Download this edition as PDF
We'll email a 6-digit access code. Enter it to unlock the Daily Market Scan PDF.
Two claims about this month are wrong
Frontier & Industry Intelligence : Regulated Sectors - Financial Services, Healthcare, Manufacturing, Energy. Both errors are circulating in trade press and both are load-bearing for compliance roadmaps being drafted right now.
Six items that mattered, 10 to 15 August 2026
A week weighted toward distribution and physical AI rather than model launches. The single most consequential development is not an announcement at all, it is a widely repeated regulatory claim that does not match the statute.
Article 50 transparency is the only EU AI Act duty that changed status this month
The transparency and information obligations under Article 50 became generally applicable and enforceable by national competent authorities on 2 August 2026. They apply whether or not a system is high-risk, provided it talks to users, generates image, audio, video or text, or infers emotion or biometric categories. Penalties reach €15 million or 3% of worldwide annual turnover, whichever is higher.VERIFIED C01
LG and NVIDIA sign humanoid MOU
Signed 13 August 2026 at NVIDIA headquarters in Santa Clara. LG's next-generation bipedal humanoid runs Jetson Thor for onboard compute and is built on Isaac GR00T, unveiling targeted for the first quarter.VERIFIED C09
Unitree H1 Pro completes its three-region rollout
Europe on 22 July, Asia on 5 August, North America on 12 August 2026. A sequenced regional launch rather than a single global date is a supply and certification signal, and it sets lead-time expectations.CITED C11
Agent distribution moved into the tools people already have
Salesforce made the rebuilt Slackbot generally available to Business+ and Enterprise+ customers, placing agent access inside the messaging surface rather than a separate console. Anthropic released Cowork on Monday 10 August, extending Claude Code style agents to non-technical users.CITED C38CITED C17
Engineering and manufacturing get a dedicated agent platform
L&T Technology Services announced AgenticIQ on 11 August 2026, an agent platform aimed at engineering, product development, industrial operations and customer experience. Notable because it targets the manufacturing workflow directly rather than generic knowledge work.CITED C37
Scheduled retirement worth diarising: 20 August 2026
The Grok 4.1 model family on the Gemini Enterprise Agent Platform is deprecated and shuts down on 20 August 2026. Gemini 3.5 Flash was removed from the global region in the Gemini Enterprise app on 4 August 2026. Two retirements inside three weeks on a single enterprise platform.VERIFIED C19
Architect's note: four of the six items are distribution or lifecycle events, not capability events. Capability is no longer the scarce input. Getting a capable model safely adjacent to a regulated workflow is.
Two claims in wide circulation that will not survive an audit
We check every number against a primary source. This week that caught two errors being repeated across trade coverage and vendor content. Both are the kind that ends up in a board paper.
Source C01, C02, C03, C04 and C05. The only obligation that changed status on 2 August 2026 is Article 50 transparency.
Claim one: "EU AI Act high-risk obligations became fully enforceable on 2 August 2026"
This is incorrect. 2 August 2026 was the original date. The AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved Annex III high-risk obligations to 2 December 2027.VERIFIED C04 What became applicable on 2 August is Article 50 transparency, and only that.VERIFIED C01VERIFIED C02VERIFIED C03
Why it matters. Two failure modes follow, in opposite directions. A team that believes high-risk duties are live will spend Q3 and Q4 budget on conformity work before the harmonised standards settle. A team that later finds out about the deferral may stand the programme down entirely and lose the sixteen months of runway the Omnibus handed them.
Claim two: "The FSB published its first operational framework for governing agentic AI"
This overstates the document's status. On 10 June 2026 the FSB published a consultation report with 12 sound practices. Responses closed 22 July. The final report is due October 2026.VERIFIED C06
Why it matters. A consultation report is not a supervisory expectation. The report accepts that continuous human review of individual agent decisions does not scale, pointing instead toward AI that monitors other AI. That is the right design direction and worth building against now. Citing a consultation as settled policy in a regulatory response is a credibility risk, and the text can still change.
