AI Pulse Daily Brief

AI Pulse Daily Brief

Archives
Log in
Subscribe
November 13, 2026

AI Pulse Daily Brief | 2026-11-13

Reading time ~10 mins

The EU's privacy regulators name AI fraud detection as a possible legitimate interest and ask fraud checks to weigh customers wrongly flagged. Two benchmarks find that AI finance agents deliver convincing work that fails when it is recomputed, and Stanford's AI100 panel says AI is not yet paying its way. Mistral's own documents qualify its self-hosted agent story, Bank of America and Citi back a price index for AI chips, and McKinsey finds fixed rules beat agents on two onboarding steps.

Regulatory

The EU's data protection board names AI fraud detection as a possible legitimate interest, and asks fraud checks to weigh people wrongly flagged. Authority

The European Data Protection Board (EDPB), where the EU's national privacy regulators sit together, adopted the final Version 2.0 of its Guidelines 1/2024 on 10 November. They explain Article 6(1)(f) of the GDPR, the legal ground that lets a firm use personal data without consent when it has a legitimate interest. The draft went to public consultation on 8 October 2024, and the EDPB's own comparison file shows what the final text adds. A new paragraph 90 says that developing an AI system to detect fraudulent content or behaviour, a chat agent that assists users, or better threat detection may constitute legitimate interests. It points to the board's Opinion 28/2024 on AI models for how to assess them. The fraud section, paragraphs 103 and 104, now asks controllers to be specific about which fraud they are trying to prevent and what data they really need. It adds that fraud screening can affect people not involved in any fraud, which the balancing test must weigh, and that more serious fraud carries more weight in that test. The three cumulative conditions stay: a legitimate interest, strict necessity, and a documented balance between the firm's interest and people's rights and expectations. One point is not new. Paragraph 80 repeats from the draft that legitimate interest alone does not authorise a solely automated decision under Article 22(2)(b). On 21 October this brief reported a Dutch parliamentary motion, adopted on 8 October, against an EU Council plan to make AI training on personal data a legitimate interest. These guidelines interpret the GDPR as it stands, not that proposal. National regulators, the Dutch one included, now have a final text to hold a bank's AI fraud models against. It asks of each model which fraud it targets and how it weighs the customers it flags in error.

European Data Protection Board: Guidelines 1/2024 on processing of personal data based on Article 6(1)(f) GDPR, Version 2.0

Industry & competition

Bank of America and Citi invest in Silicon Data, which builds price indices for the chips that run AI. Vendor

Silicon Data announced on 12 November that Bank of America and Citi joined its Series A funding round, without disclosing amounts. The New York firm publishes daily rental price indices for AI chips such as Nvidia's H100 and B200, along with forward price curves and estimates of a chip's resale value. The round's first close of $30.5 million in August was led by the Valor Atreides AI Fund, with CME Ventures among the backers. Mickey Bhatia, Citi's global head of spread products, said the infrastructure for compute's "pricing, financing and risk management will need to evolve". Jim DeMare, co-president of Bank of America, said investors will need tools "to price and manage risk". The release does not say that either bank uses Silicon Data's products. On 5 November this brief reported both banks arranging first-lien debt, the slice repaid first, in a $63 billion vehicle that buys AI chips and leases them to Anthropic. In that deal Broadcom, the chip supplier, supports the hardware's resale value. Lending against AI chips needs a value for them that neither the borrower nor the supplier sets, and two of the banks arranging such debt now back a firm that builds one.

Silicon Data: Bank of America and Citi Invest in Silicon Data

McKinsey's bank cases found fixed rules beat AI agents on two onboarding steps. Advisory

McKinsey's article Agentic AI in financial-crime prevention: Lessons from winners, published on 26 October, reached readers on 4 November through its cost, sign-off and approval figures. The full text, drawn from engagements at European banks in 2025 and 2026, holds further findings. Banks often assign 10 to 15 percent of staff to KYC and anti-money-laundering work, the checks on who a customer is and where their money comes from. A manual source-of-wealth assessment can take up to 35 hours, and a pipeline of several agents did it in around half the time, with analysts reviewing set checkpoints. In one such run, about 60 subtasks finished in 15 to 30 minutes. People checked three points: the employment history, important claims that were not fully corroborated, and the final facts and figures. In one onboarding build, two head-to-head comparisons went against the agent. Rule-based routing of follow-up questions beat questions chosen by an agent, and a structured data service beat an agent searching the web, on both speed and accuracy. The article sorts agent uses into risk tiers by four drivers. They are how much the agent does without a human, the harm if it is wrong or manipulated, how sensitive its data is, and what it can act on. It says the AI Act's high-risk label depends on what each system does. In its example, onboarding that shapes a customer's access to banking differs from an internal tool that sorts alerts for an investigator. McKinsey sells this work, and the banks are unnamed. In McKinsey's own cases the agent was not the best tool for every step, and one financial-crime programme can hold systems that fall on different sides of the high-risk line.

