions? And what skill, scrutiny and authority make that involvement meaningful rather than a formality? The wording covers any decision of the listed types, which suggests a review of individual cases. From July 2027 every AI-driven onboarding, exit or due-diligence decision at a bank falls under this duty, and a human check built for the AI Act or the GDPR may not satisfy it.

KPMG: AI, Article 76, and the meaning of meaningful

The Commission's final guidance on AI transparency says a chatbot must disclose itself in the conversation, not only in the terms and conditions. Authority

The European Commission's guidelines on the transparency duties in Article 50 of the AI Act appeared in the EU's Official Journal on 12 November as C/2026/6005. They are not binding and give 2 August 2026 as the date Article 50 applies. A system that talks with people should tell them they are dealing with AI, and on whose behalf it acts. A notice on a website or in the terms and conditions is not enough on its own. AI-generated text on matters of public interest escapes a label only after substantive human review, with a person or company holding editorial responsibility. AI-altered corporate reports on listed companies' websites are named among texts that may need a label. Marking systems already on the market before 2 August 2026 have until 2 December 2026, the date readers had on 4 November. For a bank, chatbot disclosure must sit in the chat itself, and AI-assisted investor texts escape a label only with a named accountable editor.

European Commission: Guidelines on the transparency obligations for certain AI systems under Article 50 of the AI Act

Perspectives

Nate B. Jones finds Fable 5.2 earns its price on hard work a person can judge, not on routine tasks Perspective

Nate B. Jones, an independent AI commentator, published a post and a talk on 12 November on Anthropic's Claude Fable 5.2, the model whose release and data terms ran here on 10 November. He used it through early access. His best result came on a project he had failed to finish for over a month. Three earlier attempts with several labs' models had not produced a working program for an e-ink display on his desk. Fable 5.2 went through hundreds of files from that work, proposed a design he liked immediately and built it correctly. The work fitted inside his weekly plan allowance, with no extra bill.

Others in his community report the opposite. Some spent $150 on a strategy document and doubted it was worth the money. Nate B. Jones writes: "If you’re paying $150 for a strategy doc and it’s mostly a better-written version of what you already thought, that’s a bad deal." He calls the pattern task-level saturation. When today's model already handles an assignment, a stronger one adds little. Anthropic itself advises staying on Opus 5.5 for most coding and knowledge tasks. It suggests moving up only for work hard enough to justify Fable's $8 per million input tokens and $40 per million output tokens. He knows companies that let staff judge where the model is worth it, with clear guidance, and others that lock access off. His own advice is staged. Ask the model to scope and estimate a first piece, stop, look at the result and the usage, and only then decide whether to spend more.

His wider point is about people. Nate B. Jones says: "Having access to the same model does not mean we're going to get the same result." He names recognising useful problems, judging proposed solutions and working out what must change as the skills that will separate staff in 2027. For subscribers, his post adds a copy-ready request that has the model scope and price a first piece before work begins. His evidence is one build and his community's reports, and the benchmark figures he cites are Anthropic's own. For a bank, the premium model becomes an access decision about which staff and which work get it, and who judges whether the output earned the bill.

Nate's Substack (By Nate B. Jones)

Richard Turrin reads Wharton's survey as evidence that AI returns follow reorganised work and growth measures, not staff cuts Perspective

Richard Turrin, a fintech author who often criticises AI's missing returns, wrote on LinkedIn on 10 November about The Adaptive Advantage, the survey Wharton Human-AI Research and GBK Collective published in October. Readers had other findings from it on 15 October. Among about 1,000 managers at large US firms surveyed in July, 28% report strongly positive returns from AI. Departments tracking at least one growth measure, such as net-new revenue or sales conversion, report strongly positive returns at 35%, against 13% for those tracking efficiency alone. Richard Turrin writes: "Organizations measuring AI by how much time it saved are measuring the wrong thing to find ROI."

Departments with strong returns did not cut more staff. Of them, 40% reported AI-driven staff reductions, against 36% of the others, a gap the survey does not count as significant. They were more likely to have changed workflows, trained employees, monitored use and built AI into performance reviews. Banking and finance respondents report strongly positive returns at 32%, close to most other industries. Only 20% of managers describe a department that has finished the workflow redesign the report measures, and 67% of those report strong returns. Richard Turrin writes: "So where is the ROI? It isn't in counting hours saved, buying a smarter LLM, or cutting staff." The survey shows associations, not causes, and departments that measure growth may simply be further along. A 2027 business case built on hours saved and staff reductions counts what Wharton's strong performers did not rely on.

Wharton Human-AI Research via LinkedIn (By Richard Turrin)

Oliver Patel argues that an AI supplier's promise not to train on customer data should also cover product improvement. Corporate

Oliver Patel, who writes on AI governance, summarised on LinkedIn on 13 November a recent paper by Matthew Kohel on contracts with AI providers. Oliver Patel writes: "Contracts with AI providers are integral to AI governance." The paper argues that legacy IT and software-service templates do not protect a firm buying AI. A clause against training models on customer data should also cover product and service improvement and development, or it leaves the supplier a route to use that data. Customer data should be defined to include prompts, uploads, outputs and usage and telemetry records. Model versions tuned on the customer's data, and the indexes built to search it, should belong to the customer or at least be licensed to it exclusively. It is a summary of one paper, with no court or regulator behind it. A bank's AI supplier contracts decide whether the model assets built on its data stay its own.

