Download this edition as PDF
We'll email a 6-digit access code. Enter it to unlock the Daily Market Scan PDF.
The week capability stopped being the pitch
Frontier & Industry Intelligence : Regulated Sectors - Financial Services, Healthcare, Energy, Manufacturing. Week in review, 10 to 14 August 2026, including OpenAI Ultrafast, the Cursor–SpaceX close, and a four-sector read on where agent and robot deployment actually stands.
Seven moves that mattered, 10 to 14 August 2026
The control-plane story still holds: who controls execution, where data sits, and which clock is running. Thursday and Friday added two facts the first draft missed. OpenAI shipped Ultrafast, and SpaceX closed the Cursor acquisition. Grok 4.6 also shipped Wednesday, inside Cursor.
1. Article 50 enforcement moved from theory to practice
Transparency duties under the EU AI Act became applicable and enforceable by national authorities on 2 August 2026, and law firm guidance published across the last two weeks is consistent that the Digital Omnibus did not touch them. Annex III high-risk duties moved to 2 December 2027 and Annex I duties to August 2028. Generative systems already in market have until 2 December 2026 to meet the machine-readable marking requirement.VERIFIED C01
2. Google crossed a billion, and retired three things
Sundar Pichai said on 11 August that the Gemini app passed one billion monthly active users. In the same window the Gemini Enterprise platform removed Gemini 3.5 Flash from the global region on 4 August, set image generation model shutdown for 17 August, and set the Grok 4.1 family on the agent platform for shutdown on 20 August.VERIFIED C02
3. NVIDIA and LG signed on robotics
LG Corp and NVIDIA signed a memorandum of understanding on 13 August at NVIDIA headquarters in Santa Clara. LG confirmed on 14 August it is building a new bipedal humanoid on Isaac GR00T with Jetson Thor onboard compute and Halos for Robotics, with a public unveiling planned for the first quarter of next year.VERIFIED C03
4. Anthropic pushed governance into the runtime
August releases added beta security scanning of third-party skills and plugins on Enterprise plans, public beta self-hosted environments for Claude Code that run inside the customer network, Compliance API coverage extended across Cowork and Claude Code, and Claude in Chrome defaulted off on Enterprise with admin domain allowlists.VERIFIED C04
5. Two agent APIs are being retired this month
OpenAI deprecated Atlas on 9 August as browser agent capability folded into ChatGPT and Codex, and the Assistants API retires on 26 August 2026 at both OpenAI and Microsoft into two successors that are not interchangeable. Anyone with a validated workflow pinned to that API has twelve days.CITED C05
6. OpenAI shipped Ultrafast on Cerebras
On 13 August OpenAI opened a limited API preview of Ultrafast, a new service tier that runs GPT-5.6 Sol at up to 750 output tokens per second, up to 14 times Standard processing, on Cerebras wafer-scale chips. No quality cut, no smaller model. Access is waitlist-gated while capacity grows. Aimed at incident response, financial research, live support and coding.VERIFIED C16
7. SpaceX closed the Cursor acquisition
The merger became effective on 14 August. Cursor, Anysphere Inc., is now a wholly owned SpaceX subsidiary in a $60 billion all-stock deal announced 16 June. Cursor said this completes a process that started in April with SpaceXAI. Grok 4.6 shipped Wednesday inside Cursor, matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index. The day before close, Cursor published AIUC-1 certification for agent security.VERIFIED C17
Architect's note: five of the seven items above are still lifecycle or control-plane events. The last two are a speed tier and a vendor-map change. If your AI programme has a model change calendar but no model retirement calendar, this week is the argument for building one. It is also the argument for a coding-agent concentration review. Cursor is now a SpaceX subsidiary, and that is a third-party risk fact, not a product fact.