McKinsey: Agentic AI in financial-crime prevention: Lessons from winners

Innovation

Mistral's own documents show its new Studio agents keep their state and traces with Mistral, even when self-hosted. Independent

On 6 November this brief carried Mistral's launch of managed agents for Studio, its enterprise AI platform, and BNP Paribas's statement that it runs Studio on its own infrastructure. An analysis by Rajesh Beri in The DAILY BRIEF on 7 November, checked against Mistral's documentation, qualifies that picture. The agents, workflows and apps launched on 5 November are a public preview hosted by Mistral, with no published price. Mistral's page on self-hosted deployments says that when a customer hosts the work, routing, recovery after crashes, scaling, and the logs and traces stay with Mistral. The page states: "Self-hosting a deployment is not the same as running the whole Studio platform on your own premises." Apps run only on Mistral Cloud. A fully on-premises Studio exists as a separate Enterprise contract, priced on request, and no Mistral document says when the new features reach it. Mistral does not say which arrangement BNP Paribas uses. Its documentation warns that behaviour and limits can change during the preview. Beri sets the missing price against rivals that publish meters. Anthropic charges $0.08 per running session-hour for its managed agents, plus model use, and Amazon lists hourly rates for its AgentCore service. Beri notes two safeguards. Workflow data is encrypted before it leaves the customer's servers, and large inputs can stay in the customer's own storage. The 6 November reading of Studio as a self-hosted European route for agents holds for the older on-premises product, not yet for the new agent features. A sovereignty case for those features rests on contract terms that Mistral has not published.

The DAILY BRIEF: Mistral Studio Self-Hosted Agents Send State and Traces to Mistral

Research

Two new benchmarks find that AI finance agents deliver convincing work that fails when someone recomputes it. Institute

FinCosmos, posted on 5 November by Yupeng Cao and 16 co-authors from The Fin AI, Stevens Institute of Technology, the University of Manchester and others, tests agents on investment-strategy work. A simulated user holds a hidden strategy and reveals details only when asked. The agent must find its own data on the web and test the strategy without using future data. Four coding agents and a model without tools ran 150 sessions, and none passed all 15 checks. Each of the 35 sessions that the simulated user signed off still failed a programmed check. None of the 120 sessions with tools delivered trades that reproduced the returns the agent reported. Agents asked about only 54 to 64 percent of the details they had not been told, and in a sample of 60 failures, experts blamed the agent for 53. With the benchmark's feedback, an agent with persistent memory eventually passed all ten tasks. It needed a median of 34.5 sessions per task, and the authors caution that the feedback can leak the hidden strategy. DeepFinance, posted on 7 November by four authors at Julius Research, gave 22 model set-ups 21 private-equity and investment-banking tasks. Each takes a professional 27.1 hours on average and comes with a data room of about 19 files, graded on about 84 expert criteria. The best set-up, OpenAI's GPT-6 Astra at maximum reasoning, met 57.6% of them, and the top four were statistically indistinguishable. The authors report that models often calculate correctly from the wrong inputs and find decision-relevant evidence without acting on it.

Different teams, tasks and methods point to one failure. The output reads as finished, while the inputs or the trades behind it do not hold. On 11 November this brief carried a Columbia and NYU study in which about four AI-written investment memos in ten held at least one wrong claim. Both new papers are unreviewed preprints. FinCosmos uses simulated users and US shares only, and Julius sells AI data analysis. In FinCosmos the requester's own approval caught none of the wrong deliverables, so a sign-off from the person who asked for an AI analysis says little about whether its numbers are right.

arXiv: FinCosmos, An Open-World Evaluation Paradigm for Agents in Finance | arXiv: DeepFinance, Can AI Diligence a $100MM Deal?

Stanford's AI100 panel says AI is not yet paying its way, while AI-related bond debt has passed $500 billion. Institute

Stanford University's One Hundred Year Study on AI published its third panel report, AI Goes Mainstream: The Scaling Era and the Rise of Generative AI, on 10 November. It is the first since 2021, and its 12-member panel includes Michael Wooldridge, Peter Norvig, Sara Hooker and Toby Walsh. It concludes that AI's economic impact is not yet evident in aggregate economic data or in companies' returns on investment. It cites estimates that AI infrastructure investment made up about half of US GDP growth in 2025, and that AI-related corporate bond debt passed $500 billion by mid-2026. It warns that a bursting of the AI bubble would have major ripple effects. The panel expects the next phase to rely less on ever-larger training runs, because good data is scarce and training costs are very high. It expects more from reasoning at the time of use, memory, smaller models and agents. It also notes falling employment of entry-level workers in industries more exposed to AI, such as financial and information services. The report is a synthesis of existing evidence, not a new survey. Readers had the gap between AI spending and measured returns from Evident on 7 October. What is new is the source and the pairing. A panel that reports once every five years puts the missing return next to the debt that finances the build-out. A bank sits on both sides, as a spender on AI and as a lender to the firms building it.

Stanford University: AI100 2026 Study Panel Report, AI Goes Mainstream

On the radar

  • Notion said on 11 November that a breach at Fireworks AI, an AI service it used, exposed mostly unlabelled Notion AI chat transcripts from under 1% of workspaces, and it revoked access keys pasted into them. Notion
Don't miss what's next. Subscribe to AI Pulse Daily Brief:
← Newer AI Pulse Daily Brief | 2026-11-16 Older → AI Pulse Daily Brief | 2026-11-12
Powered by Buttondown, the easiest way to start and grow your newsletter.