01ProductionSep 29 · Banking
HSBCnio
Clients’ agents get a transaction-banking interface
HSBC launched HSBCnio at Sibos in Miami on September 29: a digital transaction-banking platform for corporate and institutional clients. Clients can reach authorized account and transaction information through HSBC’s own Model Context Protocol server.
HSBC says existing permissions and controls are maintained, and the interface helps treasury teams find and interpret banking data. An “Ask HSBC” AI experience is planned.
Why it mattersThe reverse of the usual agent story. The agents belong to the bank’s clients. Sharpest questions: what identity a client’s agent acts under, how agent permissions are distinguished from the human user’s, and how to revoke a client’s agent mid-session without cutting off the client.
Confirmed · questions · sources
Confirmed: HSBCnio launched Sept 29 at Sibos, Miami; AI-enabled access through HSBC’s own Model Context Protocol server; HSBC says existing permissions and controls are maintained; treasury use case stated; an “Ask HSBC” AI experience is planned. Confidence is medium-high: all three sources are trade outlets, all dated Sept 29, and no HSBC primary press release was located in this pass.
Open: The identity a client’s agent acts under; how agent permissions are distinguished from the human user’s; mid-session revocation for a client’s agent without cutting off the client.
02ProductionSep 29 · Fintech
Plaid x Decagon
Bank connections inside the AI conversation
Plaid Link is now embedded inside Decagon’s conversational AI agents. Consumers connect or reconnect bank accounts mid-conversation without leaving the chat. Decagon says its agents are already in production at Square, Affirm, Chime, and Valon. The agents act on bank connections, failed payments, and disputed charges, and tasks spanning days or weeks resume with full context.
Decagon frames the access as consumer-permissioned. Plaid’s Head of AI GTM, Kathleen McGuirk, said that across Plaid’s AI customers, thousands of consumers are linking billions of dollars in accounts to AI platforms every single day.
Why it mattersAgents act on consumers’ bank accounts mid-conversation. Sharpest questions: credential scope once Plaid Link is connected, where the OAuth token lives, and how a consumer revokes the agent’s access independently of the chat.
Confirmed · questions · sources
Confirmed: Partnership announced Sept 29 via Decagon’s blog and LinkedIn; Plaid Link embedded mid-conversation; production at Square, Affirm, Chime, Valon per Decagon; dialogs act on bank connections, failed payments, disputed charges; multi-day tasks resume with context; the McGuirk quote is Decagon’s published quote.
Open: Credential scope once the consumer connects; where the OAuth token is held; whether the consumer can revoke the agent’s access without exiting the chat.
03LaunchedSep 24 · Financial crime
Feedzai Farol
An agent inside bank fraud operations
Feedzai launched Farol on September 24: an AI agent embedded in RiskOps Studio for fraud investigation. It offers conversational intelligence, automated rule-performance analytics, and contextual case summarizations. The release claims “autonomous execution capabilities” with “transparent audit trails.”
Justinas Rekus, Fraud Prevention Business Owner at SEB, a northern European financial services group operating in more than 20 countries, said a single intelligent interface handling data retrieval, insight generation, and production-ready rule suggestions fundamentally transforms how the team refines fraud strategies.
Why it mattersThis is “production-ready rule suggestions” inside bank fraud operations. Sharpest questions: who approves a suggested rule before it goes live, and whether the audit trail covers the rule change itself as well as the investigation steps.
Confirmed · questions · sources
Confirmed: Farol launched Sept 24, embedded in RiskOps Studio; conversational intelligence, automated rule-performance analytics, contextual case summarizations in the release; the Rekus quote is Feedzai’s published customer voice. Note: only one primary source, the Feedzai press release.
Open: Who holds the go-live permission on a suggested rule; whether the audit trail covers the rule change itself as well as the investigation steps; what “autonomous execution” means in practice.
04ProductionSep 29 · Freight
Flexport MCP Server
Third-party agents book freight against your rates
Flexport launched an MCP server on September 29 letting third-party AI agents track shipments, surface exceptions and customs holds, search rates, and book shipments in plain language. An agent can submit cargo details, book against the company’s negotiated rate, and return a booking ID and transit time without logging into the Flexport platform.
Agents named in the release include Claude, ChatGPT, Microsoft Copilot, and company-built internal agents. Flexport says its own internal AI agent fleet already processes 21 million tasks a year autonomously.