The practical read
The deferral is runway, not relief. Article 50 applies horizontally, so the cheapest obligation to breach right now is disclosure, not risk management. Marking and user-facing disclosure are the things to have finished. The high-risk file is the thing to have designed.
One date deserves a diary entry ahead of the high-risk work: generative systems already on the market have until 2 December 2026 to meet the machine-readable marking requirement under Article 50(2).CITED C05 That is 109 days out, and it lands on the installed base, usually the hardest surface to retrofit.
Equal-weight read across the labs regulated buyers actually shortlist
Same two questions of each provider: what shipped or was scheduled in the last two weeks, and what does it change for a buyer in a regulated sector.
| Provider | What moved | Regulated-buyer read |
|---|---|---|
| Anthropic | Cowork released Monday 10 August, extending agent capability to non-technical users. Enterprise plans gained beta security scanning for third-party skills and plugins. Self-hosted runners let Claude Code sessions execute on customer-controlled machines. Claude in Chrome is off by default on Enterprise.CITED C17CITED C18 Cognizant and Anthropic expanded their enterprise partnership on 27 July 2026.CITED C40 | The self-hosted runner and default-off browser posture are the two items a bank or hospital security review will care about. Both reduce the data-egress surface that usually stalls approval. |
| OpenAI | Presence launched as an enterprise platform for realtime voice and chat agents that take approved actions and escalate to a person.VERIFIED C15 GPT-5.6 shipped with native multi-agent orchestration and programmatic tool calling. Daybreak expanded with GPT-5.6-Cyber.CITED C16 | Voice is the channel with the heaviest consent and recording exposure in financial services and healthcare. Escalation-to-human as a first-class primitive is the feature to test, not the latency. |
| Two retirements inside three weeks on Gemini Enterprise: Gemini 3.5 Flash removed from the global region on 4 August, Grok 4.1 family shutdown on 20 August. Gemini Enterprise for Customer Experience aimed at agentic retail.VERIFIED C19 | Model lifecycle velocity is now a procurement clause, not a footnote. Any agent pinned to a specific model version needs a documented re-validation path with a named owner and a date. | |
| xAI / SpaceX | Grok 4.6 is on the xAI API with a 500k context window and $2 per million input tokens, roughly half the list price of comparable frontier models.CITED C12 grok-voice-think-fast-2.0 shipped speech-to-speech from 5 August. Grok Bot launched as persistent cloud agents.CITED C13 | Price per token at this tier changes the economics of high-volume document and claims work. The open question for regulated buyers is data residency and retention terms, which the pricing pages do not settle. |
| Cursor / Anysphere | Reported agreement for SpaceX to acquire Anysphere in an all-stock transaction, expected to close in the third quarter of 2026.FLAG C14 Cursor runs Composer 2.5 as its frontier coding model and Fusion behind Tab autocomplete. Source C14. | We are carrying the acquisition as FLAG. It is reported through technology press rather than confirmed in a filing or a statement from either party in the sources we reviewed. Treat vendor-continuity questions as open until it is. |
| NVIDIA | LG MOU signed 13 August at Santa Clara for a next-generation bipedal humanoid on Jetson Thor and Isaac GR00T.VERIFIED C09 GR00T N1.7 is in early access with commercial licensing; GR00T N2, based on DreamZero research, was previewed. | The foundation-model layer for robots is consolidating faster than the fleet layer. That is good for pilot speed and bad for supplier diversity, which is a concentration risk worth naming in a board paper. |
Cross-cutting read. Four of the six providers above shipped something whose primary purpose is to put an agent closer to an existing workflow: Cowork, Presence, Slackbot distribution, Grok Bot. None of the six shipped a control plane that a regulated buyer could point a supervisor at. That gap is the commercial opportunity and the operational risk in the same sentence.
