AI Pulse Daily Brief | 2026-06-17
Reading time ~12 mins
AFM says Dutch AI Act supervision needs a sharper AFM-DNB split. Parliament moved several AI Act deadlines into 2026, 2027 and 2028. ABN AMRO, Scotiabank and Absa show AI moving through bank operating models, while OWASP, PagerDuty, BCG and a workplace-incident study show governance and security lagging adoption. The common pattern is that AI scale is becoming measurable before ownership, exit criteria and evidence routines have fully caught up.
Top signal
AFM said Dutch AI Act supervision needs a different AFM-DNB split. Authority
Autoriteit Financiele Markten published an 11 June execution test and consultation response on the Dutch AI Act Implementation Act, saying the bill is executable but needs changes for effective supervision. The AFM said its proposed tasks include AI literacy, transparency, prohibited AI practices under AI Act Article 5(1)(a-d), and high-risk credit and insurance systems under Annex III points 5(b) and 5(c). It tied the timetable to 2 August 2026 for transparency duties, 2 December 2027 for high-risk duties, and a Q3 2027 cost evaluation, and said the current budget for four staff members underestimates the work from 2027 onward.
This is the Dutch supervisory architecture becoming concrete, not a general EU compliance story. The stake is direct for credit, insurance, customer-facing AI and AI literacy evidence: the same control set may be reviewed through conduct, prudential, privacy and digital-infrastructure lenses. The AFM's request for cooperation protocols, publication powers, data-sharing rules and extra capacity makes supervision design part of the bank's present AI Act readiness picture. It also narrows the question executives need answered from "when does the AI Act apply" to "which Dutch authority can ask for which evidence."
Security
OWASP said AI agent security has moved into real incidents. Institute
OWASP GenAI Security Project published version 2.01 of its "State of Agentic AI Security and Governance" report in June 2026. The report said AI agents now have production incidents and vendor advisories attached to most major risk categories, including hidden malicious instructions, misuse of connected tools, identity abuse, unsafe agent-to-agent communication, compromised software components and agents acting outside intended bounds. Its control focus is live monitoring while agents run, agent identity, tool permissions, software-component records and clear incident routing. The source is a security standards project, so the report is a synthesis of evidence rather than one new breach disclosure.
The impact is that an AI agent can become an operational actor with permissions, memory, tool access and unclear ownership when something goes wrong. The blast radius covers any business domain that connects agents to internal systems, customer records, software workflows or external tools. For the bank, this turns agent rollout from a platform feature into a security-control question that sits beside operational resilience and third-party risk. The wider the agent can act, the more the control evidence needs to cover what it was allowed to do, what it actually did and who can stop it.
Workers are putting customer and financial data into public AI tools. Vendor
PagerDuty published an 11 June Shadow AI Survey of 1,250 office professionals at large organisations in Australia, Japan, the United Kingdom and the United States. It reported that 66% had used AI tools at work while believing that use was not allowed under company policy, and that 88% had shared work-related information with public tools. The same survey said 34% entered customer data and 31% entered financial information or confidential company documents. PagerDuty also reported that formal and informal consequences had not stopped the behaviour in the surveyed group.
The survey is vendor-published, so the exact prevalence is a medium-confidence benchmark rather than a bank-specific measurement. The security stake is still immediate: ordinary productivity use can become an unlogged data flow to a public AI service. The exposure profile fits any bank function where employees handle customer data, financial information, internal documents or strategy material and have easy access to consumer AI tools. It also shows why enforcement alone is an incomplete control if approved alternatives do not fit the work.
Regulatory
Parliament moved several AI Act dates while keeping formal Council adoption open. Authority
The European Parliament gave final approval on 16 June to digital-omnibus changes to the AI Act, with 423 votes in favour, 57 against and 174 abstentions. The approved text moves stand-alone high-risk AI obligations under Article 6 and Annex III to 2 December 2027, safety-component high-risk systems covered by sectoral product law to 2 August 2028, and AI-generated-content labelling duties under Article 50 to 2 December 2026. It also adds a ban tied to Article 5 for systems that create child sexual abuse material or non-consensual intimate content, with alignment due by 2 December 2026. The press release said most AI Act provisions still start applying on 2 August 2026 and formal Council adoption is still required.
The practical signal is a changed compliance clock, not a reduced compliance perimeter. Credit scoring, customer interaction, documentation, transparency and evidence routines remain in scope, but their dates now sit across three planning horizons. The Council dependency also matters because executive planning can move on the new dates while legal finality has one step left. For a bank, the revised sequence separates near-term transparency and content-labelling work from the later high-risk system evidence package. It also creates a clearer audit trail for why roadmap dates changed this month.
Perspectives
WPP tied AI value to redesigning work across roles and leadership. Corporate
Fortune profiled Laura Weis, WPP's head of human-AI strategy and transformation, on 16 June. Weis argued that companies lose durable AI value when technology strategy and people strategy run in separate tracks, and said executive teams need a joined human-AI workforce strategy. The article framed AI value as a question of role design, judgment, discernment, decision-making and how time freed by AI is used. It also pushed against treating efficiency as the main value case, arguing that AI-created time needs an explicit destination.
This is a practitioner lens on the same pattern visible in today's research signals: adoption alone does not prove value. The stake for business-domain leaders is benefit credibility. AI initiatives that report tool usage but do not specify the work redesign, role change or business outcome behind the productivity claim remain hard to govern and hard to compare.
