NextAI+ Praxis--:----UTCThree long-running columns of close reading — how enterprises actually deploy AI, how it is being governed, and the reports worth an afternoon. Filed weekly, with receipts.
A close reading of how Fortune-500 companies actually wire generative AI into the business. Not press releases — architecture, vendor choice, where the money lands.
The new rules, in plain language. EU, UK, Singapore, US state-level. What changed, who has to act, by when, and what it means for a company that is not a hyperscaler.
The 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.
The UK’s 4 November guidance limits the EU AI Act in Northern Ireland to AI closely integrated with physical products placed on that market, leaving standalone software and online tools outside, with prohibitions and AI literacy from 10 November 2026 and the Annex III and Annex I high-risk requirements from 2 December 2027 and 2 August 2028, the Commission’s Interoperable Europe toolbox gathers standards, use cases, training and procurement guidance while PAIR remains under development, the seventh Japan–EU financial forum names safe AI use and frontier-AI cyber threats for cooperation, the DOJ’s 13 October memorandum merges its January AI task force into DOJ-SITF, and Democrats plan independent testing, liability and data-centre cost bills while September’s ASI ban stays in committee.
The World Internet Conference’s 3 November report would extend safety evaluation from foundation models to deployed systems and allocate responsibility by each party’s actual control, its readiness assessment covering 24 countries on six dimensions and 22 secondary indicators with an evidence cut-off of 31 August 2026, the Cyberspace Administration’s intelligent internet initiative seeks a trust framework across agent discovery, negotiation and execution, Singapore’s Ministry of Manpower confirms reviewing AI output brings mental fatigue, CSA, GovTech and IMDA tell consumers to limit personal agents in application settings and approve important actions individually, and Korea’s PIPC guide explains server-side API authorisation ahead of the annual vulnerability assessment and penetration test covered public systems face from 1 January 2027.
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.
The Bank of England and FCA’s 2026 survey finds 94% of respondents hold firmwide AI policies while ongoing monitoring, leakage testing and recertification are each reported by only around half, 81% lean on vendor assurance for high-materiality third-party systems and just 24% claim deep knowledge of them as the third-party share rises from 35% to 53%, HM Treasury’s Budget proposes £40 million for a National Agentic Security Unit and a Data Centre Covenant tying extra support to five policy tests, NIST’s ten agent identity tenets put permission decisions outside model reasoning and keep append-only records agents cannot edit, and the Commission has replies to every expired AI Act request across more than 30 companies without announcing findings.
Lawrence Wong’s 28 October proposal would start from common incident categories and secure exchange, covering serious failures deployers identify and near misses that trigger no existing reporting duty, Australia’s Digital Duty of Care Bill passed the House on 28 October with proposed section 27 requiring written risk assessments at least annually and six-year retention and section 25E(1)(da) treating simulated emotional intimacy that displaces a child’s human relationships as harmful, Japan’s Digital Agency opened an AI Diet-response prototype to about 180,000 staff who must still verify facts against original sources, and the Cyberspace Administration’s nineteenth deep synthesis batch registers technical support providers without naming a specific application product.
OpenAI’s robotics hiring, DeepSeek’s reported interest in embodied intelligence, Alibaba Cloud’s Qwen-Robot Suite and Google’s Gemini 4 Argon meet Unitree’s model-in-the-loop tests — read against a bracket pick-and-place task to show where foundation models cut the reconfiguration each on-site change costs.
If you want both columns delivered together, four times a year, in one quiet email — leave an address. Otherwise just bookmark this page.