RH.
← The Adoption Feed

Saturday, 15 August 2026

96% of surveyed NHS digital leaders say the national AI strategy does not reflect their priorities

HTWorld, 2026-08-12

Dr Harry Thirkettle, director of health and innovation at Aire Logic, surveyed 50 senior NHS digital and clinical leaders. Of those, 96% said the national AI strategy does not reflect their priorities, and 6% rated their organisation well prepared to adopt AI at scale; 48% named back-office operations as where AI will tell. Supplier survey, sample of 50.

Takeaway

The gap is absorptive capacity, not procurement — if your plan assumes the national strategy is sequencing for you, your peers disagree.

htworld.co.uk

UK health and care

Lewisham and Greenwich puts corporate-function AI in years one and two, agentic AI in year three

HTN, 2026-08-11

Its five-year strategy puts "pilots for AI and automation in corporate functions" in years one and two, agentic AI in year three, scaled automation by year five — all behind a 2027 EPR go-live. Blockers named: technical debt, split community/acute EPR.

Takeaway

Fourth trust in eight days to sequence AI behind the EPR; note the first tranche is deliberately corporate, where no DCB0160 case is needed.

htn.co.uk

Leeds Teaching Hospitals licenses TANGO, because stain variation breaks AI pathology

HTN, 2026-08-11

Leeds Teaching Hospitals has licensed TANGO — a slide-staining consistency tool from its National Pathology Imaging Co-operative — to Epredia for international deployment from late 2026. It exists because staining variation between labs degrades AI pathology tools reading those slides.

Takeaway

Ask any imaging or pathology AI supplier what input variation it was validated against — a model that works in Leeds can fail in your lab.

htn.co.uk

NHS England is the fifth-largest AI buyer in UK government at £384m, half of it one Palantir contract

UKAuthority, 2026-08-14

Tussell's tracker counts 2,129 AI contracts worth £5bn across UK government since 2018 — 4% of IT spend. NHS England ranks fifth at £384m, of which £182m is the Palantir Federated Data Platform deal.

Takeaway

Strip out that one deal and eight years of national AI buying is ~£200m — corporate automation is still yours to scope, fund and buy locally.

ukauthority.com

Adoption practice

HM Treasury put Copilot on all 2,000-plus staff and published the enablement scaffolding

PublicTechnology, 2026-08-14

HM Treasury rolled Microsoft 365 Copilot to all 2,000-plus staff in 2025/26, alongside its own HMT-GPT. Around 1,000 attended "Introduction to AI" sessions, backed by a prompt guide, use case library, community of practice and an AI champions group.

Takeaway

Copy the scaffolding, but note a department that trained half its workforce still reports productivity gains with no measured figure behind them.

publictechnology.net

HMRC's AI assurance deputy director explicitly does not absorb the existing control functions

PublicTechnology, 2026-08-12

HMRC is recruiting a deputy director for AI assurance at £86,000–£110,000 in its Chief AI Office. The advert states the post coordinates and "will not replace existing specialist functions such as data protection, cybersecurity, legal, commercial, architecture, risk or ethics".

Takeaway

Lift that sentence into your Centre of Excellence role descriptions before someone drafts a post making one person accountable for IG, clinical safety and cyber.

publictechnology.net

The tools we use

Microsoft 365 Copilot

Authoritative Sites lets you tell Copilot which SharePoint sites to trust — and it is off by default

Microsoft, GA June 2026, carried in the 11 August release notes

Admins can mark a SharePoint site as authoritative, ranking its content higher in Copilot Chat and Copilot Search. GA, needs a Microsoft 365 Copilot (Premium) licence, off by default, and set only via SharePoint Online PowerShell (`Set-SPOSite -IsAuthoritative $true`).