Equal-weight read on the providers regulated buyers actually shortlist
Same question asked of each: what shipped or was scheduled in the last two weeks, and what does it change for a bank, payer, plant or utility that has to evidence its decisions to a regulator.
| Provider | What moved this month | What it changes for a regulated buyer |
|---|---|---|
| Anthropic | Self-hosted Claude Code environments in public beta for Team and Enterprise, running inside the customer network alongside internal services and security controls. Skill and plugin security scanning in beta on Enterprise. Compliance API extended across Cowork and Claude Code. Opus 5 is now the default Opus model on Max, Team Premium, Enterprise pay-as-you-go and the API.VERIFIED C04 | Self-hosted execution removes the data residency objection that stalls most bank and payer pilots at legal review. Plugin scanning is the first vendor-side answer to supply chain risk in agent tooling, which is where your third-party risk team will look next. |
| OpenAI | Ultrafast opened 13 August as a limited API preview: GPT-5.6 Sol at up to 750 output tokens per second on Cerebras, up to 14 times Standard, no quality cut. Presence, announced 22 July, still packages policies, connections, evaluations and guardrails with OpenAI deployment engineers. Atlas deprecated 9 August; Assistants API retires 26 August.VERIFIED C16 | Speed is not a new model, it is a new serving path and a new silicon dependency. Treat Ultrafast as a change-controlled tier, not a toggle: pin the service tier, record that inference now sits on Cerebras, and do not put a waitlisted preview on a validated workflow. The retirement clock is unchanged. |
| Gemini 3.7 Flash launched 13 August, three weeks after 3.6 Flash, as the workhorse for coding and agents, at an introductory $0.75 / $3.75 per million input / output tokens through 31 December. Gemini app past one billion monthly active users as of 11 August. Three retirements still scheduled inside three weeks.VERIFIED C18 | Consumer scale is not the regulated story. The retirement cadence is, and so is the Flash release tempo. A platform that ships a new workhorse every three weeks needs a contractual notice period and a pinned-version path, or your validation evidence expires faster than your audit cycle. | |
| NVIDIA | LG memorandum of understanding signed 13 August, humanoid built on Isaac GR00T, Jetson Thor and Halos for Robotics, unveiling planned for Q1 next year. Isaac GR00T reference humanoid announced 1 June at GTC Taipei on a Unitree H2 Plus chassis with Sharpa Wave tactile hands.VERIFIED C03 | The humanoid stack is consolidating on one foundation model and one compute module. That reduces integration risk and concentrates supplier risk. Ask your robotics vendor which layers they own and which they rent. |
| SpaceXAI / Cursor | SpaceX completed the $60 billion all-stock acquisition of Cursor on 14 August. Cursor is now a wholly owned subsidiary. Grok 4.6 released 12 August, available in Cursor, Grok Build, the SpaceXAI API, OpenRouter, Vercel and Cloudflare, matching GPT-5.6 Sol on Artificial Analysis. Cursor published AIUC-1 agent-security certification on 13 August.VERIFIED C17 | If Cursor sits in a regulated engineering environment, Friday was a successor-entity event. Open a third-party risk file: who the contracting party is now, where code and prompts go, which model path is defaulting to Grok, and whether AIUC-1 is in-scope evidence or marketing. Do not treat close as a product upgrade. |
The pattern, stated plainly
Across the original four, the shipped work this month still sits in the control plane: where the model runs, who may invoke it, what gets logged, what gets scanned, and when it disappears. Thursday and Friday added a speed tier and a vendor-map change. Neither cancels the control-plane read. Both belong on the same ledger, because Ultrafast is a serving-path decision and Cursor is now a SpaceX subsidiary. For a regulated buyer the useful part is still the part you can put in a policy, and the part an examiner will ask to see.
The agent reached the core ledger, and the control gap showed up with it
Banking and insurance lead every adoption cut we can find. They also carry the highest cost of an unlogged autonomous action. Both things are true at once.
What is actually running
Fiserv agentOS, positioned as an operating system for agentic AI in banking, is on track for broad availability this month, co-developed with six financial institutions, two of which have agents in beta. CITED C09 That is the meaningful shape: agents arriving through the core processor rather than as a side application, which changes who owns the control evidence. If the agent runs inside your core platform, your model risk documentation depends on a vendor's logs.