Adoption is running ahead of the control plane, and the regulator has not landed yet
Banking and insurance lead nearly every adoption cut available, and sit closest to a supervisory expectation that is not final. That is an uncomfortable place to scale from.
The win
Fraud and financial crime remain the strongest first agent use case here, for structural rather than technical reasons. Fraud already has a labelled outcome, a human review queue, tolerance for false positives and an established second line. An agent dropped into that workflow inherits a control environment instead of needing one built.CITED C24
Aggregated write-ups this week put cost reductions across a named group of global banks in a 20 to 40% band with revenue uplift of 10 to 30%. We are not carrying those as fact: the band is wide, the methodology is not published, and no named institution has confirmed them.FLAG C22 The deployments are real. The precision is not. Source C22.
On the vendor side, Salesforce reported Agentforce annual recurring revenue of $800 million in Q4 FY2026, up 169% year on year, with accounts running Agentforce in production rising close to 50% sequentially. Q2 FY2027 results are scheduled for 26 August 2026.CITED C25 Read that as a monetisation signal rather than an outcome signal: it measures what buyers are paying for, not what the agents achieved.
The constraint
Gartner projects that by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance gaps found only after a production incident.CITED C21 The same body of work warns that applying uniform governance across all agents is itself a failure mode: a low-stakes retrieval agent and an agent with write access to a ledger cannot sit under the same control set without one being throttled and the other being under-controlled.
The mechanism showing up in incident data is unrevoked build-phase access. Agents get write permissions to production during development, and nobody removes the grant when the build ends.CITED C20 That is a privileged access management problem wearing new clothes, which is good news: the control already exists in every bank.
Three questions to put to your agent programme on Monday
- Which agents still hold write access granted during build? Ask for the list, with grant dates and last-review dates. If the list cannot be produced in an afternoon, that is the finding.
- What is the kill path, and who has walked it? Not whether a kill switch exists. Whether a named person has executed it in the last quarter, in production, with a timed result.
- Which model versions are our agents pinned to? Google retired two model families on Gemini Enterprise inside three weeks this month.VERIFIED C19 Model lifecycle is now an operational resilience question.
Architect's note: the FSB consultation is worth reading precisely because it is not final. An institution that designs against the June draft will be close enough to the October text to adjust, and will have spent the months building rather than waiting.
Adoption is broad, device classification is narrowing, and the evidence base is getting more honest
Around four in five hospitals now use AI in at least one clinical or operational function. The movement this year is not adoption. It is the FDA narrowing what counts as a device, and the outcome literature revising its own early numbers.
The classification shift
The FDA's January 2026 revision narrows which decision-support software counts as a regulated device. Software that informs a clinician who can independently review the basis for its recommendation can qualify as Non-Device CDS, with enforcement discretion signalled for some single-recommendation tools.CITED C27 Roughly 1,250 AI or ML-enabled devices had been cleared as of May 2025.CITED C26
The trap inside the relief
Falling outside device classification does not remove liability, and it does not remove the obligation to monitor. It removes the external forcing function that would otherwise have made someone build the monitoring. A tool that exits regulated status without an internal owner picking up post-deployment surveillance is strictly worse governed than it was the day before.
The evidence base is correcting itself, and that is a good sign
Ambient documentation is the most-deployed clinical AI category, and its outcome numbers have moved. A 2025 JAMA Network Open study across six US health systems found ambulatory clinician burnout fell from 51.9% to 38.8% after thirty days.CITED C29 Mass General Brigham reported a 21.2% reduction after 84 days.CITED C31
A larger 2026 study of roughly 1,800 clinicians across five academic medical centres found more modest time effects: about 16 minutes saved per eight hours of patient care.CITED C30 Both can be true. Burnout is a subjective load measure, documentation time is a clock measure, and the gap is the insight: the benefit shows up in cognitive load more than in minutes recovered.
