Writing · 27 August 2026
Eight reports, one conclusion.
In 2026, BCG, Deloitte, Stanford, McKinsey, IBM, Accenture, Microsoft and Anthropic each published their major study of AI in the enterprise. Different methods, different samples, different incentives. They arrive at the same finding: the barrier to AI success is not the technology. It is the operating model.
01 · THE CONVERGENCE
Nobody blamed the technology.
McKinsey found that nearly two-thirds of enterprises have experimented with agents — and fewer than one in ten has scaled them to tangible value. Deloitte found three-quarters of organisations aspire to grow revenue with AI while one in five actually does. Stanford counted documented AI incidents rising 57% in a year. Anthropic, whose models sit inside many of these deployments, concluded that the hardest part is not intelligence but “secure and reliable access to production systems”.
Each report then names the same remedy from its own angle: governance designed into the system rather than bolted on, accountability that stays with a person as capability scales, and processes redesigned around the work rather than tools added to legacy ones. Microsoft’s version is the cleanest: 67% of AI’s measured impact comes from organisational factors — twice the contribution of individual behaviour.
02 · THE NUMBERS
Gates are an economic instrument.
The most striking finding in the set is IBM’s. Organisations that embed control into their architecture — governance by design, rather than manual review — deploy sixteen times more agents, spend four times less of their AI budget doing it, and post 18% higher operating margins. In the same study, enterprises averaged 54 agent incidents in a year, and the leaders were separated by what IBM calls portfolio discipline: managing AI as a portfolio, reallocating capital on evidence rather than annual budget cycles.
Read that against Stanford’s incident curve and Accenture’s premise that “responsibility does not scale — it remains distinctly human”, and the shape of the answer is consistent: the organisations winning with AI are the ones that can say, in writing, what was decided, by whom, and on what basis — before the build, not after.
03 · WHAT WE TAKE FROM IT
This is the method, independently priced.
This practice did not start from these reports — the operating model came from running portfolios, and the site describing it predates most of this year’s publications. But the mapping is direct. One front door and staged gates are what IBM calls portfolio discipline. A person signing every consequential write — held as a database gate, not a convention — is the accountability line Accenture and Anthropic say scaling depends on. Redesigning the demand-to-delivery process rather than adding tools to it is the behaviour Deloitte’s transforming third have in common.
Two boundaries, stated plainly because the reports also cover ground we deliberately do not. The governed data model here is decision data — the register, the gates, the trail — not an enterprise data estate; and workforce enablement is not what this practice sells, although the operating model does define who decides what, which is the part of role redesign that belongs to governance.
The eight, with their numbers
BCG — AI Radar 2026
72% of CEOs are now their organisation’s primary AI decision-maker; trailblazers direct ~60% of AI budgets to agents.
Deloitte — The State of AI in the Enterprise 2026
3,235 leaders across 24 countries; 74% aspire to grow revenue with AI, 20% do; skills and governance separate the transforming third.
Stanford HAI — The 2026 AI Index Report
Documented AI incidents up 57% year on year (233 → 362); agent adoption still single-digit per business function.
McKinsey — Building the Foundations for Agentic AI at Scale
Nearly two-thirds of enterprises have experimented with agents; fewer than 10% have scaled to tangible value; 8 in 10 cite data limitations.
IBM Institute for Business Value — 2026 Tech Leader Study
Governance by design: 16× more agents, 4× less budget, 18% higher margins. Average of 54 agent incidents per organisation per year.
Accenture — The Age of Co-intelligence
Only 11% of organisations are equipped for human–AI co-learning; responsibility “remains distinctly human”.
Microsoft — 2026 Work Trend Index
67% of AI’s impact comes from organisational factors vs 32% from individual behaviour; agents in use up 15× year on year.
Anthropic — The 2026 State of AI Agents Report
80% report measurable economic impact from agents; the top blockers are integration (46%), data (42%) and security (40%) — not model capability.
Figures are the publishers’ own, from the editions current at 27 August 2026. They describe the market, not this practice’s results — the distinction matters here, of all places.
Start here
Start with the portfolio you already have.
If the research says the operating model is the barrier, the useful question is what yours would say in writing. Bring the portfolio as it stands.
The first read ends in writing: which of the eleven gaps are live in your portfolio, and the first gate that would expose them.