Why it mattersAn agent acting as the shipper commits real freight spend against negotiated rates. The release describes no revocation mechanism or per-transaction approval. Who bears liability when an agent books the wrong lane or rate, and can the shipper revoke the agent mid-flight?
Confirmed · questions · sources
Confirmed: MCP server launched Sept 29; agents can track, surface exceptions and customs holds, search rates, and book outright; booking against the company’s negotiated rate without platform login; Flexport’s internal fleet processes 21 million tasks a year per the company.
Open: Liability when an agent books the wrong lane or rate; whether the shipper gets a per-action audit trail it can revoke mid-flight; no such mechanism is described in the release.
05Rolling outSep 28 · Commerce
Shopify WebMCP checkout
Browser agents can now check out with your money
Shopify added WebMCP support for checkout, including Shop Pay, for all eligible merchants on September 28. Browser-based AI agents can read the checkout screen, update it, and submit the transaction with the buyer’s authorization, using three tools: get_checkout, update_checkout, complete_checkout.
The launch came alongside a Shopify and Meta partnership for agent shopping through Shop Pay checkout, and a personal shopping agent from Instinct that shops and checks out inside one conversation.
Why it mattersThe agent is authorized to spend the buyer’s money. The sharpest question is what “the buyer’s authorization” binds to: a one-time, merchant-scoped, amount-capped confirmation, or a reusable credential the agent could redeploy. The announcement does not describe the authorization’s scope, persistence, or revocation.
Confirmed · questions · sources
Confirmed: WebMCP checkout Sept 28 for all eligible merchants; Shop Pay included; three tools: get_checkout, update_checkout, complete_checkout; buyer’s authorization stated; Shopify x Meta partnership and Instinct personal agent alongside.
Open: The authorization’s scope, persistence, and revocation; whether it binds to one transaction, one merchant, an amount cap, or something reusable.
06Limited betaSep 30 · Corporate services
DoorDash office ordering
Company agents order against corporate benefits
DoorDash launched an MCP-based connector on September 30 letting companies bring DoorDash into internal AI tools or agents they build themselves. Agents can find items, build a cart, place an order, and track arrival, including recurring meal deliveries against corporate meal benefits.
Early beta partners are SpaceXAI, Vercel, Cognition, Tempo, and Mercor. Access is at company level; employees sign in to their own DoorDash accounts before connecting an agent. A company can wire an agent to inventory systems so it restocks snacks, drinks, or supplies as they run low.
Why it mattersA delegated ordering authorization becomes open-ended recurring spend. The release describes company-level access but no per-employee spend limits or revocation controls. It is unclear how a company caps or revokes an agent that reorders autonomously.
Confirmed · questions · sources
Confirmed: MCP-based connector Sept 30; agents can find items, build a cart, place an order, track arrival, and run recurring deliveries against corporate meal benefits; beta partners SpaceXAI, Vercel, Cognition, Tempo, Mercor; company-level access with employee sign-in first; inventory-linked restocking described.
Open: Per-employee spend limits; how a company caps or revokes an agent that reorders autonomously; neither is described in the release.
07General availabilitySep 28 · Healthcare
Cognizant TriZetto agentic claims
Agents adjudicate routine claims on PHI-bearing data
Cognizant reached general availability on September 28 with Workflow Agentic Processing and Enterprise MCP Servers for the TriZetto Facets and QNXT core administration platforms. AI agents follow a health plan’s standard operating procedures to process eligible routine pended claims.
Denials, exceptions, and complex cases route to human reviewers. The release ships a library of more than 100 MCP tools for health plans’ own AI initiatives, including prior-authorization turnaround. No named health-plan customer has been disclosed.
Why it mattersAgents act on PHI-bearing claims inside core payer systems, with write authority over routine adjudications. The release cites visibility, auditability, and human oversight, but the real authority boundary is the classification step deciding what counts as “eligible routine.” A misclassification could finalize payment on a claim that should have been denied or reviewed.
Confirmed · questions · sources
Confirmed: General availability Sept 28; Workflow Agentic Processing and Enterprise MCP Servers on TriZetto Facets and QNXT; agents carry out a health plan’s SOPs on eligible routine pended claims; denials, exceptions, complex cases route to humans; 100-plus MCP tools; prior-authorization turnaround included; no named health-plan customer.
Open: The classification boundary for “eligible routine”; who approves changes to the routing rules; what the cited human oversight covers in practice.
