Reading time ~5 mins

Accenture becomes the sixth major research house in three months to find that governance and workforce readiness, not model access, is what holds AI agents back. Microsoft has put a date on charges for a part of its AI hosting that is free today. MITRE has drafted a way to keep proving that a running AI system is still the approved one. A three-year study locates stalled AI plans inside the leadership team, and a panel of 272 AI experts ranks finance and insurance among the sectors most exposed to AI risk.

Perspectives

Three-year study locates the brake on AI plans inside the leadership team. Institute

Harvard Business Review published research on 11 August by Morgan Blangeois and Thomas Roulet of Cambridge Judge Business School. It draws on three years of work inside eleven European IT services firms, with 23 interviews and three structured leadership workshops. The authors report that unresolved disagreement within the leadership team, rather than employee fear or missing skills, was what held back adaptation to AI in the firms they studied. The evidence is qualitative, confined to one sector, and the article body sits behind a paywall. The finding still lands here because a slipped AI initiative in a business domain is almost always explained as readiness, and a readiness diagnosis leaves disagreement at the sponsor level unmeasured and untreated.

Harvard Business Review

Innovation

Microsoft will start charging on or after 21 August for a part of its AI hosting that is free today. Vendor

The pricing page for Azure OpenAI, the service that runs OpenAI models on Microsoft's cloud, states that charges for cache writes are not yet active and are expected to begin on or after 21 August. A cache write is what the service does when it stores a block of reused text so the same context does not have to be sent and processed again. Microsoft has published the date but not the price. That gap matters for anyone running assistants or agents on the platform, because those workloads write to the cache constantly. Once billing starts, the new charge and ordinary growth in usage become indistinguishable on the invoice, unless a volume baseline was taken beforehand.

Microsoft Azure (publication date unverified)

Research

A sixth research house in three months lands on governance, not model access, as the limit on AI agents. Institute

Accenture surveyed 650 senior corporate development and private equity dealmakers across 12 industries and 24 countries, and found that data interoperability, clear accountability and workforce readiness are what constrain agentic AI in practice. Two thirds said interoperable data systems strongly affect success while only about 20% used structured data well at the underwriting stage, and 47% said explicit human decision rights would materially raise their willingness to deploy. The number that is new is the readiness split: 54% of C-suite respondents judged their workforce ready to work alongside AI agents, against 30% of the people doing the work.

Five other houses reached the same conclusion within 90 days from separate data sets, which moves this from one firm's opinion to a repeatedly measured pattern. The 24 point gap is the part that bites, because an executive sponsor's assurance that a domain is ready is the evidence most agentic business cases actually rest on.

Accenture: The Dawn of the Agentic Deal (publication date unverified) | McKinsey & Company: State of AI trust in 2026 | Boston Consulting Group: AI at Work | IBM Institute for Business Value: The blueprint for agentic operations | Bain & Company: Your AI Budget Is Growing, Your Returns Aren't | Grant Thornton: 2026 AI Impact Survey Report

An expert panel ranks finance and insurance among the sectors most exposed to AI risk. Institute

A three round expert elicitation with 272 AI researchers, presented by one of its authors, asked participants to assess how vulnerable each sector is across 24 categories of AI risk. Finance and insurance came out among the five most exposed, alongside information technology, national security, health care and education. The panel put a probability of 10% or higher on catastrophic outcomes in 18 of the 24 categories within five years if nothing changes, and five stayed above that line even after realistic mitigation. The underlying preprint was not retrieved, so this rests on the author's own account of the work. The ranking is worth knowing about regardless, because a short list that names finance travels into board conversations with none of those caveats attached.

YouTube: three-round Delphi study on sector exposure to AI risk

Security

A draft framework sets out how to keep checking that a running AI system is still the approved one. Institute

MITRE, the American non-profit behind several widely used security frameworks, released a public review draft on 16 July. It extends integrity checking from a single test at start-up to a repeated one, by default every sixty seconds. Its AI section sets out how to prove a model's weights and origin are unaltered, restrict the documents a model may consult to a registered list, and measure the safety guardrails separately. It also proposes that a failed check should block a workload from renewing its credentials rather than waiting on a human incident call. The draft is voluntary and leaves model drift outside its boundary, which is the assurance gap any institution adopting it would still have to give an owner.

MITRE