AI Pulse Daily Brief | 2026-08-25
Reading time ~6 mins
Europe's three financial supervisors want frontier-AI risk folded into the resilience controls banks already run. Amsterdam has written the first practical Dutch guide to the AI disclosure duties that took effect on 2 August. OpenAI moves enterprise ChatGPT to usage-based billing and drops an older model tomorrow. Amazon opens a specification for finding and governing agents across environments. Five independent publishers land on the same finding about where agent value actually comes from.
Top signal
Europe's three financial supervisors fold frontier-AI risk into the resilience rulebook banks already run. Authority
On 31 July 2026 the European Banking Authority, the European Insurance and Occupational Pensions Authority and the European Securities and Markets Authority published a joint statement. It covers the technology risks that frontier AI models create for financial firms. Rather than open a separate AI regime, the three supervisors ask firms to extend the controls they already operate under the EU's Digital Operational Resilience Act (DORA). The named expectations are updated inventories of AI and machine-learning assets, secure-by-design measures, continuous vulnerability monitoring, more frequent testing, and incident and continuity planning. The statement also asks management bodies to review their Risk Appetite Framework metrics and thresholds, and says AI risk is being folded into 2027 oversight planning for critical technology providers.
There is no filing date and no firm-level deadline, which is what makes the 2027 reference the operative one, because the baseline supervisors will measure against is being set now. The inventory and testing work extends machinery the bank already runs. The Risk Appetite Framework threshold review does not, and it sits with the Managing Board rather than the technology function.
Regulatory
Amsterdam writes the first practical Dutch guide to AI disclosure duties. Media
Amsterdam has shared an internal guide with other Dutch municipalities covering the EU AI Act transparency obligations that took effect for many uses on 2 August 2026, Binnenlands Bestuur reports. Staff must disclose AI use when a system talks to residents or employees, generates new content, informs people without a human check, or produces deepfakes. Disclosure stays optional for limited assistive uses such as spelling checks and minor edits that do not materially change the content. The guide is written for product owners, project teams, developers and information managers rather than for lawyers. That line, drawn at whether the AI materially changes what the reader sees, is a workable test for the low-stakes assistive uses already running in customer-contact workflows.
Industry & competition
Thomson Reuters spent $40 million on a legal AI model whose final training run cost $450,000. Media
Thomson Reuters has launched its first in-house AI model as the default engine for table analysis in its legal research product, while outside models keep handling other tasks. The company says it spent about $40 million over two years on people and computing. The final training run cost about $450,000 after starting from a freely available open model and adding its own content and expert review, though SiliconANGLE notes the internal comparisons lack extensive independent validation. The $450,000 figure is the one that gets quoted in build-versus-buy discussions, but the $40 million is the real entry price. The transferable decision is the architectural one, keeping third-party models for every task the in-house model could not demonstrably beat.
Goldman Sachs asks which junior work AI should not be allowed to take. Media
Goldman Sachs is examining how to embed AI in trading and banking work without eroding the apprenticeship model, PYMNTS reports from a bank podcast and CNBC coverage. That model is how junior staff learn judgment from senior practitioners. Goldman's Chris Churchman warned that leaning too heavily on the tools could turn traders into passive operators and dull the reasoning they are hired for, even where near-term profitability improves. The framing treats workforce learning and controlled delegation as part of operating-model design rather than as a productivity programme. Goldman is naming a cost that appears in no AI business case, the tacit judgment juniors gained by doing work an agent now does, and it only shows when the senior bench thins.
Innovation
OpenAI moves enterprise ChatGPT to usage-based billing and retires an older model tomorrow. Vendor
OpenAI has put new enterprise ChatGPT customers on usage-based billing, charging by volume of text processed rather than per message or credit across its chat, work and coding products. The published rate card prices its most capable model at $4 per million tokens of input and $20 per million of output, where a token is the small text fragment a model reads or writes. The same page says the older o3 reasoning model leaves ChatGPT on 26 August 2026 while staying available to developers. Two separate exposures sit here and only one has a clock. Any team that pinned a workflow to o3 loses it tomorrow, while the shift to per-use charging is what makes enterprise ChatGPT cost move with usage at the next renewal.
Amazon opens a specification for finding and governing AI agents across environments. Vendor
Amazon Web Services has published an open specification for discovering AI agents, tools and services across cloud, on-premises and partner environments, released under a licence that lets anyone implement it. Alongside it Amazon previewed a central catalogue that adds approval workflows, access control and search, and lets one organisation's catalogue federate with another while each keeps local control over publishing and revoking access. The federation cuts both ways. The same mechanism that lets a bank find and govern its own agents makes that catalogue discoverable to whatever it federates with. The question to settle is which side of the trust boundary enforces revocation, and that is cheaper to answer while the catalogue is still in preview.
Research
Five independent publishers find agent value gated by process redesign and supervision, not model capability. Advisory
Deloitte surveyed 501 leaders and found 43% expanding agents across functions while only 16% call their business processes ready. Accenture surveyed 510 executives and 2,085 employees, and found 28% able to monitor or audit agent behaviour at scale, with 57% unable to supervise agents in real time. Anthropic's survey of more than 500 technical leaders put integration and data quality at the top of the barrier list. MIT CISR argued that governing an autonomous agent means underwriting a capability rather than signing off a specification. TIAA's operating chief credited years of core-systems work for the 85% daily internal use its AI platform now reports.
Four of these five have no shared commercial interest in the conclusion, which is what makes the agreement worth more than any single survey. The constraint all five name, evidence that an agent can be supervised and audited once it is in production, is the same evidence a resilience control owner has to produce.
Deloitte Insights: Path to Agentic Transformation | Accenture: Generating Impact (publication date unverified) | Anthropic: The 2026 State of AI Agents Report (publication date unverified) | MIT CISR: Five AI Value Propositions | Fortune: TIAA and Accenture on rewiring for AI
On the radar
- Researchers showed that a crafted link could make Microsoft's workplace AI assistant search a user's own mailbox and send what it found to an attacker, and Microsoft says it deployed broader fixes on 18 August with no customer action required. Ars Technica
- A widely used open-source toolkit for building AI applications in Java shipped a flaw that let a hidden instruction trigger a tool the request had never been authorised to use, and patched releases are now available. Spring Security Advisories