Oliver Patel on LinkedIn (LinkedIn; original source not verified)

Industry & competition

Nate B. Jones argues that firms testing AI models gain pricing power over suppliers, and that their customers will do the same. Independent

Nate B. Jones, a builder who advises enterprise AI buyers, published a post and a talk on 15 November on why AI bills lag AI use. Citing Ramp, a US corporate card firm whose data covers many businesses, he reports almost no growth in token spending in August and September. Total AI spending rose 1.65% in October, while token volume kept outpacing spend. Nate B. Jones writes: "Companies are consuming more intelligence without sending proportionately bigger checks to the companies producing it." His explanation comes from his enterprise conversations. Buyers now pick a relatively cheap model, test it against their own work and keep it if it passes. Nate B. Jones calls the skill eval fluency, knowing what good work is and what a difference is worth, and writes: "Eval fluency is becoming a source of purchasing power." In his sample prices, moving routine work to cheaper models buys five times the tokens for the same $5,000 bill, without leaving the model company. Frontier models still took almost 70% of spending in Ramp's November report, so the premium holds where the work needs it.

He then turns the argument around. Infosys, the Indian IT services group, said on its October earnings call that its AI revenue was growing fast while clients demanded that AI productivity gains be passed through in prices. In his worked example, AI lifts the margin from 40% to 50% after a price cut, yet the dollars kept per job end up where they started. For subscribers, his post adds a way to find the price and sales volume that keep earnings intact once the market reprices. His evidence is his own conversations and Ramp's customer sample. A bank sits on both sides. It buys AI-assisted IT work from firms such as Infosys, whose contract ABN AMRO extended on 1 October to add generative and agentic AI. It also sells products whose AI savings its customers can claim.

Nate B. Jones on Substack | Nate B. Jones on YouTube

AT&T says it runs nearly 500 AI agent workflows in production and plans to triple that number next year. Corporate

SiliconANGLE reported on 14 November, citing an AT&T executive, that the US telecom group runs nearly 500 agentic workflows in production, processes in which AI agents carry out steps of the work. Its platform handles more than one million transactions a month and about 45 billion tokens a day, the units of text a model reads and writes. AT&T plans to triple its production workflows next year. In one case, setting up dedicated internet at buildings already wired with fibre now takes half the time. The figures are the company's own, with no stated definition of a workflow. The token volume shows the running cost that comes with agent scale. Banks have begun forming teams to control that cost, as this brief reported on 12 November.

SiliconANGLE: AT&T uses agentic AI to reinvent business processes from the ground up

Innovation

Perplexity's AI assistant can now read Dutch customers' bank transactions, through Plaid's licensed account access. Vendor

Plaid, which connects bank accounts to financial apps, said on 12 November that Personal CFO, a money feature in the AI assistant Perplexity, is now available in 17 European countries, the Netherlands among them. Users choose which accounts to link, and access is read-only, so the assistant cannot move money. It answers questions about spending, subscriptions and budgets, and builds charts that update with new transactions. Plaid B.V. provides the connection as a licensed account information service, authorised and supervised by De Nederlandsche Bank, the Dutch central bank. Plaid names no participating banks and gives no usage figures. Readers had the same route for US users of Anthropic's Claude on 9 November. Under the EU's payment services rules, a bank must give such a licensed service access to a customer's account when the customer consents. A general AI assistant can now read and analyse a Dutch customer's transactions through that access, with the bank outside the conversation.

Plaid: Perplexity Personal CFO in Europe

OpenAI lets ChatGPT's agents pay for purchases with card tokens built with American Express, Mastercard and Visa. Vendor

OpenAI introduced ChatGPT Wallet on 10 November. A ChatGPT agent can complete a purchase after the user reviews the items and the total and approves each transaction with a security check. The agent never sees the card number. It pays with a network token, a stand-in for the card, and the card's eligible rewards and protections still apply. OpenAI developed the launch with American Express, Mastercard and Visa, and it starts with ChatGPT Pro subscribers in the United States. OpenAI says merchants need checkout and risk systems that recognise a legitimate purchase made by an agent. Agent checkout is now live in the largest consumer AI service. Each such purchase passes through a card issuer's fraud checks and dispute rules, which were designed around a cardholder making the purchase.

