UK health and care
University Hospital Southampton has deployed 185 Copilot licences and is taking AVT into the emergency department
HTN, 2026-08-13
The trust's intelligent automation update reports 185 Copilot licences deployed so far, a SharePoint automation hub built to share practice and support adoption, and guidance in development on permitted AI uses to "encourage adoption and safeguard against inappropriate use". The AVT trial has expanded into ED with further onboarding, a business case is written and rollout is scheduled for Q3. National funding bids of around £1m have gone in across AVT and outpatient administration automation. Market engagement on outpatient admin automation is complete, with procurement estimated at three to six months, and RPA work has started in maternity and outpatient partial booking.
Takeaway185 licences is the real unit of a Copilot deployment against a 505,000-seat national headline — and the enablement asset here is a SharePoint hub plus a written permitted-use guide, not a training course.
htn.co.ukHumber Teaching sets an AI scribe target of 90% of clinical notes for a 4–6% admin time saving
HTN, 2026-08-14
The trust's 2026/27 digital plan commits to piloting and rolling out an AI scribe starting in mental health and community services, targeting 90% of clinical notes automatically transcribed and a four to six per cent reduction in administrative time in pilot areas. The wider plan also names cognitive AI for chatbots and virtual assistants, deep learning for outcome prediction and risk stratification, and AI-driven security monitoring. No suppliers are named.
TakeawayThis is the most useful number published this week — a trust putting a 4–6% admin time reduction in writing against near-total transcription coverage, which is far below the time-saved figures suppliers quote. Baseline against this, not against the vendor deck.
htn.co.ukSt George's, Epsom and St Helier puts an AI Centre of Excellence at 2027 and names 200+ fragmented systems as the blocker
HTN, 2026-08-13
The newly integrated group's digital and data strategy to 2030 makes "AI and automation to improve productivity" one of six priorities, with Copilot and other generative AI for drafting and summarising, AVT for clinical documentation, AI tools introduced during 2026 for administrative work, an AI Centre of Excellence launching in 2027, and advanced AI and automation used widely by 2030. The strategy is unusually direct about the starting position: over 200 fragmented systems with limited interoperability, ageing infrastructure, underused data and a history of underinvestment.
TakeawayNote the sequencing — foundations in 2026, Centre of Excellence in 2027. If you are standing up a CoE before the data estate is addressable, you are building the governance layer for a capability that cannot yet ground itself.
htn.co.ukAdoption practice
An acceptable-use policy for AI writing that fits in one rule
Simon Willison, 2026-08-11, on a post by Sophie Alpert
Willison highlights Sophie Alpert's internal policy on engineers using LLMs to write documents, and singles out one rule as crucial: "You must stand behind every idea and every sentence in your docs." The argument underneath it is that there are no lossless transformations of natural-language text — every rewrite by something that lacks your mental model of what you meant loses information. The practical consequence is that "the AI wrote that, ignore it" is not an available defence in review.
TakeawayThis is a one-line AI writing policy you can put in front of a board or a clinical governance group tomorrow, and it is more enforceable than the usual paragraph about human oversight. It applies directly to AI-drafted business cases, SOPs and incident reports.
simonwillison.netOpenAI reports 57% of its business customers running agents on multi-stage workflows, 16% across teams
OpenAI, 2026-08-12
OpenAI's economics research team combined aggregated usage data from its one million business customers with a survey of 9,000 workers across almost 100 enterprises. It reports that 57% of firms deploy agents for multi-stage workflows and 16% have reached processes spanning multiple teams, and that firms in the top 10% of usage generate 8.3× the output tokens per active user of typical firms, use plugins around twice as often and skills roughly six times as often. This is the vendor's own data on its own customers, and the usage-based definition of a leading firm is circular.
TakeawayIgnore the leaderboard framing and take the one transferable measure: cross-team agent workflows are the step almost nobody has made. If your agents all sit inside a single directorate, you are at the same stage as 84% of this sample, not behind.
openai.comCoffee break
DTAC
The **Digital Technology Assessment Criteria** is NHS England's standard front door for assessing a digital health product before you buy it. A supplier completes the DTAC and you, the buying organisation, review it. It covers five areas: clinical safety (has the supplier met DCB0129 and named a Clinical Safety Officer), data protection (DPO, DPIA, DSPT status), technical security (penetration testing, patching), interoperability, and usability and accessibility.
What it does **not** do is the work most people assume. DTAC is a self-declaration by the supplier, reviewed by you — NHS England is explicit that it "applies alongside, but does not replace, other required approvals", and it is not a certification. Passing DTAC does not discharge your DCB0160 duty, does not constitute a DPIA, and says nothing about whether the tool works clinically. It is the minimum baseline for a conversation, and inclusion in national lists such as the NHS England ambient voice technology supplier registry requires it alongside MHRA registration where the product is a device. Note the version: DTAC v2.0 landed in February 2026 with roughly 25% fewer questions, and the previous version became obsolete on 6 April 2026.