What that means for a business case. A model built on recovered clinician minutes converting to appointment slots will underdeliver. Sixteen minutes per eight-hour session does not reliably become a billable slot. A case built on retention, after-hours record time and attention during the encounter is defensible against the 2026 data.
The control that is usually missing
Ambient tools fail in a documented way: they occasionally invent detail or soften hedging language a clinician used deliberately.CITED C29 The control is not "clinician signs the note", because signature rates approach 100% regardless of review depth. It is sampling: a fixed percentage of notes compared independently against the encounter, with hedging drift as an explicit criterion.
Architect's note: 82% of health systems piloting or deploying AI in at least one clinical workflow is a KLAS 2025 figure.CITED C28 We flag the vintage deliberately. A 2025 adoption number quoted in an August 2026 board paper without its date is the most common factual defect we see in this sector.
Humanoids cleared the pilot threshold in automotive, and nowhere else yet
The clearest applied robotics story this month is automotive assembly. It is also the narrowest. Treating it as a general manufacturing signal is the mistake to avoid.
What the robots are actually doing
The task profile is consistent: installing components in constrained spaces, quality inspection, materials handling. These resist fixed-arm automation because they need physical adaptability rather than repeatability.CITED C10 BMW has a Center of Competence for Physical AI in Production and is extending to Plant Leipzig from summer 2026.
Source C20. 65% of organisations reported at least one security incident in the past year caused by an agent operating on the corporate network. Modes overlap, so shares do not sum to 100.
Why automotive first, and what that tells other sectors
Automotive final assembly has four properties most regulated manufacturing does not: a well-surveyed physical environment, an existing safety case for humans and machines sharing a line, a takt time that can absorb a slower cycle during ramp, and a workforce already trained around robotic cells. A pharmaceutical fill-finish suite has none of those in combination, and cleanroom, washdown and validated-state change control each add a constraint no humanoid vendor addresses in the product today.
The practical filter. Before scoping a humanoid pilot outside automotive, ask whether the task sits inside a validated process. If it does, the robot is not the long pole, revalidation is. Budget change control first and hardware second, or the pilot stalls at the quality gate rather than the technical one.
Architect's note: cumulative unit counts circulating this month come from deployment trackers, not company filings, so we do not carry them as verified. Fleet claims here routinely mix produced, shipped and deployed units, which are three different numbers.
Power became the gating factor, and utilities are now in the AI supply chain
The constraint on AI build-out moved from silicon to interconnection. That reframes every energy and utility conversation from a sustainability topic into a capacity and reliability topic.
Source C08 and Source C36. Secondary write-ups round the Gartner figure to 27% and attach a 132 GW absolute. The 26% growth figure is the one carried in the Gartner release itself, so that is what we use.
The structural shift
An AI-optimised rack draws 30 kW to over 100 kW against 5 to 15 kW for a traditional rack.CITED C36 That is not an incremental load change, it is a different thermal and distribution problem inside the same building envelope. Gartner projects data centre electricity consumption growing 26% in 2026VERIFIED C08 and separately projects that power shortages will constrain 40% of AI data centres by 2027.CITED C35
The response, and where it is genuinely committed
Hyperscalers had committed roughly 9.8 GW of nuclear capacity across 13 deals as of July 2026.CITED C34 Two are worth separating from the rest because their status differs sharply:
Contracted and progressing
Microsoft holds a 20-year power purchase agreement with Constellation Energy for the full 835 MW output of Three Mile Island Unit 1. A FERC transmission waiver was approved on 1 June 2026, with power targeted for the second half of 2027.CITED C32 This is a restart of existing plant with a regulatory milestone already cleared, which is a materially different risk profile from new build.
Announced, long-dated
Meta announced agreements with TerraPower, Oklo and Vistra in January 2026 supporting up to 6.6 GW by 2035.CITED C33 The word doing the work is "up to", and the horizon is 2035. This is a development pipeline, not delivered capacity, and it should not be counted as firm power in any 2026 to 2029 load model.
