AI Pulse Daily Brief | 2026-08-06
Reading time ~5 mins
- About 190 organisations signed the EU's transparency code for AI-generated content, days before the Article 50 marking duties took effect.
- Europe's systemic risk board says the weak point in AI cyber defence is how long a change takes to approve, rather than the tooling.
- Two frontier AI agents were given real unpublished research questions, and the original authors rejected both papers.
- Microsoft's filings show $24.1 billion of revenue from OpenAI and $6.0 billion still owed, Bank of America put a live assistant behind its call handlers, and a Singapore legal group argues agent logs are evidence before they are a control.
Regulatory
About 190 organisations signed the EU's AI transparency code days before the marking rules applied. Authority
The European Commission reported on 31 July that around 190 organisations had signed its transparency code for AI-generated content, days before the AI Act's Article 50 marking duties applied on 2 August. OpenAI, Google, Meta, Microsoft and Mistral are named among the signatories to the provider section. A separate section covers deployers, the organisations that put generative AI in front of customers, and two implementation task forces start in September 2026. A supplier's signature does not discharge the deployer duties. That leaves the bank's model-risk and vendor-management owners two things to establish before September: which contracted suppliers actually appear on the provider list, and whether the bank signs the deployer section itself.
Perspectives
Microsoft disclosed $24.1 billion of revenue from OpenAI, and $6.0 billion still owed. Skeptic
Writer Ed Zitron read Microsoft's 2026 annual disclosure and found $24.1 billion of revenue from its commercial arrangements with OpenAI, plus $6.0 billion that OpenAI still owed at 30 June. From there he estimates OpenAI accounts for about 70% of Microsoft's AI revenue, using a Bloomberg run-rate figure rather than audited segment reporting, so that share is an inference. The receivable is the part that is Microsoft's own reported number. A collection problem there would reach a bank as capacity rationing or repricing rather than as an outage, which is not what availability commitments in a cloud contract are written to cover. That puts supplier revenue concentration and receivables quality on the agenda for the bank's third-party risk owners at the next AI vendor review.
Industry & competition
Bank of America's call-centre assistant now prompts staff live, reportedly saving about a minute per call. Media
Bank of America has upgraded the assistant its service staff use so that it surfaces recommendations during a live call, according to the trade publication Banking Dive. The reported gain is close to a minute off each interaction, with a person still on the line rather than a chatbot in front of the customer. The figure is as reported and has not been independently audited. What matters is which business case it supports: this is a handle-time saving inside a staffed call, whereas deflection removes the call altogether. The two lead to different control, training and headcount plans, so the bank's customer-contact and workforce owners need to say which one their own roadmap is funding before the next planning cycle sets headcount.
Innovation
LendingTree has run a mortgage assistant built from cooperating AI agents in production since late 2025. Vendor
Amazon published a customer account of LendingTree, a US mortgage marketplace, running a set of cooperating AI agents in production on Amazon's enterprise AI platform. The disclosed design layers retrieval from approved documents, platform content and personal-data guardrails, a separate policy classifier, and routing to a human for complex cases. LendingTree reports 1,960 conversations through the first quarter of 2026, with more than 97% completed without a handoff to a person. That measures how often the assistant avoided a handoff rather than how often it was right, and the volume is pilot scale. The transferable part is the control stack, worth putting beside the bank's own agent design review this quarter.
Research
Two frontier AI agents were given real research questions, and the original authors rejected both papers. Institute
Princeton researchers Sayash Kapoor and Arvind Narayanan, writing at AI as Normal Technology, gave two frontier AI agents six days and thousands of dollars of compute each on a genuine unpublished research question. The original authors of those questions reviewed the resulting papers and rejected both. The failure pattern is the usable part: both agents stopped early, left more than half their compute budget unspent, and did not change course after corrective feedback. The authors call the result tentative, since it rests on two cases and reviewers knew the work was machine-generated. It still gives any sponsor running an agent pilot two things to measure that benchmark scores do not show: budget actually used, and whether the agent corrects course when told it is wrong.
A Singapore legal working group finds existing law can absorb AI agent harm, but proving fault is hard. Institute
Singapore's media and technology regulator convened more than 20 lawyers between March and May 2026 and published their conclusions as a 36-page discussion paper, Legal Responsibility for AI Agents. The group finds contract and negligence law can probably absorb most cases where an agent acts on its own and harms someone. It also finds fault is hard to establish in practice, because the evidence sits with whoever built and ran the system and several parties are involved at once. The paper has no binding force, and its practical turn is that agent logs and configuration records are evidence before they are a compliance control. Evidence the bank cannot produce after an incident is loss it keeps.
Infocomm Media Development Authority: Legal Responsibility for AI Agents (LinkedIn; original source not verified) (publication date unverified)
Security
Europe's systemic risk board says AI may help attackers faster than it helps defenders. Authority
The European Systemic Risk Board, the EU body watching for risks to the financial system as a whole, published a July report on frontier AI and cyber capability. It argues that offensive capability, technology dependencies shared across firms and slow internal approval of defensive changes can combine into structural risk for EU finance. The board presents its scenarios as illustrative rather than forecasts. The exposure it names sits in how long a defensive change takes to clear internal governance, rather than in detection tooling. That converts a governance habit into a supervisable number for the bank's resilience owners under the EU's Digital Operational Resilience Act, who can expect to be asked how long an emergency patch actually takes.