The Signal

Salesforce unveiled AIforce at Dreamforce as a new interface layer designed to make Salesforce's governed enterprise context—data, workflows, business logic, semantics, permissions and governance—available to AI experiences wherever employees and agents work. Rather than forcing users to operate through conventional CRM interfaces, AIforce positions Salesforce as an underlying context and execution layer for multiple AI interfaces. The important enterprise implication is architectural: companies could potentially reuse the permissions, workflow logic and semantic context already maintained inside Salesforce instead of recreating those controls separately for every AI agent or application.

Salesforce and Google Cloud expanded their strategic partnership to enable cross-platform agent reasoning and actions across shared enterprise data and infrastructure. Salesforce workloads will run on Google Cloud through Hyperforce, while Salesforce context and actions can connect with Google's Gemini Enterprise ecosystem. The significance is interoperability: enterprises increasingly need agents operating across CRM, cloud, collaboration and data environments rather than remaining isolated inside one application. The arrangement is designed to reduce bespoke integrations and unnecessary data movement while retaining the underlying enterprise context and controls. Google Cloud independently lists the announcement among its September 15 releases.

Salesforce and AWS expanded their partnership to connect Salesforce data and context with AWS services, bring AWS agents into Slack and support real-time bidirectional voice interoperability between Agentforce Voice and Amazon Connect Customer using Agent2Agent capabilities. The architecture is intended to let enterprises connect workflows across platforms without migrating all underlying data between systems. This moves enterprise agents beyond individual applications toward multi-system execution spanning CRM, employee collaboration, cloud services and customer-service voice environments.

Salesforce and NVIDIA announced Koa, Salesforce's first CRM-specific reasoning model for Agentforce. Koa was created by post-training NVIDIA Nemotron 3 Super using a proprietary synthetic dataset modeled on approximately 27 years of Salesforce CRM experience. It is designed to reason through multi-step CRM tasks and determine which tools should be used to complete work. The development is an important example of enterprise AI moving from general-purpose frontier models toward specialized reasoning systems incorporating domain structures, workflow patterns and institutional expertise.

Adecco announced a global rollout of Salesforce Agentforce Coworker across more than 40 countries following pilots in the UK and France. The Claude-powered AI teammate supports recruitment, sales and client/candidate engagement. Adecco said agentic AI was already deployed across recruitment workflows in 10 countries representing approximately 50% of its business revenue. The rollout is particularly significant because it moves beyond experimental access to AI toward multinational production workflows inside a labor-intensive global enterprise.

Siemens and Salesforce expanded their AI partnership by connecting Agentforce with Siemens Teamcenter Service Lifecycle Management. The integration makes engineering-grade product and service information available within sales, service and customer workflows. Siemens also said Agentforce qualifies every inbound lead for approximately 18,000 sellers. The development illustrates a strategically important enterprise-AI pattern: agents are becoming more useful when commercial workflows can access specialized engineering and product context rather than relying only on generic CRM information.

The U.S. Transportation Security Administration reported production results from Ace, an AI agent built with Salesforce Public Sector Solutions. Ace is handling approximately 100,000 routine traveler conversations per month, with 96% of routine inquiries resolved without human escalation. The agent addresses questions involving liquids, medical devices and checkpoint procedures in an operating environment supporting nearly 3 million airline passengers per day. The deployment is valuable evidence of enterprise AI producing a measurable operating outcome rather than merely generating user engagement.

Live Nation expanded Salesforce Agentforce to provide 24/7 fan support across U.S. concert venues after initially piloting its Melody AI agent at BottleRock Napa Valley. Melody supplied event-specific information to festival attendees. The broader deployment demonstrates a repeatable enterprise scaling pattern: begin with a bounded, context-rich workflow, validate performance and operating controls, then extend the agent across a larger footprint using reusable knowledge and escalation processes.

Cohere and Aleph Alpha signed a definitive business-combination agreement creating a sovereign enterprise-AI company operating under the Cohere brand, with dual headquarters in Toronto and Berlin and continued research operations in Heidelberg. The combined organization will have more than 1,000 employees, while Schwarz Group committed €500 million in structured financing. Reuters reported that the combination had been valued at approximately $20 billion when initially disclosed and that Schwarz's StackIT infrastructure plans include a German data-center campus capable of accommodating up to 100,000 AI chips. The transaction remains subject to regulatory approval.

OpenText and Cohere announced a partnership aimed at governments and regulated industries. OpenText provides the enterprise context layer spanning unstructured, operational and transactional information, while Cohere contributes its North agent platform and enterprise models. North supports private deployment, making the architecture particularly relevant for organizations that cannot send sensitive information to conventional public AI environments. The partnership reinforces the emerging separation between the enterprise context layer and the agent/model layer, with governance and proprietary information increasingly determining practical AI value.