What this means if you are the utility rather than the hyperscaler
- Interconnection queues are now a screening problem, not a processing problem. Speculative requests inflate queues and distort planning. The agentic use case with the clearest return here is queue triage against deliverability evidence, not customer service.
- Regional concentration is the real exposure. Aggregate national demand growth is manageable. Specific interconnection points are not, and that is where reliability risk and political risk both land.
- Load flexibility is the underpriced asset. A data centre that can curtail training workloads on notice is worth materially more to a system operator than one that cannot, and few contracts price that today.
Architect's note: if you are drafting an energy AI business case this quarter, separate contracted megawatts from announced megawatts on the first page. Roughly half the gigawatt figures circulating in AI energy coverage are development-stage commitments with horizons beyond 2030.
What the New York Fed actually found, published 5 August 2026
The most-cited AI labour claims run ahead of the data. The Federal Reserve Bank of New York published fresh survey work this month, and the finding is more specific and less dramatic than the coverage.
In the inaugural Street Level post on Liberty Street Economics, published 5 August 2026, New York Fed Research Director Kartik B. Athreya reported on the Bank's August regional business surveys of manufacturing and service firms. Businesses reported a notable increase in AI use over the past year, and very few firms reported AI-induced layoffs. AI was more likely to result in retraining than job loss. Among firms using AI, just over a third of service firms and 14% of manufacturing firms reported retraining workers in response.VERIFIED C07
Reading it against the other evidence
This is consistent with, not contradictory to, the entry-level findings that get more attention. Work on early-career employment finds a roughly 6% decline for entry-level workers in the most AI-exposed occupations between late 2022 and July 2025, against 6 to 9% growth for older workers in the same occupations.CITED C39 Aggregate employment shows no unemployment signal; the effect is concentrated at the entry of the pipeline, and it shows up as slowed hiring rather than displacement.
Those two findings fit together. Firms are not cutting incumbent staff. They are retraining them, and they are hiring fewer people into the roles those incumbents used to be recruited from. That is a slower and more damaging failure mode than layoffs, because it removes the training ground that produces the next cohort of senior operators without producing a headline event that forces a response.
The operational implication for regulated sectors
Model risk, clinical governance, financial crime and process safety all depend on senior judgement that was historically built through years of junior work now increasingly handled by agents. If the junior rung is thinned without replacing the learning it provided, the shortage appears in five to eight years, in exactly the roles that sign off on agent behaviour.
What to do about it now
Treat agent deployment and capability building as one programme, not two. Where an agent takes a task that used to teach something, name what it taught and rebuild that learning explicitly through review work, exception handling and structured exposure. This is a workforce design decision that has to be made at deployment time, not discovered afterwards.
Architect's note: the retraining split in the Fed data is worth sitting with. Just over a third of service firms retrain, against 14% of manufacturing firms.VERIFIED C07 Manufacturing has the longer-tenured workforce and the higher physical-safety stakes, and it is retraining less than half as often. That gap is a risk indicator.
Controls, questions and answers you can use on Monday
Everything below is drawn from the evidence in this edition. Nothing requires a new platform, and all of it can be started with people already on payroll.
The seven-control agent baseline
This is the minimum set we would expect to find before signing off an agent that touches a regulated process. It is deliberately short. A long control set that nobody completes is worth less than a short one that is actually evidenced.