Takeaway

The cheapest fix for "Copilot quoted an out-of-date policy" — flag your policy library and intranet; it is still off in your tenant today.

learn.microsoft.com

The Copilot credit consumption dashboard is on by default and visible to managers

Microsoft, GA June 2026, updated 13 July 2026

Viva Insights carries a Consumption Dashboard showing Copilot credit usage by group and team (RM566302). Enabled by default, it populates only where usage-based billing is configured, and is visible to analysts, global admins and any manager with five reports.

Takeaway

You have per-team AI spend visibility for free, and you have devolved it to every line manager with five reports — decide if that was intended.

learn.microsoft.com

Claude

Anthropic is watermarking Claude's text output, globally, with no detection tool yet

Anthropic, 2026-08-14

Anthropic is embedding an invisible watermark in Claude's text output, based on DeepMind's SynthID-Text and applied globally for future models. A detection API is promised but unavailable; it works poorly on short or factual text, and a rewrite removes it.

Takeaway

Do not write this into policy — undetectable today, defeated by a rewrite, exempt on short text; keep the accountable-author rule instead.

anthropic.com

Voices

Dr Harry Thirkettle, director of health and innovation, Aire Logic

HTWorld, 2026-08-12

Writing up his firm's survey of 50 NHS leaders, Thirkettle argues the constraint is the organisation's ability to absorb what it buys, not tool supply, and that there is "no procuring your way out of" it. He is a supplier.

Takeaway

Put an absorption line in your next business case — enablement staff, backfill for change, support model — or you are costing licences, not the change.

htworld.co.uk

Katryna Peart, AI governance researcher and AI historiographer

PublicTechnology, 2026-08-12

Peart argues council reorganisation removes the people who would catch an AI tool misreading the inherited record. Her testing of council documents found summaries turned contested decisions into consensus — "narrative flattening" that procurement checks miss.

Takeaway

Your group's shared policy library is a merged corpus; before pointing Copilot at it, test summaries against contested source documents, not just for factual accuracy.

publictechnology.net

Coffee break

Software as a Medical Device

If an AI tool is intended to diagnose, prevent, monitor, treat or alleviate disease, it is a medical device in law. **Software as a Medical Device (SaMD)** covers software doing this on its own, without being part of a physical device; the MHRA uses **AIaMD** where that software is AI or machine learning.

The test is *intended purpose*, not the technology. This is why the MHRA's July 2026 ambient voice statement lands where it does: a tool that transcribes, summarises a consultation or suggests codes for a clinician to review is not a device; one intended to support diagnosis or treatment, or that acts without clinician review, is. The same product can cross the line when a supplier ships a feature.

A device needs UKCA marking for the GB market, with conformity assessment by an Approved Body above the lowest risk class, plus post-market surveillance. CE marking remains acceptable transitionally. The MHRA's **AI Airlock** sandbox has run two phases with 11 innovators; Phase 3 is in design.

Also worth reading

  • MHRA's 29 July clarification: AVT transcription and coding suggestions are not medical devices; diagnosis or unreviewed action is. gov.uk
  • CLEARvalidate expressions of interest close 5pm Friday 21 August. Covered here 10 August; flagged because the deadline is now live. clearprogramme.org.uk
  • Doug Turnbull's trick for classifying against large fixed vocabularies: let the model invent tags, then map them with embeddings. simonwillison.net
  • Anthropic opened skill and plugin scanning in beta for Claude Enterprise on 6 August — relevant if staff publish skills internally. support.claude.com
  • NHS England's GIRFT programme has opened a £450,000 call for ICB community waiting-time interventions — a route for a triage pilot. htn.co.uk
  • 74 pooled trials: digital tools improved medication safety 46% and cut patient harm 17%, strongest with audit and feedback. psrc-gm.nihr.ac.uk
  • An admin reports Copilot audit record types 284, 373 and 328/383/384 not generating despite Purview Audit enabled. techcommunity.microsoft.com
  • Unanswered thread: Agent Builder in Copilot Chat still lets users create agents after Copilot Studio is disabled at licence level. techcommunity.microsoft.com