NextAI+ Praxis--:----UTCThe reports worth an afternoon, taken seriously at length — on AI, organisations, interaction, and the enterprise. One document per entry. We read the appendices so the argument holds.
Stanford’s SALT Lab read nearly 250,000 Claude conversations: 72% are human-led, 49.7% carry friction, and 78.7% of those show the user actively recovering. High-stakes tasks run 12.4 turns against 6.5 for ephemeral ones — so the judgment and correction people perform needs named owners, evidence and time.
AIUC co-founder Rune Kvist argues the binding constraint on AI procurement is trust, not capability — and answers it with a quarterly-updated security standard, third-party testing and Lloyd’s-backed cover. Copyright stays the hardest risk to underwrite, because downstream users cannot see what a model was trained on.
Ireland’s SOHAM team surveyed 487 employees, ran a 421-participant experiment and interviewed 32 executives: 76.8% use AI at least weekly, yet 78.9% report little change in working independently. Explicit authorisation moves willingness far more than skill does — and verification needs owners.
McKinsey’s 1,719-respondent survey finds enterprise-wide scaling up to 44% and 80% of users reporting higher productivity — while only 37% see any contribution to EBIT. Nearly three-quarters of AI high performers have fundamentally redesigned workflows; among everyone else, about a quarter have.
Emergence AI ran eight single-model agent worlds for 16 days and hit them with phishing, disinformation, and memory-leak attacks. Every world recognized the attack; almost none contained it — and persistent memory, population homogeneity, and the finding that model alignment is not compositional move the safety boundary up to the system.
Kyndryl’s 2026 People Readiness Report finds efficiency gains arrive first while innovation and revenue lag behind, and traces the gap to four organizational foundations — work redesign, formal change management, workforce capability, and governance that follows agents into production — in the order Pacesetters execute them.
OpenAI’s Enterprise Signals data show agentic output overtaking chat and the frontier-firm usage gap widening from 2.6× to 8.3× — and the context, tools, and continuous execution enterprises must give agents to delegate real work.
BCG finds only about 5% of companies earn substantial financial value from AI and traces 70% of AI value to workforce transformation — leadership, behavior change, workforce planning, and an AI-enabled operating model — with a closer look at its human cost.
Three Stanford HAI student affinity groups — on cognitive autonomy, narrative-level misinformation, and art in the age of AI — show AI safety widening from technical systems to the integrity of cognition, information, and trust.
Starting from Deloitte’s AUD 440,000 report refund, a reading of the Stanford AI Index 2026 chapter on Responsible AI — output quality, system controls, and accountability — and the three-layer self-assessment enterprises should complete before AI goes live.
A close reading of Microsoft’s 2026 Work Trend Index: as AI and agents take on more execution work, individual agency expands — and the question becomes whether organisations are built to capture that opportunity, from Frontier Professionals to an agent infrastructure that can be evaluated and improved.
A close, weekly reading of how large enterprises actually wire AI into the business. Architecture, vendor choice, where the money lands. Filed every Thursday.
The rules, in plain language. EU, UK, Singapore, US state-level — what changed, who has to act, by when. Filed every Wednesday.
If you want both columns delivered together, four times a year, in one quiet email — leave an address. Otherwise just bookmark this page.