| Control | Evidence that it exists | Why, from this edition |
|---|---|---|
| Access expiry | Every agent credential has an expiry date and a named owner. Build-phase grants expire automatically. | Unrevoked build access is the mechanism behind the largest share of reported agent incidentsSource C20 |
| Tested kill path | A dated record of a named person stopping the agent in production, with elapsed time. | A third of executives report they could not immediately stop a rogue agentSource C20 |
| Tiered governance | Agents classified by blast radius, with different control sets per tier. Not one policy for all. | Uniform governance across agents is itself a documented failure modeSource C21 |
| Model-version register | Which model version each agent is pinned to, plus a re-validation owner and trigger. | Two model families retired on one enterprise platform inside three weeks this monthSource C19 |
| Disclosure map | Every user-facing surface where the system talks, generates or infers, with its disclosure text. | Article 50 is live now and applies horizontally, regardless of risk tierSource C01 |
| Output sampling | A fixed percentage of agent outputs independently reviewed against source, with a drift criterion. | Ambient tools invent detail and soften hedging language; signature rate does not detect itSource C29 |
| Rollback boundary | Documented ability to isolate one agent, pause its pipelines and reverse only its change set. | Around 30% of autonomous agent runs hit exceptions requiring recoverySource C20 |
Did you know
The transparency duty does not care whether your system is high-risk. Article 50 applies if the system interacts with people, generates synthetic content, or infers emotion or biometric category. A low-risk internal chatbot with an external-facing surface is in scope. A deferred high-risk classification does not exempt it.VERIFIED C01
The FSB draft points at AI monitoring AI. The June consultation accepts that continuous human review of individual agent decisions does not scale, and recommends supplementing human oversight with automated monitoring layers. If your target operating model assumes a human in every loop, it is already diverging from where the supervisory thinking is going.VERIFIED C06
Frequently asked, briefly answered
| Question | Answer |
|---|---|
| Do we have to stop EU AI Act high-risk work? | No. The obligations moved to 2 December 2027, they did not disappear. Use the runway to finish design and data lineage while harmonised standards settle, and sequence conformity work rather than front-loading it.Source C04 |
| Is the FSB report something we must comply with today? | No. It is a consultation report from 10 June 2026 with a final version due October 2026. Design against it; do not cite it as a binding requirement in a regulatory response.Source C06 |
| Should we pilot humanoid robots? | Only if the target task sits outside a validated process, in a surveyed environment, with an existing human-machine safety case. Otherwise revalidation cost exceeds the hardware case.Source C10 |
| What is the single highest-return control? | Access expiry on agent credentials. It addresses the most common incident mechanism, uses privileged access management you already own, and can be evidenced in an afternoon.Source C20 |
| How should we handle model retirements? | As an operational resilience event with a named owner, a re-validation test pack and a diarised date, the same way you would treat a core banking patch window.Source C19 |
Where Ariana Digital fits
AEGIS, the Agentic Enterprise Governance and Intelligence Standard, is the framework behind the control baseline on this page. Three engagement shapes: AEGIS Diagnostic to establish where agents actually run and what holds access today, AEGIS Build to stand up the tiered control set and evidence pack, and AEGIS Run to operate monitoring and re-validation as models and rules move. Read more at ariana.digital.
Every claim in this edition, traceable
Citation chips: VERIFIED means named, dated and confirmed against a primary or authoritative secondary source. CITED means named source, not independently re-verified. FLAG means contested, imprecise or awaiting confirmation, and is stated as such in the text.
