Feasible is not deployed

1. What happened this week

This edition covers Thursday, September 24 through Thursday, October 1, 2026, America/New_York. Items outside that window are dated and labeled wherever they carry weight in the argument.

Three kinds of evidence landed inside seven days, and they point the same direction. First, capability measurement: Anthropic's robot exposure index separates what machines could do from what is economical to deploy VERIFIED Source C01, and the International Federation of Robotics (IFR) reported that full-size humanoids sold about 7,000 units in 2025 and remain in a pilot phase limited by safety standards, cost and unproven business cases VERIFIED Source C13. Second, product releases: four US frontier labs each shipped or described something with a staged-trust design, covered in section 2. Third, buyer sentiment: PwC's 22% figure and the utility survey show buyers asking for human approval and regulatory clarity before they widen autonomy.

That reading is not a prediction about adoption speed. Several of the signals are surveys with stated limits, and several of the product claims are company-reported. What the week does establish is where buyers and vendors are both placing the work: staged access, human approval on write actions, scoped credentials and records that survive an audit.

2. Frontier scoreboard: equal-weight reading

Each lab below gets the same treatment: what was released or published, what is established, and how it is labeled. Company-reported benchmarks are not independent results. A product release is not an adoption figure.

Two readings matter for a regulated buyer. First, each of the four US labs shipped a version of the same design choice inside ten days: access staged by trust, with a human or a gate in front of consequential actions. SpaceXAI describes its own support rollout that way Source C07, Google gates Argon by program Source C06, Anthropic caps and dual-approves in its government product Source C04, and OpenAI makes dots an admin-enabled beta Source C30. Second, list prices for new flagship models have converged: Google's Argon and SpaceXAI's Grok 4.7 both list $2 per million input tokens, with output at $10 and $6 respectively on the figures each company published Source C06, Source C08. Cheaper inference does not shrink the evidence requirement.

SpaceXAI also has two items that are reports, not announcements. Bloomberg reported on September 30 that SpaceXAI is weighing a four-tier unified Grok and X subscription, including a $100-a-month tier; the company has not announced it CITED Source C09. Azerbaijan's digital ministry signed a memorandum with SpaceXAI to explore Grok for government services; it covers pilots and demonstrations, not a deployment CITED Source C10.

3. The deployment gap, in numbers

Anthropic's authors used Claude to rate about 19,000 job tasks from the US Department of Labor's O*NET database against a four-tier robot exposure rubric, cross-referenced robot capabilities with web searches for commercial deployments, and compared estimated annual deployment costs with worker compensation VERIFIED Source C01. The headline numbers are measurements of tasks and working hours, not of jobs. The authors state the limit themselves: the scale assumes jobs are more at risk where today's robots already do them in some settings, and robots driven by AI could do work they cannot do today.

Set that beside the industrial base. The IFR reports that the global operational stock of industrial robots reached about 5 million units, up 9%, and that 2025 installations rose 11% to more than 600,000. China installed 354,000 units, 59% of the world total. The United States installed nearly 38,500 units, up 12%, and moved ahead of Japan into second place. IFR's own forecast, which is a forecast and not a result, is 655,000 installations in 2026 and 806,000 by 2029 CITED Source C12. In professional service robots the IFR reports 249,500 units, up 24%, and about 7,000 full-size humanoids, which it describes as a pilot-phase segment VERIFIED Source C13.

The mechanism is straightforward. A factory with a defined, repeated task and a cost case buys robots at scale. A humanoid or a general agent asked to work across tasks needs a safety case, a liability owner and a price that beats labor, and those arrive more slowly than a model release. That is the gap the numbers show.

4. Financial services

Win: two frontier labs shipped finance-specific products, and both ship with controls

Anthropic launched Claude for Financial Advisors on September 14 with connectors to Addepar, BlackRock, Charles Schwab, Envestnet, iCapital, Orion, SS&C Black Diamond and others. Its stated controls include audit logs (Enterprise plans are recommended for recordkeeping), compliance screening against the SEC Marketing Rule, and human approval on critical tasks, with every investment recommendation and client communication subject to advisor review VERIFIED Source C23. OpenAI launched ChatGPT for Financial Services on September 10 with Morgan Stanley and Evercore as design partners, 50+ connectors, SAML single sign-on, SCIM provisioning, role-based access, workspace-level information barriers and compliance log exports VERIFIED Source C24. Both are company-reported product launches. Neither is an outcome study.

On September 22, Bank of America, Capital One, ASB Bank, Commonwealth Bank of Australia, ING and NatWest published joint principles for agentic commerce: transparency, safety, privacy and data, choice, and interoperability, with an implementation paper to follow VERIFIED Source C21.

Constraint: identity, authorization, fraud and liability are still described as open

The consortium's own language names identity, authorization, fraud prevention, liability management and customer protection as what must be worked out VERIFIED Source C21. On September 28, Apollo chief economist Torsten Slok warned that agents optimizing cash balances could pull low-rate deposits out of banks; the cited contrast was an average checking rate of 0.1% against 3.3% to 5% at fintechs. That is an opinion and a scenario, not an observed event CITED Source C25. In Europe, ESMA announced that digital innovation, covering AI and tokenization, becomes a Union Strategic Supervisory Priority from 2027, starting with a mapping of firms' client-facing AI use CITED Source C22.

5. Healthcare

Win: a funded, named path to the first FDA-authorized agentic clinical system, as a pilot

ARPA-H announced ADVOCATE on September 9: $62.7 million over four years, with $33.7 million in year one, to build what it calls the first FDA-authorized clinical agentic AI system, a cardiovascular care agent. Participants are Atman Health, Tempus AI and Updoc for patient-facing agents, Stanford for supervisory AI, and Duke and Kaiser Permanente for implementation. The program targets an FDA authorization package within 24 months of award, and ARPA-H projects $28 billion in annual savings across the heart-failure population if it succeeds. That is a projection. ADVOCATE is a funded program, not a result VERIFIED Source C17.

Constraint: the regulatory path is still the 510(k) path, and the generative-AI path is in comment

In a Federal Register order dated September 17, the FDA denied a petition to exempt radiology computer-aided detection, diagnosis and triage software from 510(k) premarket notification. Manufacturers must continue to file VERIFIED Source C18. The agency's August 18 discussion paper on generative-AI-enabled medical devices is neither draft nor final guidance; comments close October 19, 2026, 18 days from today, under docket FDA-2026-N-7874 CITED Source C19. In California, bills signed on September 30 include AB 1979, which treats businesses offering health care chatbots as health care providers under the Confidentiality of Medical Information Act, and SB 503, which requires developers and deployers of clinical decision-support AI to identify foreseeable risks of biased impact. The source does not give effective dates CITED Source C20.

6. Manufacturing and robotics

Win: the installed base keeps growing, and the United States is now second

The IFR figures in section 3 are the most solid manufacturing evidence of the week: about 5 million industrial robots in operation, 2025 installations above 600,000, and the US ahead of Japan on 38,500 units CITED Source C12. IROS 2026, the robotics research conference, runs in Pittsburgh through Saturday, October 3, with more than 4,000 participants; this edition reports conference context only, and industrial results and awards are expected to publish after it closes CITED Source C35.

Constraint: humanoids are in pilots, and the safety case sits with the integrator

The IFR describes about 7,000 full-size humanoids as a pilot-phase segment limited by safety standards, costs and the lack of established business cases VERIFIED Source C13. In Europe, Machinery Regulation 2023/1230 applies from January 20, 2027, 111 days from today. According to an August 6 summary from the French agency Eurogip, AI performing safety functions in machinery falls mainly under that regulation, with a single conformity assessment, while technical AI requirements arrive by delegated act no later than August 2, 2028 CITED Source C29. Programs that fold the EU AI Act and the Machinery Regulation into one date are carrying a gap.

7. Energy and utilities

Win: most surveyed utilities say AI is deployed or being operationalized

In the National Grid Partners survey, 78% of 134 US utility innovation leaders said they are deploying or operationalizing AI applications to manage interconnection demand VERIFIED Source C14. The figure is self-reported and sponsor-commissioned, and it describes applications in use, not outcomes achieved.

Constraint: reliability pressure, regulatory frameworks and cost allocation

The same survey found 74% saying data-center load growth affects grid reliability, 87% citing regulatory and rate-case frameworks as limiting returns on innovation, and 83% saying infrastructure costs for AI data centers are passed to residential customers. The survey was commissioned by its sponsor and the figures are self-reported VERIFIED Source C14. On the standards side, a July 16 FERC order (docket RD26-7-000) directs NERC to file new or modified Reliability Standards for computational loads by December 31, 2026, 91 days from today, and a work-plan filing by March 1, 2027 CITED Source C15. Earlier, on May 12, NERC issued a Level 3 alert, its highest severity level, with seven Essential Actions and responses due August 3; the alert is not mandatory and carries no penalties CITED Source C16.

8. Implementation architecture: the gated agent path

The four industry cases above share one structure. An agent or robot policy should reach a consequential action only through a path that checks who it is, what it may touch, who approves, and where the record goes. Figure 2 shows the path; the table after it maps each case to its autonomy tier and the evidence artifact a regulator or auditor would ask for.

9. Work and talent: what the Fed and Sacramento said

Two Federal Reserve governors addressed AI and work within two days. On Monday, September 28, Governor Lisa Cook said unemployment and layoffs remain low and have been relatively flat over two years even as AI adoption picked up, but that there is evidence in some sectors, including coding jobs in software and simultaneous translation, that AI might be decreasing labor demand, and that many recent graduates are having more difficulty finding first jobs, possibly in part because of AI's ability to take on entry-level tasks. She also cited a small-business survey in which nearly half of small employer firms use AI and 71% of those report higher productivity VERIFIED Source C02. On Tuesday, September 29, Governor Michael Barr said "we are already seeing signs that AI may be a labor substitute in some sectors, especially for younger, less experienced workers," and that now is the time to begin considering how to address the effects while adoption is early. His speech also confirms that the FOMC voted unanimously earlier in September to raise short-term rates VERIFIED Source C03.

On Wednesday, September 30, Governor Newsom signed 11 more AI bills. According to the Transparency Coalition's summary, SB 951 requires covered employers to give 90 days' notice before technological changes affecting 25% or more of their workforce. The summary gives no effective dates and the bill text should be checked before it enters a compliance plan CITED Source C20.

The Bureau of Labor Statistics publishes the September Employment Situation report tomorrow, Friday, October 2, at 8:30 a.m. ET VERIFIED Source C28.

10. Platforms, other labs and policy on the record

Enterprise platforms. ServiceNow described, in an interview reported by SiliconANGLE on September 28 around Okta's Oktane conference, a five-step containment model of discover, observe, govern, secure and measure, with graduated containment in place of a binary kill switch. It is an interview-based statement, not a product release CITED Source C31. This edition's research found no Salesforce, Snowflake, Databricks, Adobe or Microsoft announcement dated September 24 through October 1 that met its verification bar, and none is reported here.

Executive order. A September 29 order titled "Inaugurating the Era of Super Intelligence" directs federal agencies to use "Super Intelligence" or "SI" in place of "Artificial Intelligence" and "AI" in non-statutory documents, and directs the Assistant to the President for Science and Technology to submit proposed statutory-definition language within 60 days, which falls on or about November 28. It states that it does not alter prior regulations and creates no enforceable right VERIFIED Source C26. For vendors selling into federal agencies, it changes document vocabulary. It does not change a statutory obligation today.

Reported investigation. Reuters, citing a senior FTC official speaking anonymously, reported on September 30 that the Federal Trade Commission is investigating Anthropic, OpenAI and METR over risks from rogue AI agents, with formal information demands and compelled executive testimony planned. The companies had not responded when the report was published, and no FTC announcement was found FLAG Source C27.

Labor-market reading. A Harvard Gazette report on September 29, on why AI has not yet triggered mass layoffs, quoted Elmendorf saying current evidence has "very little predictive power" for the next five to 15 years CITED Source C34.

11. The regulatory calendar, today to January 2027

12. Scenario planning

These are scenarios, not forecasts. Each names what would signal it and what a regulated organization can do now at low cost.

13. Self-check: five questions before an agent leaves pilot

This tool runs in your browser and stores nothing. It scores five yes-or-no questions drawn from the controls in sections 4 to 8. It is a conversation starter, not an assessment.

14. Did you know / FAQ

15. What we are watching, and what would change our view

Tomorrow, Friday, October 2: the BLS Employment Situation report Source C28. By October 19: the FDA docket Source C19. Through Saturday, October 3, and just after: IROS 2026 industrial and award coverage Source C35. Open items: whether SpaceXAI announces the reported subscription overhaul Source C09; when Gemini 4 Argon reaches broader access Source C06; and whether the reported FTC inquiry is confirmed by the agency Source C27.

Three results would change our view that deployment evidence, not capability, is the constraint. A regulated-industry deployment at scale with autonomy beyond Tier 2 and no gate in front of writes. A survey showing most buyers would approve fully autonomous agents in a high-consequence domain, against PwC's 22% Source C11. Or a robot-cost trend that moves the cost-competitive share well above Anthropic's 0.3% Source C01.

16. Source ledger

Every claim group used in this edition, with the sources behind it. Chip meanings: VERIFIED named, dated and re-read at the first-party or primary page during this run. CITED named source, not independently re-verified at the primary page. FLAG reported, single-sourced or pending confirmation; do not carry into a regulated-buyer document without checking. Figures published by a company about its own product are company-reported and labeled so in the body. Forecasts, pilots, memoranda and announced targets are not treated as completed facts.

1. C01 Anthropic, "What work can robots do?" (Robot Exposure Index), September 30, 2026. 74% of physical tasks technically feasible; 0.3% cost-competitive; 34% of working hours; 81% of employment exposed to robots or LLMs versus 50% for LLMs alone; 40-year statement; method and caveat. https://www.anthropic.com/research/what-work-can-robots-do
2. C02 Federal Reserve Governor Lisa D. Cook, "An Update on AI and the Economy," September 28, 2026, Oakland Tech Week keynote. https://www.federalreserve.gov/newsevents/speech/cook20260928a.htm
3. C03 Federal Reserve Governor Michael S. Barr, "Economic Conditions and Monetary Policy," September 29, 2026, Detroit Economic Club. https://www.federalreserve.gov/newsevents/speech/barr20260929a.htm
4. C04 Anthropic, "Claude for Government is now generally available," September 30, 2026. FedRAMP High, fixed-increment billing with hard not-to-exceed cap, two-person approval for sensitive Anthropic-side operations, early access items. Company-reported. https://claude.com/blog/claude-for-government-is-now-generally-available
5. C05 OpenAI, "Disrupting a coordinated model-distillation campaign," September 30, 2026. Dates, request counts, attribution caveat, actions taken. Company-reported. https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/
6. C06 Google, "Gemini 4 Argon: our next era of frontier intelligence," September 30, 2026. Fairwind Program access, pricing, benchmark table, pre-release process. Company-reported. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ · https://techcrunch.com/2026/09/30/google-releases-gemini-4-argon-called-its-most-powerful-model-yet/ · https://www.marktechpost.com/2026/09/30/google-deepmind-unveils-gemini-4-argon-with-1m-output-tokens-for-coding-knowledge-work-and-cyber-defense/
7. C07 SpaceXAI, "How SpaceXAI is using Grok Bot to scale customer support," September 22, 2026. Ticket, cost and rollout statements; Cursor acquisition dated August 14, 2026. Company-reported. https://x.ai/news/grok-bot-customer-support · https://thenextweb.com/news/spacex-cursor-acquisition-completed-gpu-fleet
8. C08 SpaceXAI, "Grok 4.7," September 21, 2026. Pricing; availability through Cursor, Grok Build and the API; invite-only red-team access for cybersecurity partners. Company-reported benchmarks not repeated. https://x.ai/news/grok-4-7
9. C09 Bloomberg, "SpaceXAI considers overhaul of pricing for Grok, X users," September 30, 2026, as reproduced by Yahoo Finance. Reported consideration, not an announcement. https://finance.yahoo.com/technology/ai/articles/spacexai-considers-four-tier-pricing-181516980.html
10. C10 OC Media, "Musk's SpaceXAI signs agreement for Azerbaijan government to trial Grok," late September 2026 (exact date not confirmed). Memorandum covering pilots and demonstrations. https://oc-media.org/musks-spacexai-signs-agreement-for-azerbaijan-government-to-trial-grok/
11. C11 SecurityWeek, "Enterprises struggle to prepare for AI and quantum threats, PwC says," October 1, 2026. Survey of nearly 4,000 leaders in 70+ countries; 22% figure. https://www.securityweek.com/enterprises-struggle-to-prepare-for-ai-and-quantum-threats-pwc-says/
12. C12 The Robot Report, "5 million robots now working in factories worldwide, IFR reports," reporting IFR World Robotics 2026 industrial statistics. Forecasts for 2026 and 2029 are IFR forecasts. https://www.therobotreport.com/5-million-robots-now-working-factories-worldwide-ifr-reports/
13. C13 International Federation of Robotics, World Robotics 2026 service robots release, September 30, 2026. 249,500 professional service robots; about 7,000 full-size humanoids; pilot-phase statement. The release's medical-robot figures are not used. https://ifr.org/ifr-press-releases/news/world-robotics-2026
14. C14 National Grid Partners, 2026 Utility Innovation Survey, released September 18, 2026. 134 US utility innovation leaders; fielded May 20 to July 7, 2026; sponsor-commissioned. https://www.prnewswire.com/news-releases/2026-utility-innovation-survey-industry-leaders-turning-more-to-ai-as-data-center-boom-reshapes-grid-planning-302883063.html
15. C15 FERC order of July 16, 2026, docket RD26-7-000, directing NERC filings by December 31, 2026 and March 1, 2027, as summarized by Willkie Farr and Gallagher. FERC's own text was not re-read. https://www.willkie.com/publications/2026/07/ferc-orders-new-reliability-standards-for-data-centers-and-other-computational-loads
16. C16 NERC Level 3 Alert, May 12, 2026 (seven Essential Actions; responses due August 3; not mandatory), as summarized by Morgan Lewis. NERC's own text was not re-read. https://www.morganlewis.com/blogs/powerandpipes/2026/05/nerc-alert-moves-data-centers-from-emerging-risks-to-planning-obligations
17. C17 ARPA-H, ADVOCATE program announcement, September 9, 2026. $62.7 million over four years; participants; 24-month FDA package target; $28 billion projection. https://arpa-h.gov/news-and-events/arpa-h-launches-worlds-first-bid-build-fda-authorized-clinical-ai-cardiovascular
18. C18 Federal Register, "Medical Devices; Exemption From Premarket Notification: Radiology Computer-Aided Detection and/or Diagnosis Devices and Computer-Aided Triage and Notification Devices," September 17, 2026, docket FDA-2025-P-5560. Petition denied. https://www.govinfo.gov/content/pkg/FR-2026-09-17/html/2026-19074.htm
19. C19 FDA discussion paper on generative-AI-enabled medical devices, August 18, 2026; comments close October 19, 2026, docket FDA-2026-N-7874. Read through law-firm and trade summaries; the FDA and Regulations.gov pages were not re-read. https://innolitics.com/articles/fda-generative-ai-medical-devices-discussion-paper/ · https://natlawreview.com/article/fda-seeks-public-feedback-regulatory-approach-generative-ai-enabled-medical-devices
20. C20 Transparency Coalition, "Gov. Newsom wraps California term by enacting 11 more laws on AI safety," September 30, 2026. SB 951, SB 503, AB 1979. No effective dates stated; advocacy-group summary, so bill text should be checked. https://www.transparencycoalition.ai/news/gov-newsom-wraps-california-term-by-enacting-11-more-laws-on-ai-safety
21. C21 Bank of America and five banks, joint principles for trusted agentic commerce, September 22, 2026. https://newsroom.bankofamerica.com/content/newsroom/press-releases/2026/09/global-banks-collaborate-on-principles-for-trusted-agentic-comme.html
22. C22 ESMA digital innovation supervisory priority from 2027, reported September 23, 2026 by FOW and Cointelegraph. The ESMA primary page was not re-read. https://www.fow.com/insights/esma-to-prioritise-ai-tokenisation-supervision-in-2027 · https://cointelegraph.com/news/esma-to-focus-eu-wide-supervision-on-ai-and-tokenization-from-2027
23. C23 Anthropic, "Claude for Financial Advisors," September 14, 2026. Connectors, controls, availability. Company-reported. https://claude.com/blog/claude-for-financial-advisors
24. C24 OpenAI, "Introducing ChatGPT for Financial Services," September 10, 2026. Design partners, connectors, access controls. Company-reported. https://openai.com/index/introducing-chatgpt-financial-services/
25. C25 CoinDesk, "AI agents could drain cheap bank deposits, Apollo's Torsten Slok warns," September 28, 2026. Opinion and scenario. https://www.coindesk.com/markets/2026/09/28/ai-agents-could-drain-cheap-bank-deposits-apollo-s-torsten-slok-warns
26. C26 The White House, "Inaugurating the Era of Super Intelligence," September 29, 2026; Freshfields analysis. https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the-era-of-super-intelligence/ · https://www.freshfields.com/en/our-thinking/blogs/a-fresh-take/trump-executive-order-mandates-shift-to-super-intelligence-102o403
27. C27 Reuters via BNN Bloomberg, "FTC opens probe into AI giants including Anthropic and OpenAI," September 30, 2026. Anonymous FTC official; no agency announcement or company response found. https://www.bnnbloomberg.ca/business/artificial-intelligence/2026/09/30/ftc-opens-probe-into-ai-giants-including-anthropic-and-openai/
28. C28 US Bureau of Labor Statistics, Employment Situation release schedule: September 2026 report on October 2, 2026 at 8:30 a.m. ET. https://www.bls.gov/schedule/news_release/empsit.htm
29. C29 Eurogip, "Machinery Regulation 2023/1230: what the Digital Omnibus on AI changes," August 6, 2026. Secondary summary; EUR-Lex text should be checked before reliance. https://eurogip.fr/en/machinery-regulation-2023-1230-what-the-digital-omnibus-on-ai-changes/
30. C30 OpenAI, "Introducing dots," September 29, 2026. Enterprise availability described as admin-enabled beta. Company-reported. https://openai.com/index/introducing-dots/
31. C31 SiliconANGLE, "ServiceNow ties agent containment to business context at Oktane," September 28, 2026. Interview-based. https://siliconangle.com/2026/09/28/servicenow-ties-agent-containment-business-context-oktane
32. C32 Meta, "Introducing Muse for small business," September 29, 2026. Company-reported. https://about.fb.com/news/2026/09/introducing-muse-small-business/
33. C33 The Star, reporting Bloomberg on DeepSeek's Huawei Ascend toolkit, September 30, 2026. Data-center and fundraising items are reports. https://www.thestar.com.my/tech/tech-news/2026/09/30/deepseek-unveils-huawei-ai-chip-tools-that-may-replace-nvidias
34. C34 Harvard Gazette, "Why AI hasn't triggered mass layoffs yet," September 29, 2026. https://news.harvard.edu/gazette/story/2026/09/why-ai-hasnt-triggered-mass-layoffs-yet/
35. C35 Carnegie Mellon University, IROS 2026 preview, September 21, 2026; conference September 27 to October 3, Pittsburgh. https://www.cmu.edu/news/stories/archives/2026/september/global-conference-in-pittsburgh-puts-carnegie-mellons-robotics-leadership-in-the-spotlight

Prepared by Ariana Digital LLC. Anthropic Claude Partner — Ariana Digital LLC. Coverage of frontier labs is based on each lab's published activity and does not reflect partner status. This edition is market intelligence, not legal, financial or clinical advice. Dates are America/New_York. Edition date: 2026-10-01. © Ariana Digital LLC. All rights reserved.