AI Pulse Daily Brief | 2026-09-04
Reading time ~10 mins
OpenAI's newest model loses its low-latency mode under EU data residency, and Amazon's zero-retention terms for its newest hosted model run out on 31 December. ABN AMRO has published its AI register and named the assistants it governs. Europe's cyber agency now expects AI-written code to pass the same checks as human code. Three consultancies, three samples, one conclusion: AI value stalls on operating-model redesign, not on tooling.
Perspectives
A bank's 500% return on AI is the top of the range, not the middle. Media
PYMNTS reported on 3 September that Discovery Bank credits a behavioural-AI system, built on Databricks and Microsoft Azure, with more than 500% return on investment. The same report carries claims of twenty-times-faster data work and more than 300 AI models built per day. PYMNTS then set that case against its own benchmark figures, which show 90 to 100% of enterprises reporting some return from AI but only 5 to 10% reporting full payback. At least half expect to wait five or six years, and both Microsoft and Databricks present the 500% figure as a customer-reported outcome rather than an audited one. The number will travel into vendor decks, and the distribution it sits in will not travel with it.
A Bain survey shows AI budgets still rising while realised savings fall short. Skeptic
Fortune interviewed strategist Amy Webb about companies buying the promise of AI output without budgeting for what integration, governance and workflow change actually cost. The article cites a Bain survey of 951 global companies in which nearly 40% of those measuring AI cost savings came in below 10%, against targets of 11 to 20%. In the same survey, 90% were still increasing their AI budgets. Webb describes one client that ran 14 or 15 generative-AI and agent pilots without scaling any of them, and says executives often cannot say how freed capacity would be redeployed. Budget growth and realised savings have come apart, and most business-case forms carry no field that would tell those two states apart.
AI’s decision dividend requires an operating model, not just a tool rollout Advisory
Tony Moroney’s captured post shares a McKinsey analysis arguing that AI’s largest economic gains need not come primarily from reducing labor. The source frames AI as a way to make cheaper and better decisions: compressing decision cycles, improving the use of existing assets, and allowing organizations to evaluate and capture opportunities before circumstances change. Its central warning is that most organizations have seen little measurable earnings impact because they layer copilots, chatbots, and dashboards onto existing processes instead of redesigning end-to-end workflows around AI.
The analysis describes decision-making as a large but poorly visible operating cost spread across information gathering, analysis, coordination, approvals, meetings, and governance. It recommends starting with economically consequential decisions, measuring decision throughput and outcomes such as asset utilization, revenue growth, and margin improvement, and reusing implementation assets as deployments expand. This is more useful than counting licenses, pilots, or models, which can show activity without showing value. The source also notes that agentic decisions may incur materially higher costs because verification and refinement loops are required, so the economics depend on the execution design and the value of the decision being supported.
For a bank’s Managing Board, the durable question is where decision throughput could change preparation, deployment, procurement, or monitoring stance. The source’s examples point toward planning, pricing, maintenance, resource allocation, and customer operations, while its governance recommendation is explicit: define when AI may act independently, when human escalation is required, and who remains accountable. That turns governance into an operating condition for scale rather than a retrospective approval gate. The analysis cautions that few companies have replicated early benefits at scale and that regulation, capital intensity, switching costs, and change management affect timing. The appropriate stance is therefore to test workflow-level economics and control effectiveness in priority domains, then expand on measured outcomes rather than assume that AI adoption alone creates value. This also means tracking whether faster action actually improves outcomes, whether existing assets are used more effectively, and whether accountability and escalation remain clear as decision authority moves toward AI. A bank can use those measures to distinguish durable operating improvement from a growing inventory of AI activity.
McKinsey & Company (Shared by Tony Moroney)
Netherlands & Sovereignty
Dutch organisations use AI more than their neighbours and embed it least. Advisory
Consultancy Eraneos published survey findings on 31 August covering 658 leaders and teams in the Netherlands, Germany, Spain and Switzerland. Just 3% of Dutch respondents say AI is embedded in their workflows, against 13% across the four countries, and 83% discard 40% or more of the recommendations AI gives them. Only 20% report clear AI guidelines, and 22% clear accountability when AI gets something wrong. Trust rose from 15% where training sat apart from the job to 56% where learning was embedded in daily work and paired with professional development. This is a vendor-sponsored survey and directional at best, but it separates two states that a licence count reads as identical: people using AI, and work redesigned around it.
A Chinese court froze the China stakes of a Dutch chipmaker until 2029. Media
Reuters reported on 31 August that a court in Dongguan froze up to 2.14 billion yuan, about 300 million dollars, in assets held by Dutch semiconductor company Nexperia and its equipment arm. The order covers Nexperia's stakes in four China-based businesses and runs to August 2029 unless suspended. It strengthens the position of Chinese parent Wingtech, which lost control of Nexperia when Dutch authorities intervened last year and is still disputing the outcome. Nexperia said the measures do not affect day-to-day operations, management or business continuity, and both statements are true at once. A supplier review that tests only service continuity marks this company green while its ownership, transferability and exit rights sit frozen for three years.