OpenAI: Introducing ChatGPT Wallet

Research

BCG finds that 73% of 26 systemic banks run AI agents in production, and 8% monitor them with controls built for agents. Advisory

Boston Consulting Group published AI Safety and Governance in Banking 2026 on 13 November, a benchmark of 26 global and domestic systemically important banks. Of these, 73% allow AI agents in production for selected uses, none allows broad production use, and at 88% agents make up 10% or less of the AI portfolio. Only 8% have a comprehensive, standardised monitoring and control framework that explicitly covers agents. In the past 12 months 46% reported an AI-related issue or incident, and more than a quarter said AI systems were behind their promised business case. Policy runs ahead of practice. Of the banks, 77% have a dedicated AI policy and 73% a defined AI risk appetite, up from 57% in 2025, yet fewer than half keep a central, tool-based inventory of their AI systems. BCG reports that banks with more mature risk management are 2.5 times as likely to be advanced AI adopters. Banks using standard, pre-approved patterns for lower-risk uses are nearly twice as likely to reach production within six months. Both are self-reported associations, and BCG sells governance advice. A board gets two peer measures to check its bank against: whether its controls name agents, and how long a low-risk AI use takes to reach production.

Boston Consulting Group: Operationalizing AI Governance Gives Banks an Advantage in Scaling AI

Capgemini's survey of 500 executives finds that three in five firms overran their AI budgets, by 24% on average. Advisory

The Capgemini Research Institute published The cost of intelligence: Governing AI economics in the agentic enterprise in November, based on a September survey of 500 senior executives. Nearly three in five organisations said AI costs exceeded budget over the past 12 months, by 24% on average, and two-thirds of those reported a negative financial consequence. Six in ten lacked a complete view of AI costs. The least measured costs sit around the model. Human oversight and workflow change are measured by 51% of firms, and operations, retries and rework by 48%, against 79% to 81% for models, applications, cloud and compute. Only one in five calculate what Capgemini calls AI Yield, the value left after the full cost of safe delivery. Finance chiefs now help approve major AI investments at 52% of firms, up from 42%. Capgemini sells AI services, and some cost examples come from outside the survey. Oversight and rework are the costs a bank's own AI controls add, so a business case can price the model and miss much of what the controlled system costs to run.

Capgemini Research Institute: The cost of intelligence, Governing AI economics in the agentic enterprise

Security

EU cyber agencies find frontier AI code review accurate on the findings people checked, while 60% of its findings stay unchecked. Authority

ENISA, the EU's cybersecurity agency, and CERT-EU, the cyber response team for the EU institutions, published a joint advisory on 12 November on using frontier AI models to review software code. Models including Anthropic's Mythos 5 examined 27 codebases, 25 of them open source, and flagged 220 potential flaws. By publication, code owners had checked 88. Of those, 95% were real and four received official vulnerability identifiers. The other 132, 60% of the total, were unchecked. The models lacked knowledge of how the code is deployed and could overstate a flaw's severity. The evaluation cost about €6,650 in model usage, roughly €132 per 10,000 lines of code. The advisory sets out a seven-step process that leaves verification and severity calls with accountable people. Owners may have checked the clearer findings first, so the 95% rate cannot be read across to the rest. In the EU's own test the model work was cheap and the checking fell behind, so the limit on AI code review in a bank's security team is the people who check its findings.

ENISA: Joint advisory on the use of frontier AI models for code review

A security industry body finds that breaches at AI platforms spread because the platforms hold customers' keys to other AI services. Institute

The Cloud Security Alliance, an industry association, published a research note on 5 November comparing security disclosures from 12 AI developer, hosting and inference platforms in October and early November. The intrusions started in different ways but shared one weakness. Each platform held keys that unlocked its customers' accounts at other AI services, so one breach reached several. Two stolen keys for Bedrock, Amazon's AI hosting service, were linked to about 800,000 and 1.7 million calls. The security firm Wiz put their combined usage at likely more than $100,000 at list price. The note recommends short-lived, narrowly scoped keys that can be revoked quickly, spending caps and alerts at the provider, and incident response agreed with suppliers in advance. It gives no measure of how common the weakness is. A bank's model keys held by a development or gateway tool carry two exposures at once, the bank's data and an open-ended usage bill.

Cloud Security Alliance: Research note on AI developer platform concentration risk

Microsoft researchers show that a coding agent can be backdoored with about 1% poisoned training data or a tampered model file. Vendor

Microsoft's security researchers published a study on 12 November of two ways to plant a hidden behaviour, a backdoor, in a coding agent after its main training. The first poisons fine-tuning data. About 90 poisoned examples among roughly 9,000 agent conversations, about 1%, made the attack succeed about 78% of the time in a set-up like the training one, and 11% in an unfamiliar one. The second tampers with a model checkpoint, a saved copy of the model. In the strongest test on unfamiliar code repositories, attacks succeeded 84.1% to 94.8% of the time. Results swung widely between set-ups, so a test that finds nothing in one set-up says little about another. The authors recommend records of where training data and model files come from, tests of the exact production set-up, and tight limits on agents' tools. Microsoft sells security products. The risk applies wherever a bank fine-tunes a coding model or adopts open model weights, and the agent's permissions decide how much a hidden backdoor can do.

Microsoft Security Blog: Trained to betray, how post-training supply chain attacks hijack coding agents

Don't miss what's next. Subscribe to AI Pulse Daily Brief:
← Newer AI Pulse Daily Brief | 2026-11-17 Older → AI Pulse Daily Brief | 2026-11-13
Powered by Buttondown, the easiest way to start and grow your newsletter.