AI Pulse Daily Brief | 2026-08-12
Reading time ~5 mins
ABN AMRO signed a strategic partnership with the French model developer Mistral AI, stating it wants to reduce dependence on non-European technology suppliers.
Microsoft's own testing puts Excel's new autonomous mode at 57% accuracy against 71% for people, and it reaches the bank through existing licensing rather than a procurement gate.
Google DeepMind published a worst-case threat model for the AI agents running inside its own operations, and separately moved agent guardrails into its hosted developer service.
The European Commission's public compute build-out now lists 19 sites, one of them Dutch. Boston Consulting Group puts the failure of procurement agents on the operating model rather than the technology.
Perspectives
Microsoft's own tests put Excel's autonomous mode well below human accuracy. Advisory
A consultant's walkthrough of new assistant features in Excel closes on a figure that outweighs the feature list. Microsoft tested Agent Mode, which builds and edits workbooks from a written instruction, across all 912 tasks in SpreadsheetBench, a public test of real spreadsheet work. It scored 57.2%, against about 71.3% for human testers on the same tasks. The controls that matter are plan mode, which makes the tool state its approach and wait for approval before editing, and grounding, which ties its numbers to the reporting layer the board already sees. This capability arrives inside existing Microsoft licensing, so it can reach month-end reporting without ever passing a procurement gate.
LinkedIn (LinkedIn; original source not verified) | Microsoft
Netherlands & Sovereignty
Europe's public AI compute build-out now lists 19 sites, including one in the Netherlands. Authority
The European Commission updated its AI Factories page on 7 August. It now lists 19 AI Factories and 13 smaller associated sites across Europe, with the Netherlands among them. At least nine new AI-optimised supercomputers are to be bought and installed, which the Commission says will more than triple the public AI computing capacity currently available through the EU's joint supercomputing programme. A separate call covers up to seven much larger facilities, each targeting more than 100,000 advanced AI chips, backed by up to €10 billion in EU and national funding. The page sets out scale and money but no access model, service levels, or pricing, so none of this yet replaces a single line of hyperscaler capacity.
Industry & competition
ABN AMRO signs a strategic partnership with Mistral AI to reduce non-European technology dependence. Media
ABN AMRO announced on 7 August a strategic partnership with Mistral AI, the French model developer, to explore and build AI applications for the bank. Chief Innovation and Technology Officer Carsten Bittner framed it around Europe staying digitally resilient, competitive and strategically autonomous. The bank said it wants AI capability developed and governed in Europe, reducing its reliance on suppliers outside the region. The announcement named no specific use cases, and described joint exploration and development rather than a production deployment. That distinction is the substance here: a named Dutch peer has put a sovereignty rationale on the record as a sourcing principle, well before it has a system to point at.
Innovation
Google moves agent guardrails inside its own hosted service. Vendor
Google DeepMind updated the managed agent service in its Gemini developer platform, which runs AI agents inside a protected environment that Google itself operates. The update adds hooks that can block, check, or record what an agent does as it works. It also adds a spending ceiling that pauses a run at its limit instead of discarding the work, scheduled triggers that reuse an environment across runs, and free access for experimentation. The direction matters more than the feature list: guardrails a bank would otherwise build, own and evidence itself are moving inside the vendor's service. That relocates the control evidence a supervisor would ask to see, from the bank's own logs into a third party's.
Research
Boston Consulting Group puts the failure of procurement agents on the operating model, not the technology. Advisory
Boston Consulting Group published a 32-page executive perspective on 6 August arguing that procurement should move from isolated assistants to supervised agent workflows. Its foundation is six elements: connected data, a knowledge layer, AI talent, organisational design, process governance, and cross-functional collaboration. The firm splits the work explicitly, keeping people strategy, exception handling, negotiation and performance management with humans while agents run routine workflows. It attributes 70% of transformation success to people, organisation and process, 20% to technology and 10% to algorithms. The report mixes its own analysis with client cases, so the part a bank can reuse is the decision-rights and exception-handling design, not the savings figures, which are not independently validated.
Boston Consulting Group: AI-First Procurement: How Autonomous Agents Drive Competitive Advantage
Security
Google DeepMind publishes a worst-case threat model for its own internal AI agents. Vendor
Google DeepMind released the first version of a control roadmap for the AI agents running inside its own operations, borrowing from security practice and assuming a hostile agent pursuing goals its operator never intended. It sorts the resulting threats into three classes, starting with loss of control, such as an agent running itself outside any sanctioned deployment. The others are sabotage of the organisation's own safety and alignment work, and direct harm such as destroying or stealing critical assets. It commits to two defensive properties that are meant to hold even as agents grow more capable. The reference point matters because a company that sells these models is setting the vocabulary supervisors and auditors are likely to borrow when they ask how a bank contains them.
LinkedIn (LinkedIn; original source not verified)