- C01 European Commission, "Safer and more transparent AI", 2 August 2026. commission.europa.eu
- C02 Goodwin, "Not Delayed, Not Deferred: EU AI Act Transparency Obligations Are Now in Force", August 2026. goodwinlaw.com
- C03 Cooley, "EU AI Act: Transparency Obligations Take Effect 2 August 2026", 3 August 2026. cooley.com
- C04 DLA Piper, "Digital AI Omnibus: deferral of high-risk obligations". Reg (EU) 2026/1744, in force 27 July 2026. dlapiper.com
- C05 EU AI Act, "Transparency Rules: A Practical Guide to Article 50". artificialintelligenceact.eu
- C06 Financial Stability Board, "Sound Practices for Responsible Adoption of AI: Consultation report", 10 June 2026. fsb.org
- C07 Athreya, K., "AI's Impact on Labor and Hiring", Liberty Street Economics, NY Fed, 5 Aug 2026. libertystreeteconomics.newyorkfed.org
- C08 Gartner, "Data Center Electricity Consumption to Grow 26% in 2026", 10 June 2026. gartner.com
- C09 "LG to Unveil Its Next-Gen Humanoid Robot, Built on NVIDIA Isaac GR00T", PR Newswire, 14 Aug 2026. MOU 13 Aug 2026. roboticstomorrow.com
- C10 Humanoid Robot Deployment Report, July 2026: Figure AI at BMW Spartanburg. humanoidapplications.com
- C11 "AI Robotics August 2026", Unitree H1 Pro regional launches. skycrumbs.com
- C12 xAI Grok 4.6 launch and pricing. basenor.com and docs.x.ai
- C13 xAI release notes, grok-voice-think-fast-2.0 and routing update from 5 August 2026; Grok Bot. x.ai/news
- C14 Reported SpaceX agreement to acquire Anysphere. FLAG, technology press only. techfundingnews.com
- C15 OpenAI, "Introducing OpenAI Presence". openai.com
- C16 OpenAI release notes, Aug 2026: GPT-5.6, Responses API, Daybreak, GPT-5.6-Cyber. releasebot.io
- C17 Anthropic Cowork release, week of 10 August 2026. anthropic.com/news
- C18 Claude enterprise updates, Aug 2026: plugin security scanning, self-hosted runners, Chrome defaults. releasebot.io
- C19 Google Cloud, Gemini Enterprise release notes: 3.5 Flash removal 4 Aug 2026, Grok 4.1 shutdown 20 Aug 2026. docs.cloud.google.com
- C20 Kiteworks, "AI Agent Security Incidents Hit 65% of Firms in 2026". kiteworks.com
- C21 Gartner, "Uniform Governance Across AI Agents Will Lead to Failure", 26 May 2026. gartner.com
- C22 Aggregated bank outcome bands, 20-40% cost / 10-30% revenue. FLAG, methodology not published. neurons-lab.com
- C23 JPMorgan Chase production AI use cases, company reported. thecgaigroup.com
- C24 Commonwealth Bank fraud loss reduction, H1 FY2026, company reported. ciodive.com
- C25 Salesforce Q4 FY2026 results and Agentforce ARR; Q2 FY2027 earnings 26 Aug 2026. futurumgroup.com
- C26 FDA AI/ML-enabled medical device clearances, count as of May 2025. pmc.ncbi.nlm.nih.gov
- C27 FDA Jan 2026 revision to decision-support classification, Non-Device CDS. livecompliance.com
- C28 KLAS Research 2025, 82% of health systems piloting or deploying clinical AI. arjunjaggi.com
- C29 "Ambient AI Scribes to Reduce Administrative Burden and Burnout", JAMA Netw Open, 2025. pubmed.ncbi.nlm.nih.gov
- C30 2026 multi-centre study, ~1,800 clinicians, five academic centres, documentation time. ama-assn.org
- C31 Mass General Brigham burnout prevalence reduction after 84 days. aha.org
- C32 Microsoft and Constellation, Three Mile Island Unit 1 PPA; FERC waiver 1 June 2026. axis-intelligence.com
- C33 Meta nuclear agreements, TerraPower/Oklo/Vistra, Jan 2026, up to 6.6 GW by 2035. about.fb.com
- C34 Hyperscaler nuclear commitments, 9.8 GW across 13 deals as of July 2026. axis-intelligence.com
- C35 Power shortages projected to constrain 40% of AI data centres by 2027. enkiai.com
- C36 AI-optimised rack power density against traditional racks. enkiai.com
- C37 L&T Technology Services AgenticIQ announcement, 11 Aug 2026. aiagentstore.ai
- C38 Salesforce Slackbot general availability to Business+ and Enterprise+. n8nlab.io
- C39 Entry-level employment effects in AI-exposed occupations, 2022-2025. naceweb.org
- C40 Cognizant and Anthropic expanded enterprise partnership, 27 July 2026. anthropic.com/news