Gartner expects many enterprise agents to be demoted or removed by 2027. Media
ZDNet reported on 12 June that Gartner expects 40% of enterprises to demote or decommission autonomous AI agents by 2027. The article said governance gaps are being discovered only after production incidents, and pointed to weak repeatable frameworks, poor context about what data means, data governance and expert oversight. The source is a secondary report of Gartner's forecast, so the exact figure carries medium confidence. Its strongest signal is the existence of a forecast large enough to make agent retirement part of the enterprise AI conversation.
The useful lens is lifecycle discipline. AI agents are being discussed as a permanent operating-model layer, but the failure mode is already visible: pilots can become production systems before anyone defines when to stop, demote or roll back the agent. That makes agent portfolio hygiene a current governance issue rather than a 2027 cleanup problem.
Industry & competition
ABN AMRO's CEO made AI an everyday operating-model story. CxO voice
ABN AMRO published a 16 June newsroom item summarising CEO Marguerite Berard's Money20/20 remarks on AI in banking. Berard framed AI as part of daily operations, and ABN AMRO repeated that more than 85% of its employees already use AI in daily work. The bank highlighted leadership by example, broad employee adoption, guardrails built on trust, human control, data and process foundations, internal innovators, AI-amplified risks and the ambition to rethink banking. The source is ABN AMRO's own newsroom, so the adoption figure is primary but self-reported.
The competitive signal is the public operating-model narrative from a direct Dutch peer. ABN AMRO is connecting employee use, executive sponsorship, controls and business redesign in one CEO-level message. That framing sets a visible comparison point for how Dutch banks explain AI adoption to boards, staff, customers and supervisors.
Scotiabank expanded employee AI tools to 71,000 staff. Corporate
Scotiabank announced expanded Scotia Intelligence capabilities on 10 June, focused on client experience, speed, efficiency and risk management. The bank said more than 71,000 employees now have access to assistive AI tools, 5,500 engineers use AI for coding productivity, and employee use of AI to reply to client questions rose 30% quarter over quarter. It also positioned the tools across risk management and client-service use cases rather than only personal productivity. The figures are primary but self-reported adoption metrics.
This gives a peer-scale benchmark for enterprise access, engineering adoption and customer-response use. The bank relevance is not that Scotiabank's tools are directly transferable; it is that large-bank AI adoption is being measured in workforce reach, engineering habits and customer-service workflows. Those are the same measurement categories that make AI adoption legible beyond isolated pilots.
Absa reported SME onboarding fell from seven hours to under 20 minutes. Vendor
Salesforce reported on 15 June that Absa Bank deployed three production AI agents across 12 African countries for SME banking, product support and incident management. The vendor-published case said SME onboarding fell from about 420 minutes to under 20 minutes, productivity rose 85% year over year and digital new-to-bank account acquisitions doubled. It names production workflows rather than a generic assistant rollout, which is why the case survived source caution. Because Salesforce is the source, these figures are low-confidence deployment claims rather than independent outcome evidence.
The item still earns a place because the use case is concrete, banking-specific and quantified. SME onboarding is a domain where time, document completeness, customer drop-off and control quality can all be measured. Even with source caution, the case sharpens the question of whether AI agent benefits are being tracked against full process outcomes rather than narrow task speed.
Research
BCG found agent use doubling while human-AI team governance lags. Advisory
Boston Consulting Group published "AI at Work: Strategy Matters More Than Tools" on 3 June, based on 11,749 respondents across worker levels, markets and industries. BCG found that AI agent integration into workflows rose from 13% in 2025 to 30% in 2026, while 50% of respondents said their company had not put clear guidance in place for managing human-AI teams. It also found that 42% of regular AI-using frontline employees save at least one workday per week, while 66% receive limited or no guidance on what to do with saved time. The report's operating model separates tool deployment, end-to-end work redesign and new business-model invention.
This is a medium-confidence single-publisher survey, but the pattern is durable: value is leaking between adoption and operating model. The report gives hard numbers for a familiar executive problem, namely that saved time, agent use and frontline adoption do not become business value without accountability, workflow redesign and value measures. That makes it useful quarterly material as well as a daily operating signal. It also aligns with today's WPP and workplace-incident signals: the bottleneck is increasingly the design of work around AI, not access to AI tools. The repeat across independent source types strengthens the editorial weight of the theme.
Boston Consulting Group: AI at Work: Strategy Matters More Than Tools
A workplace-AI study found most incidents start with poor task fit. Institute
ACM FAccT / arXiv published a workplace-AI incident study by researchers from King's College London, Nokia Bell Labs, the University of Oxford and Politecnico di Torino. The authors reviewed AI Incident Database material covering 1,256 incidents and 6,163 news reports, narrowed the set to 214 workplace AI incidents and 482 task occurrences, and found that 83.4% involved a mismatch between what the AI system did and what workers needed. The study also attributed 73.6% of task occurrences with mismatched AI to developer preferences that differed from worker preferences. The authors found a shift after generative AI adoption, with more incidents linked to systems that were too imaginative for precision-heavy work.
The finding shifts AI failure from model capability to work design. In regulated workflows such as HR, legal review, customer service, data analysis and engineering, a tool can be fast and still be wrong for the job if workers need precision, explanation, personal context or tolerance for edge cases. That makes worker-task fit a governance control, not a user-experience afterthought. It is also a useful corrective to vendor demonstrations, because the failure pattern starts before model output reaches a dashboard.
ACM FAccT / arXiv: Workplace AI incident study
On the radar
- KPMG and INSEAD framed AI governance as two-speed board work: immediate control readiness on one track, and longer-term transformation, workforce and competitive-position choices on the other. KPMG in Malaysia
- Richard Turrin argued that strong 2025 bank shareholder returns may mask delayed transformation on AI, non-bank competition and digital assets, but this is a single LinkedIn practitioner commentary rather than a primary report. LinkedIn (LinkedIn; original source not verified)