Download this edition as PDF Email verification · about 30 seconds

We'll email a 6-digit access code. Enter it to unlock the Daily Market Scan PDF.

Friday, September 11, 2026 · 2026-09-11

Weekly Digest · Frontier & Industry Intelligence

The capital moved to the jurisdiction layer. The deadline nobody is counting is 82 days out.

In the five business days from September 7 to September 11, the largest disclosed AI rounds did not go to a frontier model. They went to sovereignty and to regulated professional work. In the same week OpenAI turned its agent harness into public infrastructure and Anthropic published evidence that the skill floor for sustained cyber operations has collapsed. Underneath all of it, one EU obligation that survived the Digital Omnibus is 82 days away.

1. The week in one screen

This is the weekly digest for the five business days from Monday, September 7 through Friday, September 11, 2026. It is a slow week for model launches and a loud one for where money and liability are moving.

Three things happened that a regulated operator should file together. Capital concentrated on the deployment and jurisdiction layer rather than on new frontier pretraining. A major lab turned its internal agent harness into a public API, which quietly makes the audit primitive you have been hand-building cheap. And a second major lab published evidence that the skill and tooling floor for running a sustained, multi-victim cyber campaign has collapsed. Those are not three stories. They are one story about who can now act at scale inside your perimeter, and what evidence you will have when they do.

Weekly digest: the five days that set up your Monday. All dates America/New_York.
Date	Event	Why it lands on a regulated operator
Mon Sept 8	Mistral AI closed a €3.0 billion Series D led by Samsung Electronics at a post-money valuation above €21 billion, the largest equity round ever completed by a European technology company. VERIFIED C03	Jurisdiction became a fundable product category. If your EU entity has been told sovereign hosting is a niche, the market just priced it otherwise. VERIFIED C03
Tue Sept 9	Harvey raised $550 million at a $15.5 billion valuation, with annual recurring revenue the company reports above $400 million. CITED C04	The buyer is a law firm and an in-house legal team, which is to say a control function. Capital is flowing into the part of your org chart that signs off on risk. CITED C04
Wed Sept 10	OpenAI opened the Agents API in public beta, exposing the managed Codex harness as four objects: Agent, Environment, durable Session, and the event stream. VERIFIED C01	The durable session is an auditable artifact by construction. That is the first time a major vendor has shipped the evidence layer as a first-class object rather than as a logging afterthought. VERIFIED C01
Wed Sept 10	Anthropic published its September 2026 threat intelligence report covering activity disrupted between December 2025 and August 2026 across seven harm areas. VERIFIED C02	The report's central finding is that AI collapsed the labor and tooling gap that separated state-sponsored operations from individual operators. Your threat model's top tier now has a much wider base. VERIFIED C02
Fri Sept 11	Edition date. The EU AI Act Article 50(2) machine-readable marking obligation is 82 days away, on December 2, 2026. VERIFIED C08	The Digital Omnibus deferred the expensive high-risk obligations and left this one alone. Most programs read the headline and filed it. VERIFIED C08

Architect’s read

The pattern across all five days is that the frontier labs shipped runtime and the market funded jurisdiction. Neither of those is a model capability story. Both are procurement and control stories, which means both land on functions that in most regulated firms are still reviewing agentic systems under a framework written for scorecards.

2. Where the capital went

Disclosed AI capital events, September 7 to 11, 2026. Mistral figure converted at the rate reported in contemporaneous coverage. Source C03, C04.

Mistral’s round was co-led by the EQT-managed Scaleup Europe Fund and existing investor PSG Equity, with new money from Advent, funds managed by BlackRock, and the Grand Duchy of Luxembourg. The company says it operates across 20 countries with more than 125 enterprise customers including Airbus, ASML and HSBC, and chief executive Arthur Mensch has said Mistral is on track to pass $1 billion in annual recurring revenue before year end. That last figure is forward guidance from the company, not a reported result. VERIFIED C03

Harvey’s round was co-led by Diffusion and Lightspeed Venture Partners. The company reports that 80 percent of Am Law 100 firms and five Fortune 10 companies use its tools, and that annual recurring revenue has passed $400 million. Both are company-reported. Coverage of the valuation splits between $15.5 billion and $15.6 billion; we have used $15.5 billion as the more widely reported figure and flag the variance rather than smoothing it. FLAG Source C04

The useful contrast is what did not get funded this week. No new frontier pretraining round was disclosed in the same five days. That is one week of data and not a trend, but it is consistent with a market that has stopped paying for raw capability and started paying for the ability to deploy capability inside a regulated boundary. CITED C15

3. The frontier ledger, global

Equal editorial weight, not equal praise. Everything below is what each organization shipped, said, or closed, with the operator consequence attached.

