ARIANA.DIGITAL

Daily Market Scan · Edition 2026-09-23
Enterprise Agentic AI & Governance

1. The 60-second scan

What moved this week — Monday, September 21 through today, Wednesday, September 23, 2026 — and the tail of last week, September 14 through 18. Anthropic shipped Claude Opus 5.5 on Tuesday at a lower list price than Opus 5, and closed off the ability to disable thinking on that model entirely VERIFIED C01. OpenAI moved GPT‑6 Sol into general API availability on Tuesday at roughly half the prior flagship tier and published its priorities and principles for third-party safety assessments the same day, extending outside evaluation into the training phase rather than only pre-release VERIFIED C02 CITED C03, CITED C13. SpaceXAI released Grok 4.7, which it says runs at the same price and speed as Grok 4.6 on a larger base model CITED C11, CITED C12. Google's Gemini 3.8 Flash has been generally available since September 2 at an introductory $0.75 per million input tokens through December 31, 2026, and AlphaEvolve reached general availability on the Gemini Enterprise agent surface VERIFIED C04. Alphabet's Intrinsic open-sourced Intrinsic Core VERIFIED C05 CITED C06. On the policy clock: the President addressed the General Assembly on Tuesday, September 22, rejecting what he called a globalist scheme to control AI and stating he would substitute the term “super intelligence” on US government documents CITED C09; other heads of government at the same gathering pressed for a single global standard CITED C10; and the Security Council briefing followed this morning with Yoshua Bengio of the UN Independent International Scientific Panel on AI, OpenAI's Sam Altman, Anthropic's Dario Amodei and Hugging Face's Clément Delangue among the anticipated briefers, with DeepSeek and Moonshot AI also invited to speak VERIFIED C07.

2. Two curves, two clocks: the Wednesday thesis

Put the two curves on one axis and the asymmetry is easy to see. The capability curve is set by competitive dynamics between a handful of labs and moves in weeks. The accountability curve is set by rulemaking, supervisory guidance, standards bodies and case law, and moves in years. Both are real. Only one of them is under your control.

Now the second clock. The three supervisory regimes that matter most to the buyers this publication serves all produced the same shape of output in 2026: an instrument that governs the models an institution was already governing, and an explicit or implicit deferral on the agentic case.

3. Frontier ledger: equal-weight, what actually shipped

This section gives each lab the same treatment: what is verifiably shipped, what is announced but not delivered, and the one thing a regulated buyer should care about. Editorial weight is equal by design. That is not the same as equal endorsement.

4. The robot stack went permissive

The most consequential engineering event of the week is the least covered one. On Tuesday, September 22, at ROSCon 2026 in Toronto, Intrinsic — the industrial robotics software company owned by Alphabet — published Intrinsic Core on GitHub under an Apache 2.0 license VERIFIED C05. The company states that the released components are the same capabilities and services it uses day to day in real manufacturing deployments VERIFIED C05.

What is in the box matters more than the license headline. Independent coverage and the company's own description agree on the contents: a hardware-agnostic real-time control framework that adjusts a robot's path mid-motion from sensor feedback; pose estimation built on NVIDIA's FoundationPose, so a robot can locate a part without a rigid fixture; collision-free motion planning; grasp planning that adapts the gripper to how an object is actually sitting; plus simulation, calibration services and ROS 2 drivers VERIFIED C05 CITED C06.

Read this next to NVIDIA's physical-AI releases — Cosmos world models, Isaac simulation frameworks, and the Isaac GR00T N line, with GR00T N1.7 in early access under commercial licensing and a GR00T N2 preview CITED C27 — and a pattern emerges that manufacturing leaders should plan around explicitly. The perception, planning and world-model layers of industrial robotics are converging on open or broadly licensed components supplied by a small number of platform vendors. Differentiation is moving up into task logic and down into hardware and safety certification. If your automation roadmap assumed a multi-year proprietary software moat in the middle of that stack, the assumption needs revisiting this quarter.

5. Financial services: the supervisory gap

The win. DBS is the cleanest publicly documented case of an agentic capability moving from pilot to institutional scale in banking this year. The bank announced on August 19, 2026 that it had rolled out an agentic credit-assessment capability to roughly 1,500 relationship managers and credit risk managers globally, following a pilot with 150 users CITED C23, company-reported. Separately, the bank has extended generative and agentic capability into its customer-facing assistants — DBS Joy for corporate customers and DBS digibot for individuals — across more than 10 million users in Singapore, Hong Kong and Taiwan CITED C23, company-reported. On the financial side, the chief executive has stated an expected revenue uplift from AI adoption of more than SG$1 billion, about US$768 million, for 2025 against SG$750 million in 2024, attributed across the bank's full portfolio of roughly 370 AI use cases rather than to agentic work specifically CITED C24, company-reported.

The instructive detail is the ratio. A 150-user pilot expanded to about 1,500 users is a ten-fold scale-up, and it happened in a credit workflow — the single most heavily supervised judgment in commercial banking. That is only possible if the pilot produced evidence a credit risk function could accept, not merely satisfied users.

The constraint. The supervisory instrument a US bank would reach for does not reach the agent. On April 17, 2026, the Federal Reserve, OCC and FDIC issued revised interagency model risk management guidance as SR 26‑2, superseding SR 11‑7 — and the revised guidance 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 on banks' use of AI CITED C16. FINRA's 2026 oversight report separately notes that autonomous AI agents are rapidly evolving and may present novel regulatory and supervisory considerations, recommending that member firms consider enterprise-level supervisory processes specifically covering the development and use of AI agents CITED C17.

Read together, those two documents describe a genuine gap rather than a loophole. Your model risk management framework will validate the model inside the agent. Nothing in it validates the agent's authority: which accounts it may touch, which thresholds it may cross without a human, what happens when it chains three permitted actions into an outcome no single action would have been approved for. Survey evidence describes the same asymmetry from the institution side, with governance maturity widely reported as trailing deployment speed FLAG Source C38.

6. Healthcare: the only sector with published denominators

The win. Healthcare is currently the only regulated sector publishing agentic outcome figures with an identifiable operator attached, which makes it the most useful evidence base even for readers outside it. Epic reports that at Summit Health its revenue-cycle AI has cut medication prior-authorization submission time by 42%, with 92% of AI-generated responses accepted without edits, and that at health systems most actively using the capability, coding-related denials have fallen by more than 20% CITED C20, company-reported. At HIMSS 2026 in March, Epic previewed Agent Factory, an integrated platform for building and monitoring agents that reason, decide and act across workflows, with traceable agent actions, a visual builder, local policy and knowledge-base attachment, and organization-controlled deployment timing CITED C21. Other vendors put numbers on the same stage — including a claim of $15 billion in prevented denials and 90% reductions in appeal workflow time, and a separate 1.1% underpayment recovery figure worth close to $1 million over three months FLAG Source C22, vendor-reported, denominators not published.

Treat those two tiers of evidence differently. The Summit Health figures have a named operator, a named workflow and a stated acceptance rate, which is the closest thing to a denominator anyone in this market has published. The aggregate vendor totals do not disclose the installed base they are summed across, which makes them unusable for a business case even if they are accurate.

The constraint. The FDA has authorized more than 1,600 AI-enabled medical devices for marketing as of this month VERIFIED C18, but the January 2025 draft guidance on AI-enabled device software functions and lifecycle management remains in draft, and the agency's approach to foundation models and agentic systems is being developed through a discussion paper and public comment rather than a final rule VERIFIED C18, VERIFIED C19. That leaves a specific operational ambiguity that clinical informatics teams are living with right now: a documentation or revenue-cycle agent sits comfortably outside device regulation, a diagnostic agent sits clearly inside it, and the fast-growing middle — agents that surface, rank or pre-populate clinical content that a clinician then signs — sits where the intended-use statement decides the answer.

7. Manufacturing and robotics: hours, not demos

Editorial note on the two programs below. The BMW–Figure and Agility deployments have appeared in this publication repeatedly over the past month, and we are returning to them deliberately rather than for want of alternatives. Today’s manufacturing news is a software-licensing event, not a deployment, and the only way to judge what a permissively licensed control stack changes is against the two programs whose operating hours are actually documented. Treat them here as the baseline, not the story.

The win. The useful manufacturing metric in 2026 is no longer a pilot announcement but accumulated operating hours against a named task at a named site. Two programs meet that bar. Figure's robots have contributed to production at BMW's Spartanburg plant across more than 30,000 vehicles, running material handling and parts transfer on ten-hour shifts five days a week; the earlier 02 units have been retired and Figure 03 moved into a logistics sequencing task at the same plant, with BMW establishing a Center of Competence for Physical AI in Production and planning extension to Plant Leipzig from summer 2026 CITED C25. Agility Robotics reports that Digit has accumulated more than 65,000 operating hours across nine customer facilities, naming GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre as commercial customers, with its RoboFab plant in Salem, Oregon designed for capacity of up to 10,000 units annually at full output CITED C26, company-reported.

Read the task descriptions, not the hour counts. Both programs are in tote and bin movement, light material transfer between stations, and inspection routes where the robot carries a sensor through an environment. High-speed, high-precision work — welding, stamping — remains with traditional fixed-arm industrial robots CITED C25. That boundary is the single most useful planning fact in physical AI right now, and it is stable across every credible 2026 deployment report.

The constraint. Adoption figures for software agents in manufacturing are rising fast from a small base — Deloitte is reported to have measured a four-fold increase in agentic AI adoption in manufacturing in 2026, from 6% to 24% FLAG Source C39, secondary citation, primary report not verified. The constraint on that curve is not model capability and not robot capability. It is the OT/IT boundary. An agent that can reason about a quality excursion cannot act on it unless it can reach the historian, the MES and the quality system, and those three systems in most plants have different identity models, different network zones and different change-control regimes. Quality-inspection results that circulate widely — a 99.7% defect-detection figure and a 40% warranty-claim reduction attributed to computer vision on electronics lines — come from vendor and secondary write-ups without published denominators and should not be used as planning assumptions FLAG Source C40.

8. Energy: the queue is the bottleneck

The win and the constraint are the same fact. Energy is the sector where AI's demand-side footprint and its operational usefulness collide most directly, and 2026 made the collision explicit. On June 18, 2026, FERC issued tailored orders under Section 206 of the Federal Power Act to each US regional grid operator, directing them to address spare generating capacity, queue management so large loads neither wait years nor jump ahead of smaller customers, containment of new substation and transmission costs so they do not land on residential bills, and full cost responsibility for grid work tied to a data center's own connection CITED C28. FERC separately moved to make NERC reliability standards mandatory for data center and other large computational loads, citing documented disturbances in which computational load contributed to bulk-power-system instability — including a July 10, 2024 Eastern Interconnection event in which a transmission fault cascaded and roughly 1,500 MW of data-center load was lost near-simultaneously CITED C29.

The scale of the queue is the operative number. ERCOT was tracking more than 438 GW of large-load interconnection requests as of late August 2026, with data centers making up close to 90% of that total, against an actual ERCOT peak demand near 85 GW; PJM received 811 requests in a single 2026 cycle CITED C30. A request is not committed load, and the gap between 438 GW of requests and an 85 GW peak is mostly speculative and duplicated applications rather than real planned demand — but the engineering effort to process each request is real regardless of whether the load materializes.

On the operational side, utility AI is furthest along in outage prediction, vegetation management, predictive maintenance and grid optimization CITED C31. Duke Energy's self-healing grid program is reported to have prevented more than 1.5 million customer outages CITED C31, company-reported. Note what that program is and is not: automated fault isolation and service restoration on distribution feeders, largely deterministic switching logic operating within engineered limits. It is a good outcome and it is not an agentic system. Conflating the two overstates how far autonomous decision-making has actually penetrated grid operations.

9. Implementation architecture: the accountability layer

Every control described in the four sector sections above is a special case of the same architecture. Here it is once, generally, as the reference we use on AEGIS Diagnostic engagements. AEGIS is the Agentic Enterprise Governance and Intelligence Standard.

10. Workforce: the bench problem nobody priced

The accountability layer is not only a software artifact. Every component in it requires a person who can operate it, and that is where the sharpest constraint in this market actually sits. Anthropic's own published usage research describes the split cleanly: on the first-party API, where businesses deploy at scale, roughly 77% of usage patterns lean toward automation with the human largely absent from individual task loops, while on the consumer product roughly 52% of conversations are augmentative — a human asking, reviewing and iterating CITED C32. Read carefully, that is a description of observed usage patterns in conversations, not a measurement of jobs or of the labor market. But it does say something concrete about staffing: enterprise deployment is where the loop closes without a human in it, which means the humans who remain are supervising outcomes rather than producing them — a different skill, and a scarcer one.

The supply side is tight on every published measure. Reported figures put AI talent demand above supply by roughly 3.2 to 1, with on the order of 1.6 million open AI-related positions against about 518,000 qualified candidates globally, and a majority of decision-makers reporting a moderate-to-extreme shortage FLAG Source C35, aggregator figures, methodology not published. The demand side is simultaneously narrowing the pipeline that would have fed it: two thirds of surveyed enterprises expect to slow entry-level hiring, a large majority report roles already changing or disappearing, and measured entry-level hiring for 22- to 25-year-olds is reported down 13–16% in relative terms CITED C34. The IMF has published staff analysis on new job creation in the AI age that frames the offsetting mechanism — new occupations forming around the technology — on a slower timescale than the displacement CITED C33.

11. Scenario planning and risk–reward

Three plausible twelve-month paths from today's position, with the decision each one rewards. These are scenarios, not forecasts.

Notice that all three scenarios reward the same work. That is unusual and it is the reason to act on it now: the accountability layer is not a bet on a regulatory outcome. It is the position that pays under every outcome, which makes the only real question how much production history you accumulate before you build it.

12. Practitioner FAQ and did-you-know

13. Sources

Research base for this edition. Primary and first-party material is marked VERIFIED where the claim was read directly from the issuing organization. Reputable secondary reporting is marked CITED. Claims marked FLAG are carried because they circulate widely in the market but lack a published denominator or primary source, and should not be used as planning assumptions without re-verification. Company-reported and vendor-reported results are labeled as such in the body. Forecasts, pilots and announced targets are not treated as completed facts.

• C01 Anthropic — Claude Platform release notes, entries dated September 22, September 18 and September 10, 2026 (Claude Opus 5.5 launch and pricing; Compliance API local-session transcripts for Claude in Chrome; Managed Agents auto permission evaluation). Primary source. https://platform.claude.com/docs/en/release-notes/overview
• C02 OpenAI — “Priorities and principles for effective third party assessments,” published September 22, 2026. Primary source. https://openai.com/index/priorities-principles-third-party-assessments/
• C03 Bloomberg — “OpenAI to Let Outside Groups Evaluate AI Models at Earlier Phase,” September 22, 2026. https://www.bloomberg.com/news/articles/2026-09-22/openai-to-let-outside-groups-evaluate-ai-models-at-earlier-phase
• C04 Google Cloud — Gemini Enterprise and Gemini Enterprise Agent Platform release notes (Gemini 3.8 Flash general availability and introductory pricing; AlphaEvolve general availability; Agent Identity, Agent Registry, Agent Gateway and Agent Observability components). Primary source. https://docs.cloud.google.com/gemini/enterprise/docs/release-notes and https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes
• C05 Intrinsic (Alphabet) — “Introducing Intrinsic Core: An open source approach to Physical AI,” September 22, 2026, ROSCon 2026, Toronto. Primary source. https://www.intrinsic.ai/blog/posts/introducing-intrinsic-core
• C06 SiliconANGLE — “Google's robotics unit Intrinsic open-sources its foundational infrastructure for intelligent robots,” September 22, 2026. https://siliconangle.com/2026/09/22/googles-robotics-unit-intrinsic-open-sources-its-foundational-infrastructure-for-intelligent-robots/
• C07 Security Council Report, What's In Blue — “Artificial Intelligence: High-level Briefing,” September 2026 (France presidency; chair; anticipated briefers; DeepSeek and Moonshot AI invited). https://www.securitycouncilreport.org/whatsinblue/2026/09/artificial-intelligence-high-level-briefing-2.php
• C08 UN Web TV — “Artificial intelligence and international security,” Security Council 10228th meeting. Primary record. https://webtv.un.org/en/asset/k1c/k1cdceg9sg
• C09 Scientific American — “Trump rejects AI regulation, citing parallels with climate change, in U.N. address,” September 2026. https://www.scientificamerican.com/article/trump-rejects-ai-regulation-citing-parallels-with-climate-change-in-un-address/
• C10 Euronews — “Leaders push for stronger AI safeguards at UN as Trump touts Super Intelligence,” September 23, 2026. https://www.euronews.com/2026/09/23/leaders-push-for-stronger-ai-safeguards-at-un-as-trump-touts-super-intelligence
• C11 SpaceXAI — API pricing documentation and enterprise API controls (SSO, audit logging, SOC 2, HIPAA-eligible deployment under a business associate agreement, data residency). Primary source. https://docs.x.ai/developers/pricing and https://x.ai/api
• C12 TechRepublic — “SpaceXAI Launches Grok 4.7: Low Prices, Heavy Token Use,” September 2026. https://www.techrepublic.com/article/news-spacexai-grok-4-7-pricing-token-use/
• C13 Digital Applied — “AI Model Releases: September 2026 Tracker and Dated Ledger” (GPT‑6 Sol and Luna availability and pricing tier; GPT‑5.5 retirement scheduled October 14, 2026). https://www.digitalapplied.com/blog/ai-model-releases-september-2026-tracker
• C14 Gibson Dunn — “EU AI Act Omnibus Agreement: Postponed High-Risk Deadlines and Other Key Changes” (Regulation (EU) 2026/1744, Official Journal July 24, 2026, in force July 27, 2026; Annex III to December 2, 2027; Annex I to August 2, 2028; Article 50 unchanged). https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/
• C15 DLA Piper — “The Digital AI Omnibus: Proposed deferral of high risk AI obligations under the AI Act.” https://knowledge.dlapiper.com/dlapiperknowledge/globalemploymentlatestdevelopments/2026/The-Digital-AI-Omnibus-Proposed-deferral-of-high-risk-AI-obligations-under-the-AI-Act
• C16 CIMCON Software — “SR 26-2 Regulates Your Models, Not Your AI Agents: What Banks Need to Know” (interagency revision dated April 17, 2026, superseding SR 11-7; generative and agentic AI stated out of scope; request for information signaled). https://cimcon.com/sr-26-2-regulates-your-models-not-your-ai-agents-what-banks-need-to-know/
• C17 360factors — “Agentic AI Updates Reshaping Compliance in 2026” (FINRA 2026 oversight report treatment of autonomous AI agents). https://www.360factors.com/blog/agentic-ai-updates/
• C18 US Food and Drug Administration — “Artificial Intelligence-Enabled Medical Devices” and the AI-enabled device list (authorized device count; January 2025 lifecycle draft guidance status; training-data description expectations). Primary source. https://www.fda.gov/medical-devices/digital-health-center-excellence/artificial-intelligence-enabled-medical-devices
• C19 US Food and Drug Administration — “FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices.” Primary source. https://www.fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices
• C20 Epic — “Real Results, Right Now: How Epic AI Is Reducing Costs, Improving Care, and Helping Patients” (Summit Health prior-authorization submission time and acceptance rate; coding-related denial reduction at high-usage systems). Company-reported. https://www.epic.com/epic/post/real-results-right-now-how-epic-ai-is-reducing-costs-improving-care-and-helping-patients/
• C21 HIT Consultant — “The Agentic EHR: Epic Unveils Agent Factory and Custom Models,” HIMSS 2026. https://hitconsultant.net/2026/03/10/epic-ai-himss-2026-agent-factory-curiosity-foundation-models/
• C22 Reveleme — “HIMSS 2026 Recap for Medical Practices” (Waystar and FinThrive figures). Vendor-reported at conference; denominators not published. https://www.revelemd.com/blog/why-himss-2026-mattered-for-medical-practices
• C23 DBS — “DBS scales agentic AI to transform way of working for corporate bankers” and “DBS' Gen AI-enabled virtual assistants reach 10 million customers and go agentic.” Company-reported. https://www.dbs.com/newsroom/DBS_scales_agentic_AI_to_transform_way_of_working_for_corporate_bankers_freeing_up_time_for_more_strategic_client_engagements
• C24 CNBC — “CEO of Southeast Asia's top bank DBS says AI adoption already paying off” (revenue uplift attributed across the bank's AI use-case portfolio). Company-reported figure. https://www.cnbc.com/2025/11/14/ceo-southeast-asias-top-bank-dbs-says-ai-adoption-already-paying-off.html
• C25 Humanoid Guide — “Humanoid deployments in 2026 favor Figure and Agility” (BMW Spartanburg vehicle count, shift pattern, Figure 03 task change, Center of Competence, Leipzig extension; task-boundary analysis). https://humanoid.guide/humanoid-deployments-in-2026-favor-figure-and-agility/
• C26 Agility Robotics — corporate site and 2026 deployment disclosures (Digit operating hours across customer facilities, named customers, RoboFab designed capacity). Company-reported. https://www.agilityrobotics.com/
• C27 NVIDIA Newsroom — “NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots” (Cosmos world models, Isaac frameworks, Isaac GR00T N1.7 early access and GR00T N2 preview). Primary source. https://nvidianews.nvidia.com/news/nvidia-releases-new-physical-ai-models-as-global-partners-unveil-next-generation-robots
• C28 American Action Forum — “FERC Data Center Orders Accelerate Grid Connection” (Section 206 orders dated June 18, 2026; the four matters each regional operator was directed to address). https://www.americanactionforum.org/insight/ferc-data-center-orders-accelerate-grid-connection/
• C29 POWER Magazine — “FERC Orders Mandatory NERC Reliability Standards for Data Center and Other Computational Loads” (documented disturbances including the July 10, 2024 Eastern Interconnection event). https://www.powermag.com/ferc-orders-mandatory-nerc-reliability-standards-for-data-center-and-other-computational-loads/
• C30 Shattered.io — “AWS Grid AI Agents Launch as Queue Hits 438 GW” (ERCOT large-load request volume in late August 2026 against actual peak; PJM request count in a 2026 cycle). https://shattered.io/aws-agentic-grid-planning-438gw-queue-2026/
• C31 Renewable Energy World — “How are utilities using AI to rewrite planning playbooks?” and associated utility AI use-case reporting, including Duke Energy self-healing grid figures. Company-reported outage-prevention count. https://www.renewableenergyworld.com/power-grid/how-are-utilities-using-ai-to-rewrite-planning-playbooks/
• C32 Anthropic — Anthropic Economic Index report, June 2026 (first-party API automation-leaning usage share; consumer augmentative share). Describes observed conversation patterns, not labor-market outcomes. Primary source. https://www.anthropic.com/research/economic-index-june-2026-report
• C33 International Monetary Fund — “New Jobs Creation in the AI Age,” Staff Discussion Note SDN/2026/001. Primary source. https://www.imf.org/-/media/files/publications/sdn/2026/english/sdnea2026001.pdf
• C34 IT Pro — “Enterprises are cutting back on entry-level roles for AI, and it's going to create a nightmarish future skills shortage” (entry-level hiring intentions; relative decline for ages 22–25). https://www.itpro.com/business/careers-and-training/enterprises-are-cutting-back-on-entry-level-roles-for-ai-and-its-going-to-create-a-nightmarish-future-skills-shortage
• C35 Qubit Labs — “AI Talent Shortage in 2026: Causes, Challenges and Solutions” (demand-to-supply ratio, open-role and qualified-candidate counts). Aggregated figures; underlying methodology not published. Treat as indicative only. https://qubit-labs.com/ai-talent-shortage/
• C36 Salesforce — “Salesforce and Anthropic Announce Claudeforce,” August 26, 2026 (pilot status ahead of open beta). Company-reported. https://www.salesforce.com/news/press-releases/2026/08/26/salesforce-and-anthropic-announce-claudeforce/
• C37 OpenAI — “Our decision on Cursor following its acquisition by SpaceX” (intent to wind down model supply; proposed shut-off date of November 12, 2026) and TechCrunch on the August 14, 2026 acquisition close. Primary plus secondary. https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/ and https://techcrunch.com/2026/08/15/spacex-officially-closes-its-cursor-acquisition/
• C38 Lyzr — “AI Agents in Banking 2026: From Chatbot Theater to Autonomous Operations” (share of global banks reporting active AI in a core function; governance-maturity lag). Vendor-published survey summary; sample and instrument not disclosed. https://www.lyzr.ai/blog/ai-agents-in-banking-2026-from-chatbot-theater-to-autonomous-operations-lyzr-ai/
• C39 IIoT World — “Agentic AI in Manufacturing: ROI vs. Reality” (adoption increase attributed to Deloitte). Secondary citation; the underlying Deloitte report was not read directly for this edition. https://www.iiot-world.com/artificial-intelligence-ml/agentic-ai-manufacturing-2026/
• C40 DigitalDefynd — “Agentic AI in Manufacturing: 5 Case Studies” (quality-inspection accuracy and warranty-claim figures). Secondary compilation without published denominators. Not suitable as a planning assumption. https://digitaldefynd.com/IQ/agentic-ai-in-manufacturing/