Industry & competition
ABN AMRO published its AI control set and named the assistants running under it. Corporate
ABN AMRO's public AI page describes an internal AI register, a set of AI standards, and an ethics and compliance framework the bank says is aligned with the European AI Act. It also names the systems those controls cover: Lenny drafts credit proposals, Anna handles customer banking actions, and GAIVA summarises voice-assistant calls and supports anti-money-laundering work. A fourth tool, Advisor Assist, cuts post-call time by up to half on the bank's own account. The page adds virtual-client testing with two partners and the co-founding of an Amsterdam AI hub this year. A Dutch tier-one peer has now put a control shape in public, at a level of detail a supervisor or a board member can quote as the local reference point.
ABN AMRO (publication date unverified)
Santander put 100,000 free AI training places into the open market. Corporate
Banco Santander said on 3 September that it and Coursera are adding 50,000 places to a free training programme, taking the total to 100,000. The courses number more than 80, span AI, data and leadership, and each carries twelve months of access. The offer is open to anyone aged 18 or over in twelve countries, with no requirement to bank with Santander and no degree needed. Santander framed the expansion against a World Economic Forum estimate that 59% of the workforce will need reskilling by 2030. The number is externally verifiable and sized to be quoted, which is what makes it awkward: internal course completions are not counted on a basis that answers the same question.
Innovation
OpenAI's newest model runs slower in Europe than the version being benchmarked. Vendor
OpenAI's model page for GPT-6 Astra says the model is rolling out to enterprise customers through its Trusted Access Program, with wider access described as arriving within days. The page lists a context window of just over one million tokens and up to 128,000 tokens of output. Standard pricing is 10 dollars per million tokens in and 50 dollars per million out, with long-context work costing more. It also states that Fast mode, the low-latency setting, is unavailable when EU data residency is switched on. That single line separates the product being benchmarked from the product a European bank can actually run, so any business case resting on response time needs re-costing.
OpenAI (publication date unverified)
Amazon's zero-retention terms for its newest hosted model expire on 31 December. Vendor
Amazon Web Services announced on 1 September that Claude Fable 5.1 is available on Bedrock, its enterprise model-hosting service. AWS classifies the model as covered by extra safety handling, and under that default it may keep prompts and outputs for up to 30 days for human safety review inside its own boundary. Those records are not shared with Anthropic. Customers eligible under a separate enterprise safeguards partnership can run the model with no data retention at all, but only through 31 December 2026. Whether this model can touch bank data therefore turns on an eligibility question carrying a date, with customer-managed keys and audit logging still described as later steps.
Research
Three consultancies, three separate samples, one conclusion about where AI value stalls. Advisory
Boston Consulting Group, Accenture and the IBM Institute for Business Value published separate studies this quarter and landed in the same place. BCG surveyed 11,749 people and found 42% of regular frontline users saving a workday, while 66% get limited or no guidance on how to use the tools. Accenture asked 3,000 executives and 3,000 employees, and found 23% seeing widespread sustained value, down from 32% earlier in the year. IBM's institute surveyed 1,000 senior executives and found 37% of initiatives delivering what was expected, concentrated where customer-facing work had been redesigned rather than merely assisted. BCG's own split names the bridge: 80% report measurable impact where strategic clarity is strong and tooling thin, against 60% where tooling is strong and clarity thin.
Security
A new open standard would make AI agents inspectable while they are running. Institute
The OWASP GenAI Security Project, part of the open-source community behind the widely used web-security risk lists, published its Agent Control Standard on 1 September. The standard proposes a common set of software hooks and written safety policies that sit between an AI agent and the framework running it. The aim is that an agent's identity, access, actions and stated reasoning can be seen, traced and interrupted while it works, rather than reconstructed afterwards. It is designed to be portable across agent frameworks. That portability is the reason to read it before the next platform decision rather than after, because inspectability designed in at the interface is cheap and inspectability retrofitted is not.
Europe's cyber agency expects AI-written code to face the same checks as human code. Authority
The European Union Agency for Cybersecurity published its Secure by Design and Default Playbook in July, after a consultation drawing on 28 public and private contributors. Across 23 practical playbooks it sets out lifecycle checklists, release gates and evidence expectations, and it makes the AI element explicit. Threat modelling for products containing AI components should cover prompt injection, model poisoning and adversarial inputs, and externally sourced models should have their source, version, integrity and provenance verified before integration. AI-generated or AI-modified code should pass the same peer review, static analysis, dependency scanning and secrets checks as code a person wrote. That last expectation is a rule a bank can write down and evidence to a supervisor within weeks.
European Union Agency for Cybersecurity (publication date unverified)
On the radar
- A supply-chain campaign sent 23 unsolicited code-change requests to open-source projects on 10 August, planting a helper server that behaved normally for three calls before redirecting AI coding agents toward stored access keys, cloud credentials and command history; none were merged through ordinary review. Cloud Security Alliance
- Amazon published a production example of capping AI model spend per person within minutes, tying model-tier limits to a named corporate identity rather than to a shared application key. Amazon Web Services
- The Dutch public-sector algorithm register listed 1,542 algorithm descriptions on 3 September, 41 of them classified as high-risk AI systems, against 1,538 descriptions a month earlier. Het Algoritmeregister
- An independent analyst argued that Europe's export-control chokepoint on advanced chipmaking equipment is held by the Netherlands rather than by the EU collectively, while noting no new restriction has been imposed. Resultsense