In this briefing
- 01Salesforce puts job definitions, data models and availability states in one product catalogue
- 02The financial-adviser product binds connectors, skills and human approval
- 03Nursing and corporate security show two different forms of high-risk embedding
- 04The job workflow is becoming a more complete unit of acceptance than the model name
- →What to watch next
- ↗Sources and verification
- Salesforce's job-ready Agents span both GA products and products still in pilot: several Agents for service, employee support, shopping and supply chain are GA, while outbound-sales Agent Hunter remains in pilot and is planned for GA in 2026-11; the long-horizon runtime is currently used first for Hunter.
- Claude for Financial Advisors and Oracle's nursing Agent both connect AI to existing business systems and retain professional approval or control; the former is available to install, while the latter's announced availability is limited to the United States and should not be extrapolated to other regions or all regulated tasks.
- Dataminr labels three corporate-security Agent capabilities as GA while listing other search, briefing and API capabilities for later release. Vendor descriptions and customer cases can define which functions require acceptance testing, but cannot directly prove that accuracy, risk or business outcomes are comparable across jobs.
Salesforce puts job definitions, data models and availability states in one product catalogue
On 14 September, Salesforce announced a set of job-ready agents covering customer service, IT and human-resources service, shopping, outbound sales, supply chain, inbound sales and customer experience. The official material says that each Agent comes with skills, actions and data models for the relevant work, and connects to customer context and business processes in Customer 360. Agent Script allows AI reasoning to be combined with deterministic rules that constrain decisions and actions.
The products are not at the same maturity. Casey, Paige, Carter, Marshall, Piper and Fin are labelled GA, while Hunter for outbound sales remains in pilot and is planned for GA in 2026-11. Salesforce also describes a long-horizon runtime that can preserve progress, continue working and adjust dynamically over days or weeks, but explicitly says that Hunter is the first Agent to run in that environment; other jobs and customer-built long-running tasks remain future scope. AI Skills remains in pilot, with GA planned for 2026-10; Agent Optimizer's planned GA month is likewise 10. Multi-Agent Orchestration is GA.
A job-specific catalogue refines the procurement object from a general assistant into a combination of job, task, data model, action permissions and runtime. Enterprises still need to lock each item to its GA, pilot or planned state and validate it within their own organisation structures, approval rules and business systems. Different availability labels within one platform mean that a GA service Agent cannot serve as evidence for a long-horizon sales Agent that remains in pilot.
The financial-adviser product binds connectors, skills and human approval
Anthropic released Claude for Financial Advisors on the same day, comprising connectors and workflow skills designed for advisers. The listed connections include CRM, custodians, portfolio reporting, financial planning, estate tools and meeting-recording systems. The plugin routes skills for meeting preparation, portfolio analysis, alternative-investment briefs, compliance and AI policy into tools that firms already use. Advisers can choose which partners to connect, and institutions can adopt or revise the public skills repository to fit their own processes and writing style.
The product writes regulatory boundaries into the workflow: investment advice, client communications, compliance judgements and other regulated activities still require human review and approval; administrative actions such as CRM updates or client communications first enter a pending-review state. The adviser plugin is available, and Anthropic recommends that registered investment advisers use the Enterprise plan with audit logs. Here, available means that the connectors and skills have been released; it does not prove that every institution's authorisation, record retention, data quality or regulatory review is satisfied automatically.
The core configuration of a professional-job Agent is not just a prompt. It also includes which systems it may read, which actions it may perform, which critical points require human approval and which audit records are retained. Encoding these fields in a job workflow can keep business objects, permissions and operational outcomes connected. Final responsibility nevertheless remains with the institution and its professionals, and a connector count cannot replace adaptation and compliance acceptance testing.
Nursing and corporate security show two different forms of high-risk embedding
On 14 September, Oracle announced that Oracle Health Clinical AI Agent for nurses was available in the United States. The capability is embedded directly in Oracle Health Foundation EHR and provides voice-driven chart navigation and search, acute-care summaries, and voice entry of structured records at the point of care. The official description keeps clinicians in control and limits its aim to reducing search, documentation and coordination work. The page supplies no randomised controlled evaluation, error rate or availability outside the United States for this set of nursing capabilities. Oracle's figure of more than 400,000 hours saved comes from an earlier clinical-note capability for physicians and cannot be treated as an outcome already achieved by the new nursing product.
On the same day, Dataminr labelled Agentic Corroboration, Agentic Context and Near-Term Predictive Intelligence as GA for Dataminr Advanced for Corporate Security customers, covering event verification, contextual organisation and assessment of how near-term risks may evolve. The company says that it uses more than 60 task-specific fine-tuned models and an event archive covering more than a decade. Real-Time Brief Builder, Watchlist Agent, Agentic Search, next-generation custom alerts and Advanced API are listed for later integrated release rather than as part of the current GA set. Claims about accuracy, speed and predictive capability come from the vendor, and the public page provides no independent evaluation under matched conditions.
Nursing documentation and corporate security are both time-sensitive, risk-sensitive workflows, but one is embedded in an electronic health record under clinician control, while the other operates around public information, a proprietary event archive and security-team decisions. Their costs of error, human-review points, evidence-retention needs and update deadlines differ. A job Agent can therefore be accepted only against its specific responsibilities; the use of task-specific models does not give both jobs a common accuracy measure or level of automation.
The job workflow is becoming a more complete unit of acceptance than the model name
The four releases can be compared across four dimensions: the job objective and task set; connected business data and systems; deterministic rules or human-control boundaries; and evidence of actions and outcomes. Salesforce emphasises job data models, rules and long-horizon progress; Anthropic emphasises adviser-tool connections and approval for regulated matters; Oracle places the capability inside a nursing EHR; and Dataminr combines specific security capabilities with a proprietary event archive. The four official sources provide no common interface, state format or evaluation standard for comparison across products, and no evidence of interchangeability.
Maturity labels are also part of product semantics. Salesforce places GA, pilot and planned release dates on the same page. Anthropic's adviser plugin was available on 14 September but still depends on institutional configuration and review. Oracle announces availability only in the United States. Dataminr clearly separates three current GA capabilities from several upcoming capabilities. Vendor figures for customer proportions, time saved or model counts each have their own definitions; without task definitions, a baseline, a time window and independent review, they cannot be combined into a cross-job ranking.
An enterprise can make the job-workflow package its minimum unit of acceptance: define who sets the objective, which business objects are read, which actions are performed, when execution must pause for approval, and how results are written back and audited, then record the model and version as one dependency within that package. This preserves room to replace the underlying model while avoiding the mistake of treating 'the model is callable', 'the feature is GA' and 'the job outcome is validated' as one conclusion.
What to watch next
- When Salesforce's long-horizon runtime expands to jobs beyond Hunter, and whether recovery, approval, cost and outcome records can be exported for independent review.
- Whether financial-adviser and nursing Agents publish errors, human corrections, rejected approvals and time-to-completion measures defined against real job tasks, rather than only connection coverage or cumulative usage figures.
- Whether job-Agent products can establish a portable workflow contract that maps business objects, permissions, actions, human-control points, runtime state and outcome evidence into one versioned record.
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing