Download this edition as PDF
We'll email a 6-digit access code. Enter it to unlock the Daily Market Scan PDF.
Daily Market Pulse · 2026-08-29
Autonomy became a scored variable.
Three separate bodies stopped asking how good the model is and started asking how independently it acts. The FDA put a number on it. A Senate bill put a duty on it. NIST is writing separate control overlays for one agent versus many. If you sell into financial services, healthcare, manufacturing or energy, the buying question just changed shape.
Frontier & Industry Intelligence : Regulated Sectors. FinServices, Healthcare, Energy, Manufacturing
Top takeaway
One sentence
The unit of AI regulation shifted this month from model capability to delegated authority, and the vendors moved the same week: NVIDIA shipped a CPU it calls the first built for AI agents and an agentic infrastructure benchmark VERIFIED C01, while Salesforce pulled Claude inside its trust boundary VERIFIED C02. Autonomy is now something you benchmark, price, and answer for.
Contents
What is in this edition
01 · The lede
Three bodies started scoring autonomy, not intelligence
For three years the regulatory conversation was about models: how capable, how large, how trained. In the last six weeks that framing quietly lost its grip. Two new publications, on July 21 and August 18, score something else entirely, and they land on the same axis as an ongoing NIST standards effort that has been running since 2025. Three bodies, no coordination, one variable.
On August 18, 2026, the FDA's Digital Health Center of Excellence released a discussion paper on generative AI-enabled medical devices. Its proposed risk framework has two axes. One scores how independently a function acts, running from non-directive information through action-directing information to supervised action and then fully autonomous action. The other scores the consequence of relying on an incorrect output, from limited to severe. VERIFIED C09 Model quality is not an axis. Delegated authority is.
On July 21, 2026, S.5051 (the AI AGENT Act of 2026) was referred to Senate Commerce. Its operative definition of a "custodial user agent" is an agent authorized to act for a user in a manner that is transparent, documented, limited and revocable, that keeps real-time records of actions taken, and that cannot hand authority to another agent without permission. VERIFIED C10 That is not a model-safety definition. It is a delegation-of-authority definition, and it reads like a control objective.
And NIST's COSAiS project is developing SP 800-53 control overlays that treat "Using AI Agent Systems (Single Agent)" and "Using AI Agent Systems (Multi-Agent)" as separate use cases with separate overlays. CITED C11 The number of agents in the loop is being treated as a control-relevant variable in its own right.
Why this matters commercially. If autonomy is the scored variable, then the cheapest thing you can do to a stalled deployment is not a better model. It is to move the system one cell to the left, from autonomous action to supervised action, or from action-directing to non-directive, and ship the narrow version now. Most enterprise agent programs we see are stuck because they were scoped at the top-right cell and then had to survive a risk review designed for the bottom-left.
02 · Frontier ledger
What actually shipped, and what was only announced
We separate these deliberately. A shipped product and an announced target are different instruments, and mixing them is how procurement decks get people into trouble.
| Lab | Shipped | Announced, not yet delivered | Source |
|---|---|---|---|
| OpenAI | Site tools with WebMCP for agent-to-website interaction; Codex reported at 3 million weekly active users; APIs reported at more than 15 billion tokens per minute. CITED C05 | Enterprise revenue parity with consumer "by end of 2026." IBM Consulting Advantage embedding across legacy systems and cybersecurity. | CITED C05 company-reported |
| Anthropic | Claude embedded in the Agentforce Atlas Reasoning Engine; Salesforce plugin inside Claude with 37 prebuilt sales skills, live for select pilot customers. | Open beta of Salesforce in Claude targeted September 2026. "First LLM provider fully integrated within the Salesforce Trust Boundary" is a Salesforce characterization. | VERIFIED C02 |
| Gemini 3.7 Flash for coding and agentic workflows, three weeks after 3.6 Flash. Grok 4.6 distribution added to Google's enterprise agent platform. | Nothing pending. Note the other direction: gemini-robotics-er-1.6-preview is scheduled for shutdown August 31, 2026, two days from this edition. Check your pinned model IDs. | CITED C06 | |
| xAI / SpaceX / Cursor | Grok 4.6, August 12, 2026: 500k context, published pricing $2 / $0.50 / $6 per 1M tokens under 200k prompt tokens CITED C04. Distribution via Microsoft Foundry. SpaceX closed the ~$60B all-stock Cursor acquisition, reported August 15, 2026. VERIFIED C03 | Grok Bot is in beta. Cursor's integration into SpaceXAI teams is in progress, not complete. SpaceXAI named as a future deployer of NVIDIA Vera CPUs. | CITED C04 VERIFIED C03 VERIFIED C01 |
| Open weights, non-US | Moonshot Kimi K3 (July 17, 2026); DeepSeek V4 Flash Vision Exp (August 21, 2026); Alibaba opened Qwen 3.8 Max text weights (August 12, 2026); Mistral 3 family and Mistral Small 4. | Nothing pending. | CITED C18 |
The jurisdiction signal most US buyers are still missing
NVIDIA's own SEC filing names the open models it optimized for local deployment: DeepSeek v4 Flash, Diffusion Gemma, Nemotron 3.5 Lightning and Qwen 3.8. VERIFIED C01 Two of those four are Chinese. That is a US-listed company's regulatory filing, not a blog post.
For a regulated buyer this is not a geopolitical talking point, it is a model-provenance control question. If your architecture review does not currently record the training jurisdiction and license of every weight in the stack, add that field before your next audit does it for you. The same filing shows NVIDIA guiding Q3 at $108B while assuming zero China data center compute revenue VERIFIED C01. The supply chain is already pricing jurisdiction. Your control framework should too.
03 · Compute and capital
The quarter that reset the denominator
NVIDIA reported Q2 FY2027 on August 26, 2026, for the quarter ended July 26, 2026. Revenue of $96.221 billion, up 18% sequentially and 106% year over year. Data Center revenue of $89.0 billion, up 117% year over year. GAAP and non-GAAP gross margins both at 75.0%. VERIFIED C01
Three items in that filing matter more to a regulated-industry buyer than the headline number.
Agents got a benchmark and a CPU
NVIDIA introduced Vera, which it describes as the first CPU built for AI agents, and named AgentPerf as the industry's first agentic AI infrastructure benchmark. VERIFIED C01 Both are company characterizations. But the direction is the point: when infrastructure starts being benchmarked on agentic workloads rather than training throughput, procurement questions follow within two quarters.
The third item is physical AI. In the same filing NVIDIA announced the Isaac GR00T Reference Humanoid Robot, an open humanoid reference design on Jetson Thor; Halos for Robotics, described as a full-stack safety system unifying AI compute and safety; and Cosmos 3, an open frontier omnimodel for physical AI. VERIFIED C01 A reference design plus a named safety stack is what a supply chain looks like just before unit economics change.
04 · Regulated industry
Financial services
The win: a supervisor said the quiet part plainly
In a written reply to a Parliamentary Question for the sitting of August 5, 2026, the Monetary Authority of Singapore confirmed it applies a principles-based supervisory framework that covers all AI use cases, including agentic AI. VERIFIED C08 The practical effect for a bank is liberating: you do not need a new agent-specific rulebook to start. Your existing model risk, third-party risk and outsourcing frameworks already have jurisdiction. The gap is evidence, not authority.
The control: make revocability testable, not asserted
Borrow the S.5051 definition VERIFIED C10 as a control objective even though the bill is not law. Every agent with authority over a customer-affecting action should be able to answer four questions in an audit, with artifacts:
- Transparent. Can the customer and the reviewer see that an agent, not a person, took the action?
- Documented. Is there a real-time record of actions taken, written at the time, not reconstructed from logs afterwards?
- Limited. Is the authority scoped to an enumerated action list, and does the system fail closed outside it?
- Revocable. Has someone actually pulled the authority in a drill and measured how long it took to take effect? An untested kill switch is a diagram.
Scenario, cause and effect. A mid-size lender ships an agent that drafts adverse action notices and files them without review. It sits in the top-right cell of the autonomy matrix: fully autonomous action, severe consequence. When the supervisor asks for the revocation drill record and there is none, the remediation is not a model change; it is a six-week pause. Moving the same agent one cell left (draft, queue, human files) costs roughly two sprints and survives the review. The narrow version ships. The broad version does not.
05 · Regulated industry
Healthcare
The win: documentation burden is measurable, and it moves
A JAMIA Open study published in August 2026 ran 9 primary care physicians through 4 simulated encounters each, with and without an AI scribe. Documentation consumed 36.3% of encounter time without a scribe and 11.2% with one. VERIFIED C14 That is peer-reviewed, dated this month, and specific about its method.
The control: put the FDA's own two axes into your intake form
The FDA discussion paper VERIFIED C09 gives you a free, defensible vocabulary before it is ever guidance. Add two required fields to your AI intake:
- Autonomy tier. Non-directive information / action-directing information / supervised action / fully autonomous action.
- Consequence tier. Limited / moderate / severe if the output is wrong and relied upon.
Then set one rule: anything scoring action-directing or above crossed with severe needs named clinical ownership before pilot, not before scale. Two fields, no new committee. And file a comment before October 19, 2026. The docket is open, and the organizations that shape a framework are the ones that answered while it was still a discussion paper. VERIFIED C09
06 · Regulated industry
Manufacturing and robotics
The win: a peer-reviewed number, from one plant, on one process
Published in Production Planning & Control in 2026: a domain-adapted LLM combined with plant process data on tyre curing presses reduced mean dry cycle time from 106 to 80 seconds, stabilized the curing process and added 304 tyres of daily capacity. CITED C16 Note what made it work. Not a frontier model, a domain-adapted model wired into real process data on a single bounded process with a physical measurement of success.
The control: a supply-chain question, not a robotics question
NVIDIA's filing named an open humanoid reference design (Isaac GR00T on Jetson Thor) and a full-stack safety system (Halos for Robotics) in the same quarter. VERIFIED C01 Open reference designs collapse differentiation and compress prices, which is good for you as a buyer and dangerous if you have signed a long exclusivity. Before your next physical-AI contract, ask three things: which reference design does this sit on, what is the functional safety story independent of the vendor's own stack, and what happens to my spares and service if this vendor is acquired. That last question stopped being hypothetical this month. VERIFIED C03
07 · Regulated industry
Energy and the grid
The win: the regulator moved before the failure, not after
FERC is treating data centers as a distinct reliability class. Under Docket RM26-4-000, FERC issued an order on July 16, 2026 directing NERC to act on reliability standards for "computational loads," having already issued six Federal Power Act section 206 show cause orders on June 18, 2026 to PJM, MISO, SPP, CAISO, ISO New England and NYISO on their large-load interconnection rules. VERIFIED C12 For an energy-sector buyer, that is a rare thing: a clear, dated, docketed proceeding you can plan against.
The control: negotiate ride-through before you negotiate megawatts
If you are siting AI compute, the RM26-4 proceeding VERIFIED C12 tells you what your interconnection agreement will eventually be asked to contain: defined voltage and frequency ride-through behavior, curtailment obligations, and telemetry the operator can actually see. Put those terms in the deal now, while they are a negotiating position, rather than in 2027 when they are a compliance obligation. The same logic applies to NVIDIA's SB Energy arrangement for land, power and shell capacity at the PORTS-Pike campus in Ohio VERIFIED C01: the scarce asset in that sentence is not the silicon, it is the interconnection.
08 · The clock
The compliance clock, read honestly
This is the section most likely to correct something you currently believe.
The practical consequence. December 2, 2027 is roughly fifteen months out. A conformity assessment, a technical file and a post-market monitoring plan for a system you have not yet frozen is not a fifteen-month job with slack in it. Teams that read the deferral as permission to stop will restart in mid-2027 against the same auditor availability as everyone else.
09 · Practical
Five controls you can stand up in a fortnight
None of these require a platform purchase, a new committee, or a budget cycle. All five are things we have seen unblock a stalled agent program.
| # | Control | Why now | Evidence it produces |
|---|---|---|---|
| 1 | Autonomy tier on every AI intake form. Four values: non-directive / action-directing / supervised action / fully autonomous. | It is the FDA's own axis VERIFIED C09 and it costs one dropdown. It also gives risk a way to say yes to the narrow version. | A tiered inventory you can hand to an auditor. |
| 2 | Revocation drill. Pull an agent's authority in production, time how long until the last in-flight action stops. | "Revocable" is in the S.5051 definition VERIFIED C10 and in every supervisory conversation. Almost nobody has measured it. | A dated drill record with a measured time-to-effect. |
| 3 | Model provenance field. Record training jurisdiction and license for every weight in the stack, including embedded and open models. | Two of the four open models NVIDIA names as optimized are Chinese VERIFIED C01, and the guidance assumes zero China compute. Provenance is already being priced. | A defensible answer to a question you will be asked. |
| 4 | Single-agent versus multi-agent split in your control documentation. | NIST is writing them as separate SP 800-53 overlays CITED C11. Aligning early is free; retrofitting is not. | Control mapping that survives the overlay's publication. |
| 5 | Real-time action log, written at the time. Not reconstructed from application logs after the fact. | "Real-time records of actions taken" is the operative phrase in S.5051 VERIFIED C10 and the practical basis of task-level provenance. CITED C22 | An audit trail that answers "what did it do and on whose authority." |
Risk and reward, stated plainly
The reward for doing these five now is that you keep shipping while your competitors pause. The risk of skipping them is not a fine, because most of this is not yet law. It is a rework cycle. Every one of these five is cheap to add at design time and expensive to add after a system is in production and generating customer-affecting decisions. The gap between those two costs is the whole argument.
10 · Did you know / FAQ
Five things worth knowing this week
Did you know: a Google robotics model shuts down in two days
gemini-robotics-er-1.6-preview is scheduled for shutdown on August 31, 2026. CITED C06 If anything in your physical-AI stack pins that model ID, you have a weekend. Preview-tier model IDs are not a dependency you can treat as infrastructure.
Did you know: the "95% of pilots fail" stat is a year old
It comes from MIT NANDA's "The GenAI Divide," reported August 2025. FLAG · Source C20 It is still being quoted in 2026 without its date, and it predates the entire agent-platform generation it is now used to describe. Use it as a 2025 baseline or not at all.
FAQ: does the EU deferral mean we can stop?
No. Article 50 transparency and AI-content labeling took effect August 2, 2026, and Article 5 prohibitions and GPAI obligations were already live. VERIFIED C07 The deferral moved engineering-heavy obligations, not disclosure obligations. Your labeling and disclosure work is due now.
FAQ: is the AI AGENT Act something we need to comply with?
Not yet. S.5051 was referred to committee on July 21, 2026 and is an introduced bill, not law. VERIFIED C10 Its value today is as a control vocabulary (transparent, documented, limited, revocable) that happens to match what supervisors are already asking for informally.
On workforce numbers: use the institutions, not the headlines
AI labor-impact figures circulate faster than they are checked, and most of the ones in your inbox this week have no methodology attached. If you need a defensible baseline for workforce planning in a regulated environment, two institutional sources carry one: the OECD Employment Outlook 2026, published July 7, 2026, and IMF Staff Discussion Note SDN/2026/001, "Bridging Skill Gaps for the Future: New Jobs Creation in the AI Age." CITED C19 Neither will give you the dramatic number a slide wants. Both will survive an executive challenging the source, which is the more useful property. Anchor the plan to those, then let the vendor claims argue with them.
The concept worth carrying into Monday: authority, not capability
Every framework cited in this edition (the FDA's two axes VERIFIED C09, S.5051's custodial agent VERIFIED C10, NIST's single versus multi-agent split CITED C11) describes the same object from a different angle: how much authority has been delegated, to what, and how easily can it be taken back. Model capability is now assumed. Delegated authority is what gets governed. Teams that reorganize their AI inventory around authority rather than around use case find that most of their governance backlog collapses into three or four patterns.
11 · Corrections desk
What we are correcting, and why
| Claim in circulation | What the primary source says | Status |
|---|---|---|
| "EU AI Act high-risk obligations became enforceable August 2, 2026." | Regulation (EU) 2026/1744 moved Annex III to December 2, 2027 and Annex I to August 2, 2028. Article 50 transparency took effect August 2, 2026. | VERIFIED C07 Corrected |
| "Singapore became the first major financial regulator to put agentic AI inside binding bank rules." | MAS's August 5, 2026 parliamentary reply describes a technology-neutral, principles-based framework already covering all AI. The AIRG Guidelines are still to be finalized. | VERIFIED C08 Corrected |
| "95% of enterprise AI pilots fail." (presented as current) FLAG · Source C20 | MIT NANDA, reported August 2025. Just over a year old; predates the 2026 agent-platform generation. FLAG · Source C20 | Dated |
| Banking agentic AI uplift percentages (44% adoption; 20–40% cost; 10–30% revenue). FLAG · Source C21 | Trace to vendor blogs, not supervisory or audited disclosure. The Commonwealth Bank fraud figure is company-reported. FLAG · Source C21 | Not used |
| "Fed, OCC and FDIC revised model risk guidance in April 2026 to exclude agentic AI from scope." | Could not be confirmed against a first-party Federal Reserve, OCC or FDIC release in this research window. | FLAG · Source C23 Open question |
| NVIDIA "mobilized $500 billion" for AI infrastructure. VERIFIED C01 | Announced platforms with six firms to mobilize over $500B over time, subject to definitive agreements. VERIFIED C01 | Announced intent |
12 · Research base
Sources
Every figure, date and claim above resolves to an entry below. Entries marked FLAG are reported as contested and are not used as evidence in the industry sections.
- C01 NVIDIA Q2 FY2027 results and product ledger. SEC Form 8-K press release, August 26, 2026 (quarter ended July 26, 2026). Revenue $96,221M (+106% Y/Y); Data Center $89.0B (+117% Y/Y); Edge Computing $7.2B (+27% Y/Y); gross margin 75.0%; Q3 FY2027 outlook $108.0B ±2% assuming no China data center compute. Vera Rubin in full production; NVIDIA Vera described as the first CPU built for AI agents; AgentPerf named as the first agentic AI infrastructure benchmark; Isaac GR00T Reference Humanoid Robot; Halos for Robotics; Cosmos 3; BioNeMo Agent Toolkit; compute financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR targeting over $500B, subject to definitive agreements; SpaceXAI named as a Vera CPU deployer; SB Energy partnership at the PORTS-Pike Technology Campus, Ohio. SEC filing · Fortune
- C02 Salesforce and Anthropic announce Claudeforce, August 26, 2026. Claude embedded as a reasoning model across Salesforce products and in the Agentforce Atlas Reasoning Engine; Salesforce embedded in Claude as a plugin with 37 prebuilt sales skills; company states Claude is the first LLM provider fully integrated within the Salesforce Trust Boundary. Pilot now; open beta targeted September 2026. Salesforce release · CNBC
- C03 SpaceX closed its acquisition of Cursor (Anysphere), reported August 15, 2026; all-stock transaction valued at approximately $60 billion, announced June 16, 2026. TechCrunch · CNBC
- C04 xAI Grok 4.6, released August 12, 2026. 500k context window; published API pricing $2 / $0.50 / $6 per 1M tokens below 200k prompt tokens, $4 / $1 / $12 above. Distribution added via Microsoft Foundry and Google's enterprise agent platform. Grok Bot beta announced August 11, 2026. xAI release notes · Releasebot
- C05 OpenAI enterprise position, company-reported: enterprise more than 40% of revenue and tracking to parity with consumer by end of 2026; Codex at 3 million weekly active users; APIs at more than 15 billion tokens per minute; IBM partnership embedding OpenAI models into IBM Consulting Advantage; Site tools with WebMCP for agent-to-website interaction. OpenAI · OpenAI enterprise report
- C06 Google Gemini cadence and deprecations. Gemini 3.7 Flash succeeds 3.6 Flash for coding and agentic workflows; gemini-robotics-er-1.6-preview scheduled for shutdown August 31, 2026; several image generation models shut down August 17, 2026. Gemini API changelog · Gemini Enterprise release notes
- C07 EU AI Act Digital Omnibus, Regulation (EU) 2026/1744, published in the Official Journal July 24, 2026, in force July 27, 2026. Annex III standalone high-risk obligations moved to December 2, 2027; Annex I to August 2, 2028. Article 50 transparency, Article 4 AI literacy, GPAI provider obligations and Article 5 prohibitions unchanged. AI Act explorer · Gibson Dunn · Goodwin
- C08 Monetary Authority of Singapore, written reply to a Parliamentary Question, sitting of August 5, 2026: MAS applies a principles-based supervisory framework covering all AI use cases including agentic AI. Guidelines on AI Risk Management (consulted November 2025) to be finalized; industry AI Risk Management Toolkit developed under Project MindForge. MAS reply · MAS AIRG
- C09 FDA CDRH Digital Health Center of Excellence, "Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback," August 18, 2026. Docket FDA-2026-N-7874; comments due October 19, 2026. Proposes a two-axis framework scoring how independently a function acts and the consequence of an incorrect output. Not draft or final guidance; proposes no policy change. FDA discussion paper · FDA announcement
- C10 AI AGENT Act of 2026 (S.5051), introduced by Sen. Mark R. Warner, referred to Senate Commerce, Science and Transportation on July 21, 2026. Defines a custodial user agent as transparent, documented, limited and revocable, with real-time action records and no sub-delegation of authority without permission. Covered platforms at 50 million or more US customers. Introduced bill, not law. Congress.gov · Davis Wright Tremaine
- C11 NIST agentic work. Control Overlays for Securing AI Systems (COSAiS) developing SP 800-53 overlays including single-agent and multi-agent AI system use cases, in development as of March 2026. Center for AI Standards and Innovation AI Agent Standards Initiative announced February 17, 2026; AI Agent Interoperability Profile targeted Q4 2026. NIST COSAiS · NIST AI RMF