United States

OpenAI

The Agents API entered public beta on September 10 with no fee beyond usage. Inside a session it can run code, edit files, search the web, apply skills, produce artifacts, and split work across subagents under a concurrency limit. Developers can use OpenAI-managed sandboxes or their own infrastructure, with integrations from Cloudflare, DigitalOcean and Oracle. VERIFIED C01

Operator consequence. The Session object is durable and streams events. If you have been arguing internally that agent audit trails are a custom build, that argument got weaker this week. The reciprocal risk is that the sandbox boundary is now a vendor-controlled boundary, and your data residency review needs to name which of the three deployment modes you are actually in.

Anthropic

The September 2026 threat intelligence report documents activity disrupted between December 2025 and August 2026 across cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation. Anthropic reports that a hacktivist using stolen API keys, unconnected financially motivated individuals, and a state espionage operator each sustained multi-victim campaigns that a year earlier would have required teams of skilled operators. VERIFIED C02

Operator consequence. Stolen API keys appear as an access vector in a frontier lab’s own incident set. Treat model-provider credentials as privileged infrastructure secrets with rotation, scoping and anomaly detection, not as application config. Most firms we see still hold them in the second category.

Google

The Gemini Enterprise Agent Platform, announced at Cloud Next ’26 as the evolution of Vertex AI, took September updates including CodeMender support for Gemini 3.6 Flash and 3.7 Flash, and Gemini Omni 1.1 Flash moving to public preview for video, image and text. Google has also committed a $750 million fund for partners building agents for business processes and industries. CITED C06

Operator consequence. Google is positioning the platform as infrastructure for running many agents across workflows rather than as a single assistant. The governance question that follows is inventory: a platform designed to make agent creation cheap will produce agents faster than your registration process can absorb them.

xAI, SpaceX and Cursor

SpaceX completed its all-stock acquisition of Anysphere, the maker of Cursor, on August 14, 2026 at a reported $60 billion, the largest acquisition of a venture-backed startup on record. Grok 4.6 shipped two days before the close with a stated focus on long-running agents. Grok Bot, a persistent agent surface with a bot roster, computer access, shared group chats and routines that start without a prompt, now ships through Cursor tiers. CITED C05

Operator consequence. Your coding assistant, your agent runtime and your frontier model may now be one vendor relationship. That is a concentration your third-party risk register probably still records as three separate entries. Reconcile it before your next supervisory exam, not after.

NVIDIA

NVIDIA’s physical AI stack now spans Cosmos world foundation models for synthetic world generation and action simulation, the GR00T robot foundation models for dexterous manipulation, the Blackwell-architecture Jetson T4000 module, and an open Physical AI Data Factory Blueprint for generating, augmenting and evaluating training data. Boston Dynamics, Caterpillar, Franka Robotics, LG Electronics and NEURA Robotics are named users of the stack. CITED C07

Operator consequence. Synthetic training data for physical systems is now a supported reference architecture. For a regulated manufacturer that raises a question your quality system has to answer: what is your evidence standard for a behavior that was learned in simulation and never observed on the line?

Europe

Mistral’s raise is the headline, but the structural change is consolidation. Aleph Alpha, which built its platform around explainability and German-jurisdiction hosting, completed a merger with Cohere on April 24, 2026. The European sovereign field is now fewer, larger vendors. For a European buyer that simplifies the shortlist and weakens the negotiating position at the same time. CITED C03

China

On independent leaderboards the Chinese open-weight field leads on capability per dollar. As of September 8, 2026, GLM-5.3 at maximum effort was the top open-weight entry on the Artificial Analysis Intelligence Index at 45, ahead of Kimi K3 at 44 and DeepSeek V4 Pro at 36. CITED C20

The regulatory movement matters more than the leaderboard. China’s humanoid robot national standard took effect on September 1, 2026, and the binding rules for intelligent agents issued jointly by CAC, NDRC and MIIT have been in force since July 15, 2026. If you buy robotic hardware or agent components from Chinese suppliers, a foreign regulator is now shaping the conformity evidence you will hand your own. CITED C14

First half 2026 humanoid shipment estimates. Tracker figures carry a wide band and are not audited; treat as a planning range. Source C13.

4. Regulated-industry read

One win, one constraint, one control for each sector. The control is the part you can start on Monday.

Financial services

Win. Capital is flowing to the control function. Harvey’s $550 million round at a $15.5 billion valuation is a bet that contract analysis, research and workflow automation inside legal and compliance is a durable enterprise category, not a pilot. CITED C04

Constraint. The governing model risk guidance changed in April 2026. SR 26-2, issued jointly by the Federal Reserve, OCC and FDIC, supersedes SR 11-7 and reaffirms model inventories, independent validation, ongoing monitoring and third-party model governance. It also states that generative and agentic AI are novel and rapidly evolving and are not within its scope, with the agencies signaling a forthcoming request for information. FINRA’s 2026 oversight report separately flags autonomous agents as presenting novel supervisory considerations. CITED C10 VERIFIED C11

Control. Create a provisional model tier. Put every agentic system into your model inventory now, under full SR 26-2 documentation discipline applied voluntarily, plus three artifacts the guidance does not yet ask for: a tool-authorization register, a delegation-chain log, and dated revocation test evidence. When the RFI becomes guidance, you will be answering it rather than starting it.

Healthcare

Win. Ambient clinical documentation remains the one category with reproducible operational numbers. Kaiser Permanente reports more than 15,700 physician hours saved across a year with an 84 percent positive physician experience score, and Houston Methodist reports a 40 percent reduction in documentation time. Both are health-system reported. CITED C19

Constraint. FDA is still building the pathway. The TEMPO pilot gives some developers of generative AI medical devices provisional market access so the agency can observe them in real-world settings, and FDA has sought public feedback on a regulatory approach for generative AI-enabled devices that explicitly considers foundation models and agentic systems. There is no settled route yet. CITED C12

Control. Segment your clinical AI portfolio by claim, not by technology. A scribe that drafts a note a clinician signs is a documentation tool. The moment a system orders, triages, or adjusts therapy, you are in device territory and your evidence burden changes shape. Write that boundary down yourself, with your own regulatory affairs lead, before a vendor writes it for you in a datasheet.

Manufacturing

Win. Humanoid and agent-mediated industrial deployment is real but narrow. Trackers place first half 2026 humanoid shipments at roughly 19,000 to 22,000 units, up about 272 percent year over year, with more than 70 percent going to industrial or commercial applications. CITED C13

Constraint. Roughly 93 to 97 percent of that volume comes from Chinese manufacturers, and China’s humanoid national standard took effect on September 1, 2026. Separately, the outcome ranges circulating for agentic maintenance and inspection, commonly 15 to 30 percent downtime reduction and 40 to 60 percent inspection labor reduction, are vendor-reported and should be treated as a planning band rather than a benchmark. FLAG Source C22

Control. Build a two-column conformity map for every robotic asset: what the unit was certified against at origin, and what your jurisdiction and your insurer require. Every row where the columns differ is an integration test you owe, not a procurement footnote. This takes an engineer about a day per asset class and it is the single cheapest thing on this page.

Energy

Win. Predictive grid work is the most operationally mature utility AI category. Duke Energy reports more than 1.5 million outages prevented through self-healing grid automation, and PG&E has deployed more than 630 AI-enabled cameras for wildfire detection. Both are company-reported. CITED C18

Constraint. On June 18, 2026 FERC issued show-cause orders under Federal Power Act Section 206 to all six US regional grid operators, covering how loads above 20 megawatts interconnect, with reliability and resource adequacy information due in roughly 30 days and revised tariffs or a justification in roughly 60 days. The orders reach more than 30 states. AI compute siting is now a tariff question. VERIFIED C09

Control. If you are siting AI compute, model the interconnection tariff as a live variable with three scenarios rather than a fixed line in the business case, and put a named owner on the RTO docket. If you are the utility, the agentic opportunity this cycle is interconnection study throughput, not customer chat.

5. The 82-day clock

EU AI Act runway measured from September 11, 2026. Dates reflect the Digital Omnibus as adopted. Source C08.

The sequence is settled. The European Parliament approved the Digital Omnibus on AI on June 16, 2026, the Council adopted it on June 29, it was signed on July 8 and entered into force on July 27. It moved the Annex III high-risk obligations to December 2, 2027 and the Annex I obligations to August 2, 2028. It did not touch Article 50. Deployer-facing transparency obligations have applied since August 2026, and the Article 50(2) machine-readable marking deadline for synthetic content remains December 2, 2026. VERIFIED C08

Practitioner tip

Article 50(2) is a pipeline problem, not a model problem. The work is an inventory, and the inventory is broader than most teams assume. Every surface where your systems emit synthetic text, image, audio or video to a person counts, and that includes agent-drafted customer emails, meeting summaries, call notes and generated imagery in marketing. Build the surface list first, decide per surface whether you mark, disclose, or stop, and only then scope engineering. The inventory takes about two weeks. The engineering does not.

The scenario worth planning against: the marking obligation lands on December 2, which is inside most Q4 change freezes. If your release calendar closes in mid-November, your real deadline is eleven weeks out, not twelve.

The governance frameworks will not carry you here. ISO/IEC 42001 provides management-system framing, the NIST AI Risk Management Framework provides risk-function structure, and the EU AI Act provides the binding requirement, but all three were finalized before agentic deployment scaled. None of them addresses tool authorization, delegation-chain integrity, prompt injection or emergent multi-agent behavior as first-class controls. You will be writing supplementary controls inside the management-system structure, and you should plan for that as build work rather than as a mapping exercise. CITED C21

Agent deployment compared with agent governance maturity, 2026 survey-reported shares. Source C15.

The gap in that chart is the whole commercial argument for doing this now. Roughly 31 percent of enterprises run at least one agent in production, rising to about 47 percent in banking and insurance, while about 21 percent report a mature governance model for autonomous agents. Analysts expect around 40 percent of enterprise applications to embed task-specific agents by the end of 2026, up from under 5 percent in 2025. These are survey figures with the usual self-selection caveats, but the direction is not in dispute. CITED C15

6. Implementation architecture

This is the reference control pattern we use for an agentic deployment inside a regulated boundary. It is five layers, and this week made three of them cheaper or more urgent.

Reference control architecture for agentic systems in a regulated environment.
Layer	Question it answers	Artifact you produce	Accountable signer
L1 Intent and entitlement	Who may invoke this agent, on whose authority, for which population?	Entitlement matrix mapped to existing access control; agent registered in the model inventory	Business process owner plus model risk
L2 Tool authorization	What can it touch, at what blast radius, and how is that revoked?	Tool-authorization register: one row per tool, scope, rate ceiling, data class, revocation path	Security architecture
L3 Durable session ledger	What actually happened, in order, reconstructable months later?	Immutable session and event log with delegation chain; retention aligned to your records schedule	Technology control owner
L4 Evidence and marking	Can you show a supervisor the basis of a decision, and is synthetic output marked?	Decision evidence pack plus an Article 50 surface inventory with a mark, disclose or stop decision per surface	Compliance and regulatory affairs
L5 Human checkpoint and kill switch	Who stops it, how fast, and when was that last tested?	Named checkpoint owners, documented stop procedure, dated revocation test evidence	Operational risk

What changed this week. L3 got cheaper: a durable session that streams events is now a vendor primitive rather than a custom build. VERIFIED C01 L2 got more urgent: a threat report from a frontier lab now shows stolen API keys sustaining multi-victim campaigns, which is an argument for scoped, revocable tool credentials rather than broad service accounts. VERIFIED C02 And L4 acquired a date: December 2, 2026. VERIFIED C08

Cause and effect worth naming

Cheaper agent infrastructure does not reduce your control burden. It moves it. When creating an agent takes an afternoon, the binding constraint stops being engineering capacity and becomes registration throughput. Firms that do not put a lightweight intake in front of agent creation this quarter will spend next year doing discovery on their own estate. We have watched that happen with RPA, with unmanaged spreadsheets, and with shadow SaaS. The shape is identical and the remediation cost is higher each time.

7. What we would do Monday

Start the Article 50 surface inventory. One spreadsheet, one owner, every place your systems emit synthetic content to a human. Two weeks of calendar time. This is the only item on this list with a fixed external deadline. VERIFIED C08
Reclassify model-provider credentials as privileged secrets. Rotation schedule, scope reduction, anomaly alerting. A frontier lab’s own incident data now shows stolen keys sustaining campaigns. VERIFIED C02
Open a provisional agentic tier in the model inventory. Do not wait for the interagency RFI to become guidance. Register what you already run. CITED C10
Reconcile vendor concentration. If your coding assistant, agent runtime and frontier model now resolve to one corporate parent, your third-party register should say so in one entry, not three. CITED C05
Run one revocation test and date it. Pick your highest-privilege agent, pull its tool authorization, and record how long it took and what broke. That single artifact answers more supervisory questions than any policy document you can write this month.

Where we would help

Ariana.Digital runs principal-led engagements for regulated operators: the AI Readiness Brief for a fast, evidence-based read on where your estate actually stands, and the Reference Architecture Studio when the answer is that you need the control layer built rather than assessed.

AI Readiness Brief · AI Governance · Reference Architecture Studio

8. FAQ and did-you-know

Did the Digital Omnibus delay everything in the EU AI Act?

No. It moved the Annex III high-risk obligations to December 2, 2027 and Annex I to August 2, 2028. Article 50 transparency was not amended. Deployer obligations have applied since August 2026 and the Article 50(2) machine-readable marking deadline is December 2, 2026. VERIFIED C08

Does SR 26-2 cover our agentic systems?

Not by its own terms. The April 2026 interagency guidance that supersedes SR 11-7 states that generative and agentic AI are novel and rapidly evolving and are outside its scope, with a request for information signaled. That is not an exemption, it is an absence. Supervisors will still ask how you govern the systems, and you will be answering from first principles. CITED C10

Did you know the operator floor collapsed before the model floor did?

The September threat report describes a hacktivist with stolen API keys, unconnected financially motivated individuals, and a state espionage operator each sustaining multi-victim campaigns that a year earlier needed teams. The capability that changed hands was not novel exploitation. It was orchestration, tooling and persistence. Defensive programs built around the assumption that sustained campaigns imply a well-resourced adversary need to be rewritten. VERIFIED C02

What is the workforce read?

The wage signal is clearer than the displacement signal. PwC’s AI Jobs Barometer puts the pay premium for AI-skilled workers at 62 percent over non-AI peers in equivalent roles, up from 56 percent a year earlier. On displacement the evidence is genuinely contested: Stanford work finds a 13 to 16 percent relative decline in entry-level hiring for 22 to 25 year olds, while economists at Google and at Apollo Global Management argue interest rates and a low-hire, low-fire labor market explain the pattern better. We report both and flag the dispute rather than picking a side. CITED C16 FLAG Source C17

For talent-side work, runs deep-domain AI talent supply and upskilling at myndQ.com, with employer tooling at hr.myndQ.ai and talent.myndQ.ai.

9. Sources and research base

Every figure in this edition maps to an identifier below. Chips mark confidence: VERIFIED means named, dated and publicly checkable; CITED means named source, not independently re-verified; FLAG means contested or imprecise and pending re-verification.

C01 OpenAI, “Introducing the Agents API,” September 10, 2026. Primary. https://openai.com/index/introducing-the-agents-api/
C02 Anthropic, “Countering misuse of AI: September 2026,” threat intelligence report, September 10, 2026. Primary. https://www.anthropic.com/threat-intelligence-report-september-2026
C03 Mistral AI Series D, September 8, 2026. Bloomberg, “Mistral AI Boosts Valuation to €21 Billion in Samsung-Led Round”; TechCrunch, “Mistral raises €3B as sovereign AI becomes big business.” Two independent outlets. https://www.bloomberg.com/news/articles/2026-09-08/mistral-ai-raises-at-21-billion-valuation-in-samsung-led-round and https://techcrunch.com/2026/09/08/mistral-raises-e3b-as-sovereign-ai-becomes-big-business/
C04 Harvey funding round, September 9, 2026. PYMNTS and Tech Startups. Valuation reported as $15.5B and $15.6B across outlets; revenue and customer figures are company-reported. https://www.pymnts.com/news/investment-tracker/2026/harvey-raises-550-million-dollars-bring-ai-law-firms
C05 SpaceX acquisition of Anysphere (Cursor), closed August 14, 2026; Grok 4.6 and Grok Bot. DevOps.com and Techzine. https://devops.com/spacex-to-acquire-ai-coding-leader-cursor-in-60-billion-blockbuster-deal/
C06 Google Cloud, Gemini Enterprise Agent Platform, Cloud Next ’26 announcement and September 2026 updates. https://cloud.google.com/blog/topics/google-cloud-next/welcome-to-google-cloud-next26
C07 NVIDIA Newsroom, “NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots.” Primary. https://nvidianews.nvidia.com/news/nvidia-releases-new-physical-ai-models-as-global-partners-unveil-next-generation-robots
C08 EU AI Act Digital Omnibus and Article 50 transparency obligations. European Commission digital strategy pages and Gibson Dunn client alert on the Omnibus agreement. https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act and https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/
C09 FERC show-cause orders to six regional grid operators, June 18, 2026. White & Case insight alert and Utility Dive. https://www.whitecase.com/insight-alert/ferc-orders-grid-operators-promptly-revise-or-justify-interconnection-rules-data and https://www.utilitydive.com/news/ferc-doe-data-center-interconnection/823360/
C10 SR 26-2 interagency model risk management guidance, April 2026, superseding SR 11-7; scope statement on generative and agentic AI and signaled request for information. Consumer Finance Insights summary of OCC signaling. https://www.consumerfinanceinsights.com/2026/05/19/4745/
C11 FINRA, 2026 Annual Regulatory Oversight Report, generative AI section covering governance, recordkeeping and autonomous agents. Primary. https://www.finra.org/rules-guidance/guidance/reports/2026-finra-annual-regulatory-oversight-report/gen-ai
C12 FDA TEMPO pilot for generative AI-enabled medical devices, STAT, September 3, 2026; FDA press announcement seeking public feedback on a regulatory approach for generative AI-enabled devices. https://www.fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices and https://www.statnews.com/2026/09/03/tempo-fda-pilor-generative-ai-medical-device-regulation/
C13 First half 2026 humanoid shipment estimates and origin share. Market-tracker ranges, not audited. KraneShares and Technology.org. https://kraneshares.com/humanoid-robotics-in-2026-the-race-from-pilot-to-platform/
C14 China humanoid robot national standard effective September 1, 2026; CAC, NDRC and MIIT intelligent agent rules in force since July 15, 2026. 36Kr and Ariana.Digital regulatory ledger. https://eu.36kr.com/en/p/3967071657465732
C15 Enterprise agent deployment and governance maturity survey figures, 2026. Deloitte State of AI in the Enterprise and Gartner forecast as summarized by McKinsey QuantumBlack State of AI. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai and https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html
C16 PwC AI Jobs Barometer, wage premium for AI-skilled workers. https://www.pwc.com/gx/en/1/services/ai/ai-jobs-barometer.html
C17 Contested evidence on entry-level hiring effects. Federal Reserve Bank of New York Liberty Street Economics, “Do Job Postings Show Early Labor-Market Effects of AI?”; Goldman Sachs on AI and the US labor market. Contested and pending re-verification. https://libertystreeteconomics.newyorkfed.org/2026/05/do-job-postings-show-early-labor-market-effects-of-ai/ and https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-us-labor-market
C18 Utility AI deployment results, company-reported. EY power and utilities analysis on AI outage prediction; Overstory ignition and outage models, April 2026. https://www.ey.com/en_us/insights/power-utilities/ai-can-help-utilities-predict-grid-outages
C19 Ambient clinical documentation outcomes, health-system reported. Becker’s Hospital Review, health systems using AI. https://www.beckershospitalreview.com/healthcare-information-technology/ai/health-systems-using-ai-50-examples/
C20 Open-weight model standings as of September 8, 2026, Artificial Analysis Intelligence Index as reported by Morph and Turing Post. https://www.morphllm.com/best-open-source-coding-model-2026
C21 Gap analysis of ISO/IEC 42001 and the NIST AI Risk Management Framework against agentic system risks. Cloud Security Alliance research and Help Net Security. https://www.helpnetsecurity.com/2026/06/12/nist-iso-frameworks-govern-ai-agents/
C22 Vendor-reported outcome ranges for agentic maintenance and inspection in manufacturing. Planning band only, not a benchmark. Customertimes AI Agents in Manufacturing 2026 report. https://www.customertimes.com/ai-agents-in-manufacturing

Method and correction policy

Every edition is researched fresh against sources published within the preceding seven days where the item is time-sensitive. Figures carry a chip: VERIFIED means named, dated and publicly checkable; CITED means named source, not independently re-verified; FLAG means contested and pending re-verification. Where market commentary conflicted with primary legal sources this week, notably on EU high-risk applicability, we followed the primary legal sources and said so.