08Production · 3 of 5Sep 30 · Distribution
TwinMed x DeepFabric
Agents transact with 260 suppliers under your name
Twin Med, LLC, a national medical supply distributor, is deploying five DeepFabric AI agents across procure-to-pay, with three already in production. The agents handle purchase-order follow-up, shipment tracking, delivery scheduling, discrepancy resolution, and invoice matching: roughly 3,200 purchase orders a month across 260 suppliers.
The agents run on TwinMed’s ERP and act through EDI, email, and a shipment tracking feed, sharing context with each other. TwinMed’s internal data claims more than 90 percent fewer late receiving-discrepancy reports and about 60 percent faster supplier-issue resolution in the initial production months; those figures are company-reported, not independently verified.
Why it mattersThe agents transact with 260 external suppliers via email and EDI under TwinMed’s name while TwinMed’s supply chain team keeps decision authority. Unresolved: what identity the agents present to suppliers, how supplier-facing actions are attributed and audited, and who can revoke an agent’s authority mid-cycle.
Confirmed · questions · sources
Confirmed: Five DeepFabric agents across procure-to-pay, three live; PO follow-up, shipment tracking, delivery scheduling, discrepancy resolution, invoice matching; approx. 3,200 purchase orders a month across 260 suppliers; agents act through ERP via EDI, email, and a tracking feed; TwinMed’s supply chain team keeps decision authority; reported performance figures are TwinMed’s internal data and are not independently verified.
Open: The identity the agents present to suppliers; attribution and audit of supplier-facing actions; who can revoke an agent mid-cycle.
09RFI · responses due Oct 16Sep 25 · Federal government
ICE agentic software factory
The sharpest procurement design rule this cycle
Immigration and Customs Enforcement posted a request for information on September 25 seeking industry help to build an “agentic software factory”: AI agents would handle a significant share of planning, implementation, testing, and review across the software development lifecycle. Vendor responses are due October 16.
The RFI calls for an enterprise control plane covering governance, architecture, and infrastructure; an agentic orchestration layer with tool access protocols, human escalation, and stop conditions; and an independent assurance boundary for cybersecurity and mission-acceptance testing before production. Separately, ICE’s STELLA platform (Secure Technology Environment for Large Language Agents), implemented earlier this summer, already has at least a dozen fully deployed production-ready AI applications live.
Why it mattersICE’s design rule is the sharpest in any federal procurement seen this cycle: “generation, assurance, and authorization must remain distinguishable and no autonomous component may silently expand its own permissions or approve its own production release.” The live question is whether the selected vendor preserves that separation while ICE hands a single vendor full continuous monitoring responsibilities for its enterprise data and analytics ecosystem.
Confirmed · questions · sources
Confirmed: RFI posted Sept 25; vendor responses due Oct 16; enterprise control plane, agentic orchestration layer with human escalation and stop conditions, independent assurance boundary for cybersecurity and mission-acceptance testing; STELLA has at least a dozen fully deployed production-ready AI applications live. Note: this is market research, not a deployment. The factory is the desired future state; the STELLA applications are production per ICE’s own documents.
Open: Whether the awarded design preserves the generation/assurance/authorization separation; concentration risk where one vendor holds full continuous monitoring of the data and analytics ecosystem.
10Contract signed · pilots Oct 2026Sep 30 · Telecom
Samsung AI-RAN with KT and SK Telecom
Network AI at the edge of agency
Samsung Electronics signed contracts on September 30 to join South Korea’s government-backed Hyper AI Network program: a two-year initiative testing whether private 5G networks and edge AI can support robots doing real industrial work.
Samsung is the sole global vendor for KT’s project and a main vendor for SK Telecom’s, deploying its NIS and CognitiV NOS stack for KT. Field pilots begin in October 2026. The consortia were selected in July by the Ministry of Science and ICT and the National Information Society Agency.
Why it mattersGovernance caveat: this is network AI more than autonomous agents, so the agentic framing is thin. It earns a record for where the authorization boundary lands: when AI systems inside a carrier’s network begin reconfiguring RAN behavior for industrial customers, which party holds authority if a network action degrades a production robot cell? The carrier, the industrial operator, or the regulator?
Confirmed · questions · sources
Confirmed: Contracts signed Sept 30; Hyper AI Network is a two-year government-backed initiative testing private 5G and edge AI for industrial robots; Samsung is the sole global vendor for KT and a main vendor for SKT; NIS and CognitiV NOS stack for KT; field pilots begin October 2026; consortia selected in July by MSIT and NIA.
Open: The authorization boundary between carrier, industrial operator, and regulator if a network action degrades a production robot cell; kept on record with the network-AI caveat attached.