The gap nobody is pricing
Only 21% of organisations report a mature governance model for autonomous agents, while 74% expect at least moderate agent use by 2027.CITED C07 In financial services that spread is not a maturity curve, it is an examination finding waiting to be written.
Gartner projects that more than 40% of agentic AI projects will be cancelled by 2027, citing escalating cost, unclear return and weak risk controls.CITED C07 The cancellations we have seen up close were not capability failures. They were evidence failures, discovered at second-line review after the build was done.
Three questions to put to your agent programme on Monday
- Can you produce a per-action ledger? For any agent action in the last 90 days: which model version, which prompt, which data classification, which human approved, which policy applied.
- What is the reversal path? Not the kill switch. The specific procedure to unwind a completed agent action on a customer account, and the maximum time it takes.
- Who signs the model change? When your provider deprecates a model with thirty days notice, name the person who authorises the revalidated version into production.
Architect's note: fraud and financial crime remain the strongest first agent use case in this sector, because the outcome is measurable, the reversal path is well established, and the control environment already exists. Start where the audit trail is already a habit, then extend the same ledger to servicing and credit. The reverse order is how programmes end up in the 40%.
Adoption is broad, classification is narrow, and the two are drifting apart
Around 80% of hospitals now use AI in at least one clinical or operational function, yet only 18% of healthcare organisations report an AI agent in production. That distance is where the risk sits.
What the regulator has done
The FDA had cleared or approved roughly 1,250 AI or machine-learning enabled medical devices by May 2025. A January 2026 revision narrowed which decision support software, expressly including AI-enabled clinical decision support, counts as a regulated device.CITED C10
ARPA-H is soliciting proposals for two agentic clinical care assistants, with teams to be selected within six months of the June 2026 solicitation.CITED C10
Where the drift happens
A narrower device definition does not mean lower risk. It means a larger share of clinical AI now sits outside device review and inside your own quality system, your own validation, your own liability.
Meanwhile EU Annex I duties, which cover regulated products including medical devices, moved to August 2028. Longer runway, same eventual evidence burden, and two more years of installed base to retrofit.VERIFIED C01
The deployments with published outcomes
Sepsis early warning remains the highest-acuity clinical decision support in wide use, with Epic's deterioration index deployed across hundreds of health systems and mortality reduction documented in ICU settings. Readmission risk prediction has shown 20 to 30 percent reductions in 30-day readmission in documented deployments.CITED C10 Both are prediction, not action. The agentic step, where the system orders, schedules or messages without a human in the loop, is where 2026 procurement is stalling.
The practical control that unblocks payer and provider pilots
Separate recommendation from action at the architecture level, not the policy level. An agent that drafts and queues is a documentation tool. An agent that commits is a clinical device conversation. Most stalled pilots we review are stalled because one service does both, so the whole thing inherits the stricter classification. Splitting the service is usually two sprints and it moves the entire programme out of a category it did not need to be in.
Evidence pack a health system should be able to produce today
- Intended use statement per model, in clinical language, signed by a named clinical owner.
- Population drift monitoring with a defined trigger and a named person who acts on it.
- Override log: how often clinicians disagreed with the model, and what happened next.
- Disclosure record: where and how patients are told AI was involved, mapped to state law and to Article 50 for any EU footprint.
Past the pilot threshold, well short of autonomy
The honest read on humanoids in August 2026: real units, real hours, narrow tasks, supervising technicians nearby, and edge cases deliberately designed out of scope.
| Deployment | Status as of August 2026 | What it tells you |
|---|---|---|
| Figure at BMW Spartanburg | 40 Figure 03 units in place following an 11-month Figure 02 trial. BotQ in California producing roughly one Figure 03 per hour as of May 2026, with Figure 03 past 1,000 cumulative units.CITED C11 | Eleven months of trial before a 40-unit commitment. Treat any vendor promising a 90-day path to production line work as unpriced risk. |
| BMW AEON, Europe | First humanoid deployment in Europe, targeting full production on high-voltage battery assembly by the end of 2026.CITED C11 | Battery assembly is the tell. High consequence, highly repeatable, tightly fixtured. That is the shape of a task humanoids win in 2026. |
| AgiBot | Approximately 15,000 cumulative units, the largest reported fleet number in the segment.CITED C11 | Volume is concentrating outside the US. Supply chain and parts availability belong in your risk register, not just your capex model. |
| LG on Isaac GR00T | Memorandum of understanding with NVIDIA signed 13 August, bipedal humanoid on Isaac GR00T and Jetson Thor, unveiling planned Q1 next year.VERIFIED C03 | A major appliance and electronics manufacturer entering on the NVIDIA stack accelerates standardisation. Good for integration cost, concentrating for supplier dependency. |
The constraint that decides your business case
Reporting across the segment is consistent that these are not fully autonomous operations. Supervising technicians remain nearby and task assignment is scoped to avoid complex edge cases.CITED C11 So the unit economics are not robot cost divided by displaced labour hours. They are robot cost plus supervision cost plus fixturing cost plus the engineering to keep the task inside its envelope. Programmes that model the first equation and discover the second one at month four are the ones that get paused.
Where the safety case actually lives
Functional safety for a fixed-cell robot is a solved discipline. A mobile humanoid that shares floor space with people, changes task by instruction, and is driven by a foundation model whose behaviour is not fully enumerable is a different assessment. The practical answer we see working is envelope control: constrain the physical zone, constrain the task list, and log every instruction that changed either. That gives your safety engineer a bounded system to certify and gives your AI governance function a change record it can read.
The interconnection queue is mostly fiction, and that is this week's most useful number
Bloomberg reported on 12 August that more than two thirds of the electricity being sought for the AI build-out is unlikely to materialise. Utilities and grid operators are expected to commit to roughly 28% of the 1,066 GW requested for data centre projects.
Read it as a planning instrument, not a headline. Phantom and long-shot requests inflate the queue because a developer can apply in many places for one project at no real cost. If you are a utility, the actionable consequence is that queue position is a poor demand signal and you need a filter. If you are an industrial or data centre buyer, the consequence is that queue length is a poor scarcity signal and you may be paying a premium against a shortage that will not arrive in your region.VERIFIED C12
The reliability event that changed the conversation
NERC issued a rare Level 3 alert in May 2026 after large data centre loads dropped offline simultaneously, with documented incidents of demand blocks above 1,000 MW disconnecting within seconds in response to minor grid disturbances. Corrective action was directed at utilities with key fixes targeted for this month.CITED C13
What it means operationally
A load that disconnects faster than generation can respond is not a customer, it is a contingency. Expect ride-through requirements, interconnection agreements with performance clauses, and load-side obligations to appear in new contracts. If you are negotiating a data centre power agreement this quarter, assume ride-through language is now in scope and price it.
Architect's note: the interesting arbitrage this quarter is not generation, it is flexibility. A workload that can be time-shifted or geographically shifted is worth materially more to a constrained utility than one that cannot. Very few AI platform teams can currently answer which of their jobs are interruptible. That inventory is a two-week exercise and it becomes a negotiating asset.
The adoption gap, measured
Same survey, four sectors, one uncomfortable spread. Production agent deployment is three times more common in banking and insurance than in government, and the governance maturity number is low everywhere.
The labour read, without the drama
The Indeed Hiring Lab outlook survey fielded in July 2026 and published 5 August collected 120 economists and experts. Asked whether AI will be a net positive or negative for US employment over the next year, 52% said at least a mild drag, 35% said a net gain and 13% expected no effect. Nearly all expect productivity to rise over three years, but 70% expect only a modest to moderate boost and just 4% expect a large one.VERIFIED C14
That is a genuinely split expert view, and it is worth quoting accurately when your executive team asks. The consensus is not that AI eliminates work at scale in the next year. The consensus is that productivity gains will be real and unspectacular, and that white-collar task composition shifts faster than headcount does. Workforce plans built on either extreme will be wrong.
The Four Gates: a control pattern that survives all four sectors
Built from what we see pass second-line review and what we see fail it. Four gates, each with one artifact. If an agent cannot clear a gate, it does not pass it, it queues at it.
| Gate | The question it answers | The single artifact | Failure mode it prevents |
|---|---|---|---|
| 1. Identity | Which agent, which model version, acting on whose authority? | Agent registry entry with a named human owner and an expiry date. | Orphaned agents running on deprecated models after the sponsor left. |
| 2. Boundary | What is this agent permitted to touch, and what is it structurally unable to touch? | Scoped credential plus a written negative list. | Scope creep by integration. The most common cause of an unplanned incident review. |
| 3. Disclosure | Does the person on the other side know, and can you prove they were told? | Disclosure record at the output boundary, tagged by jurisdiction. | Article 50 exposure, state disclosure law exposure, and the trust failure that precedes both. |
| 4. Reversal | How is a completed action undone, by whom, and in what maximum time? | Reversal runbook with a tested time-to-unwind figure. | The board question you cannot answer. Kill switches stop future actions, not completed ones. |
Why this order
Most programmes start at disclosure because it is the visible regulatory ask, then discover at gate two that they cannot describe what the agent can reach. Identity and boundary are cheap and they make the other two writeable. A team that does gates one and two properly can usually produce gate three in a sprint. A team that skips them ends up writing disclosure language about a system it cannot characterise.
How to run it in the next 30 days
- Week 1. Inventory. Every agent, script and copilot with production credentials, including the ones a business unit stood up without telling you. Assume the shadow inventory is larger than the sanctioned one.
- Week 2. Assign an owner and an expiry date to each. Anything without an owner gets disabled, not documented.
- Week 3. Write the negative list per agent. What it must never touch is faster to agree than what it may touch, and it is the version a risk committee can approve.
- Week 4. Test one reversal end to end and record the elapsed time. That number, not the policy, is your actual control.
This maps to AEGIS, the Agentic Enterprise Governance and Intelligence Standard, but you do not need the full framework to start. Four gates and four artifacts will hold a first examination.
Five real questions, answered the way we answered them
"Our vendor says EU high-risk obligations are live now. Are they?"
No. Article 50 transparency duties became enforceable on 2 August 2026. Annex III high-risk duties, which is where credit scoring, insurance pricing and employment screening sit, moved to 2 December 2027. Annex I moved to August 2028. What is genuinely running is the marking duty for generative systems already in market, due 2 December 2026, which is 110 days from today.VERIFIED C01
"Colorado went away, so US state risk is lower, correct?"
It changed shape. SB 26-189, signed 14 May 2026, repealed the original act and replaced it with a disclosure and rights framework focused on automated decision-making technology, effective 1 January 2027, with Colorado adding targeted healthcare AI provisions separately. California's frontier model law has been in effect since 1 January 2026, including safety evaluation duties and a 72-hour critical incident report to the Attorney General. Illinois disclosure duties are also landing. The compliance surface got wider and shallower, not smaller.CITED C15
"Should we wait for the humanoid market to settle before piloting?"
The evidence says the trial period is the work. BMW ran eleven months with Figure 02 before committing to 40 Figure 03 units. If you start when the market settles, you start eleven months behind whoever is running a trial now. The lower-risk version is to trial the task characterisation, the fixturing and the safety envelope, which are vendor-independent and transfer to whichever platform you eventually buy.CITED C11
"How do we handle a model deprecation inside a validated workflow?"
Treat the deprecation notice as a change control trigger on day one, not day twenty. Three things need to already exist: a pinned version record, a regression set specific to your use case, and a named approver. This month gave three live examples across a single platform, so this is not hypothetical planning.VERIFIED C02
"Does SpaceX closing on Cursor change our coding-agent vendor file?"
Yes. It is a successor-entity event, not a feature release. Confirm the contracting party, data-processing terms, default model path now that Grok 4.6 ships inside Cursor, and whether AIUC-1 is evidence you will actually rely on. If Cursor holds production credentials or touches source that cannot leave your perimeter, this is a change-control trigger on Friday, not a wait-and-see.VERIFIED C17
"Where is the fastest defensible first agent in a regulated firm?"
Somewhere the audit trail already exists and the reversal path is already practised. In banking that is usually fraud and financial crime. In healthcare it is revenue cycle and prior authorisation drafting, not clinical action. In manufacturing it is quality inspection and maintenance triage. In energy it is outage and work order triage. In every case the pattern is the same: pick the workflow where being wrong is recoverable and already logged.
Research base for this edition
All items reviewed during the week of 10 to 14 August 2026. Citation IDs correspond to the chips used throughout this brief.
- C01 Goodwin, "Not Delayed, Not Deferred: EU AI Act Transparency Obligations Are Now in Force", Aug 2026. https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force
- C01 Gibson Dunn, "EU AI Act Omnibus Agreement, Postponed High-Risk Deadlines". https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/
- C01 Jones Walker, "Yes, August 2 Still Matters". https://www.joneswalker.com/en/insights/blogs/ai-law-blog/yes-august-2-still-matters-the-eu-approved-a-high-risk-ai-delay-but-most-trans.html
- C01 Morgan Lewis, "EU AI Act's Transparency Rules: What Went Into Effect on 2 August", Aug 2026. https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august
- C01 Cooley, "EU AI Act: Transparency Obligations Take Effect 2 August 2026", 3 Aug 2026. https://www.cooley.com/news/insight/2026/2026-08-03-eu-ai-act-transparency-obligations-take-effect-2-august-2026
- C02 Google Cloud, Gemini Enterprise release notes. https://docs.cloud.google.com/gemini/enterprise/docs/release-notes
- C02 Google Cloud, Gemini Enterprise Agent Platform release notes. https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes
- C03 TNGlobal, "LG to Unveil Its Next-Gen Humanoid Robot, Built on NVIDIA Isaac GR00T", 14 Aug 2026. https://technode.global/prnasia/lg-to-unveil-its-next-gen-humanoid-robot-built-on-nvidia-isaac-groot-2/
- C03 NVIDIA Newsroom, Isaac GR00T reference humanoid robot. https://nvidianews.nvidia.com/news/nvidia-open-humanoid-robot-reference-design
- C04 Anthropic release notes, August 2026, via Releasebot. https://releasebot.io/updates/anthropic
- C04 Anthropic Newsroom. https://www.anthropic.com/news
- C05 OpenAI release notes, August 2026, via Releasebot. https://releasebot.io/updates/openai
- C06 OpenAI, "Introducing OpenAI Presence". https://openai.com/index/introducing-openai-presence/
- C06 Databricks, "OpenAI and Databricks at DAIS 2026". https://www.databricks.com/blog/openai-and-databricks-dais-2026-making-enterprise-ai-real
- C16 Cerebras, "Accelerating GPT-5.6 Sol Ultrafast with OpenAI", 13 Aug 2026. https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai
- C16 Cerebras IR, "Cerebras Powers Ultrafast Mode for OpenAI's GPT-5.6 Sol", 13 Aug 2026. https://investors.cerebras.ai/news-releases/news-release-details/cerebras-powers-ultrafast-mode-openais-gpt-56-sol
- C16 TechCrunch, "OpenAI introduces Ultrafast", 13 Aug 2026. https://techcrunch.com/2026/08/13/openai-introduces-ultrafast-a-new-mode-that-makes-gpt-5-6-sol-work-at-14x-the-speed/
- C17 Cursor, "Cursor is now a part of SpaceX", 14 Aug 2026. https://cursor.com/blog/joining-spacex
- C17 SpaceX Form 8-K, completion of Cursor / Anysphere merger, 14 Aug 2026. https://www.sec.gov/Archives/edgar/data/1181412/000162828026056945/spcx-20260814.htm
- C17 The Verge, "Cursor is officially part of SpaceX", 14 Aug 2026. https://www.theverge.com/ai-artificial-intelligence/980274/cursor-is-officially-part-of-spacex
- C17 SpaceXAI, "Introducing Grok 4.6", 12 Aug 2026. https://x.ai/news/grok-4-6
- C17 Cursor, "Cursor earns AIUC-1 certification for agent security and reliability", 13 Aug 2026. https://cursor.com/blog/aiuc-1
- C18 Google, "Introducing Gemini 3.7 Flash", 13 Aug 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/
- C07 Deloitte State of AI in the Enterprise 2026, McKinsey and Gartner agentic AI figures as compiled. https://joget.com/ai-agent-adoption-in-2026-what-the-analysts-data-shows/
- C07 "89% of AI Agent Pilots Never Scale: Gartner's 2026 Data". https://www.beri.net/article/ai-agent-adoption-enterprise-2026-gartner-idc
- C08 Neurons Lab, "Agentic AI in Financial Services: A Research Roundup for 2026". https://neurons-lab.com/articles/agentic-ai-in-financial-services-2026/
- C09 Fiserv, "Fiserv Launches agentOS: The Operating System for Agentic AI in Banking". https://investors.fiserv.com/news-releases/news-release-details/fiserv-launches-agentos-operating-system-agentic-ai-banking
- C10 Live Compliance, "AI in Healthcare: The Regulatory Landscape". https://livecompliance.com/learn/ai-healthcare-regulations/
- C10 Fierce Healthcare, clinical AI agents and FDA approval timeline. https://www.fiercehealthcare.com/ai-and-machine-learning/trump-administration-creating-clinical-ai-agents-3-year-fda-approval
- C11 Solid Market Research, "Humanoid Robots Cross the Pilot Threshold". https://www.solidmarketresearch.com/post/humanoid-robots-cross-the-pilot-threshold-where-factory-deployment-actually-stands-in-2026
- C11 Humanoid Robot Deployments Tracker, milestones 2026. https://humanoidapplications.com/deployments/
- C12 Bloomberg, "Most Electricity Sought for AI Data Centers in US Will Never Materialize", 12 Aug 2026. https://www.bloomberg.com/news/articles/2026-08-12/most-electricity-sought-for-ai-data-centers-in-us-will-never-materialize
- C12 Environmental Protection, "Investors Warned Of Utility Risks From AI Data Center Growth", 4 Aug 2026. https://eponline.com/articles/2026/08/04/investors-warned-of-utility-risks-from-ai-data-center-growth.aspx
- C13 Belfer Center, "AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment". https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid
- C13 Utility Dive, "AI data center growth could force US utilities to rethink generation plans". https://www.utilitydive.com/news/ai-data-center-growth-utilities-generation-plans/825541/
- C14 Indeed Hiring Lab, "Economists Expect a Cooled Labor Market and an AI Reshuffling of White-Collar Work", 5 Aug 2026. https://hiringlab.indeed.com/2026/08/05/q2-labor-market-outlook-survey/
- C15 Seyfarth Shaw, "Artificial Intelligence Legal Roundup: Colorado Postpones Implementation". https://www.seyfarth.com/news-insights/artificial-intelligence-legal-roundup-colorado-postpones-implementation-of-ai-law-as-california-finalizes-new-employment-discrimination-regulations-and-illinois-disclosure-law-set-to-take-effect.html
- C15 Norton Rose Fulbright, "Colorado enacts revised AI law". https://www.nortonrosefulbright.com/en-us/knowledge/publications/18733d31/colorado-enacts-revised-ai-law
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.