- C12 FERC Docket RM26-4-000, Interconnection of Large Loads to the Interstate Transmission System. July 16, 2026 order directing NERC on reliability standards for computational loads; June 18, 2026 section 206 show cause orders to PJM, MISO, SPP, CAISO, ISO New England and NYISO. NERC has documented a July 2024 Eastern Interconnection event in which a line fault led to the loss of roughly 1,500 MW of data center load. FERC RM26-4 · White & Case
- C13 Data center electricity demand forecast: US 31 GW (2025), 41 GW (2026), 66 GW (2027); global consumption 565 TWh in 2026 versus 447 TWh in 2025. Vendor-published market report; treat directionally. US share of global data center electricity reported at nearly 40%. Bloom Energy 2026 power report · Forbes
- C14 JAMIA Open, August 2026. Simulation study, 9 primary care physicians, 4 simulated encounters each, with and without an AI scribe. Documentation consumed 36.3% of encounter time without a scribe versus 11.2% with one. Simulation, not a live-practice randomized trial. JAMIA Open
- C15 Pew Research Center, "From diagnoses to treatments: Why Americans use AI chatbots for health," August 25, 2026. Survey of 3,488 US adults, fielded June 22–28, 2026. 34% use AI chatbots for at least one health reason; 47% call answers extremely or very helpful; 29% are very comfortable sharing personal health data with the tools; 44% of adults 18–29 versus 17% of adults 65 and older are chatbot health users. Pew Research Center · Healthcare Dive
- C16 Production Planning & Control (Taylor & Francis, 2026), peer-reviewed case study. A domain-adapted LLM combined with plant process data on tyre curing presses reduced mean dry cycle time from 106 to 80 seconds, stabilized the curing process and increased daily curing capacity by 304 tyres. Single-plant case study. Production Planning & Control
- C17 Humanoid robot deployment data, company-reported. Figure 02 at BMW Spartanburg: 11-month deployment, more than 90,000 sheet metal parts across roughly 1,250 operational hours. Agility Robotics Digit: more than 65,000 operating hours across nine customer facilities, named customers including GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre. BMW AEON targeted at high-voltage battery assembly, extending to Plant Leipzig from summer 2026. Task scope remains material handling, bin picking and simple assembly. Deployment tracker · Technology.org
- C18 Open-weight releases outside the US. Moonshot AI Kimi K3, July 17, 2026 (reported 2.8-trillion-parameter open model). DeepSeek V4 Flash Vision Exp, August 21, 2026. Alibaba opened Qwen 3.8 Max text weights August 12, 2026. Mistral 3 family (Mistral Large 3, 675B total / 41B active, 256K context, December 2025) and Mistral Small 4 (March 2026). Cross-checked against C01, in which NVIDIA independently names DeepSeek v4 Flash and Qwen 3.8 among optimized open models. Release timeline · Open-weights history
- C19 Labor and macro reference base. OECD Employment Outlook 2026, published July 7, 2026. IMF Staff Discussion Note SDN/2026/001, "Bridging Skill Gaps for the Future: New Jobs Creation in the AI Age." OECD · IMF SDN/2026/001
- C20 The "95% of enterprise AI pilots fail" figure originates in MIT NANDA's "The GenAI Divide: State of AI in Business," reported August 2025. It is just over one year old and predates the 2026 agent-platform generation, but is still republished in 2026 commentary without its date. Fortune, August 2025
- C21 Circulating 2026 banking agentic AI figures, including 44% of finance teams using agentic AI, 20–40% cost reductions and 10–30% revenue uplift at named banks, and a Commonwealth Bank report of a 20% reduction in fraud losses in the first half of FY2026 using an internally built agent. Bank figure is company-reported; aggregate percentages trace to vendor blogs rather than supervisory or audited disclosure. Not used as evidence in this edition. Vendor summary
- C22 Agent security as a gating constraint on scaling autonomous workflows, with platform-level governance, task-level provenance and human-in-the-loop checkpoints recommended; related work includes Google's Agent Payments Protocol. Google Cloud blog · Weekly agent news roundup
- C23 Contested: secondary commentary asserts the Federal Reserve, OCC and FDIC issued revised interagency model risk management guidance on April 17, 2026 placing agentic AI outside its formal scope. We could not confirm this against a first-party release in this edition's research window. Reported as an open question, not as fact. Secondary commentary