Anthropic and Accenture established an embedded-evaluation partnership under which Faculty, Accenture's specialist AI business, will work alongside Anthropic teams on frontier-model evaluation, red teaming, alignment assessments and safeguard testing. Each company expects to invest at least $1 billion over five years, creating a combined commitment of at least $2 billion. The development indicates that model assurance is evolving from periodic external testing into a substantial continuous operating capability—a trend with implications for enterprise procurement, model risk management and regulatory assurance.

Novo Nordisk and Anthropic announced a collaboration to use Anthropic models and Claude Science across pharmaceutical R&D and agentic software engineering. Novo researchers and computational teams will work with Anthropic to identify scientific problems and develop targeted AI solutions for biological reasoning and specific research workflows. The partnership demonstrates the movement of frontier models from generic employee productivity into high-value, domain-intensive scientific processes where proprietary research context and expert validation are essential.

Abu Dhabi-based M42 announced plans to adopt Oracle Health Data Intelligence alongside Oracle Health Foundation EHR to integrate clinical, genomic and real-world data across its healthcare network. Target applications include chronic-disease management, population health and precision medicine. The development is particularly relevant to GCC enterprise AI because it addresses a prerequisite for healthcare AI adoption: governed, longitudinal clinical context must exist before advanced models and agents can reliably support patient and operational workflows.

SAP made Prior Labs' TabPFN-3.5 Plus available through SAP AI Core. The model is optimized for structured/tabular business information rather than natural-language generation and can generate predictions without conventional per-use-case model training. SAP identified cash-flow forecasting, payment delays, supplier-risk scoring, upsell opportunities and churn as representative applications. The release demonstrates why enterprise model strategies are likely to become portfolios: many high-value ERP and finance problems are better suited to specialized predictive models than general-purpose LLMs.

AWS made Moonshot AI's Kimi K3 available through Amazon Bedrock for coding, knowledge work and extended agent workflows. AWS describes Kimi K3 as a 2.8-trillion-parameter open-weight model with native vision and a 1-million-token context window. It is also the first open-weight Bedrock model with explicit prompt caching, potentially reducing latency and input cost where context is repeatedly reused. Moonshot reports approximately 2.5× scaling-efficiency improvement over Kimi K2. Bedrock availability puts the model inside AWS's enterprise access-control, encryption and auditing framework.

AWS released a new generation of Amazon Bedrock AgentCore Runtime featuring elastic memory management and substantially more consistent cold-start performance. AWS reports P75 cold starts of approximately 1.9–2.0 seconds across container images ranging from 200 MB to 2 GB, compared with 5.4–30 seconds for the previous runtime. The platform retains scale-to-zero operation and hardware-enforced session isolation. The development matters because long-running autonomous agents create infrastructure requirements and economics that differ materially from short conversational AI sessions.

AWS introduced SageMaker HyperPod Inference Gateway, a Kubernetes-native routing layer that uses real-time GPU conditions—including loaded model adapters and cache state—to determine where inference requests should be routed. AWS says the gateway can deliver up to an 82% reduction in first-token latency without requiring changes to client applications or model servers. The release demonstrates that enterprise AI economics are increasingly determined below the model layer by routing, caching, accelerator scheduling and utilization engineering.

Enterprise coding-agent company Factory raised $200 million at a $5 billion valuation, taking total funding above $400 million. Factory says its platform serves hundreds of thousands of developers and identifies NVIDIA, Blackstone, RBC, Palo Alto Networks, Adobe and T-Mobile among enterprise customers. Its platform is model-agnostic and supports cloud, on-premises and air-gapped deployment. Factory also reports that automatic task-level model routing has reduced token spending by more than 60%. The financing illustrates investor interest shifting from coding copilots toward governed autonomous software-engineering platforms.

Dutch semiconductor company Axelera AI launched Europa, its second-generation AI-processing architecture, extending the company's edge-AI technology toward enterprise and data-center inference. Systems have been validated with Dell and Supermicro. Axelera reports more than 600 customers and a sales pipeline exceeding $1.5 billion. Reuters reported that the company has raised more than $450 million and already has AI-factory supply contracts worth tens of millions of dollars. The launch adds another potential alternative to the dominant accelerator ecosystem for enterprises focused on inference economics and deployment control.

Delos Data raised more than $100 million to develop its Nonstop AI networking architecture for increasingly heterogeneous AI clusters. Its Data Interface targets 10× lower latency and 10× higher efficiency, while the company projects approximately 300× growth in AI inference demand by 2030. The company argues that communication among GPUs, XPUs, CPUs, memory and storage is becoming an increasingly significant bottleneck as infrastructure moves from model training toward sustained agent and inference workloads.

OpenAI's September 17 Enterprise/Edu update introduced several material changes to enterprise deployment. Tenant-wide SCIM now supports the API Platform, allowing administrators to assign synchronized identity-provider groups to API organizations and selected projects. OpenAI also made ChatGPT for Word available to Enterprise and Edu users. More strategically, affected Enterprise workspaces began the planned migration path from custom GPTs toward Plugins: migration availability was targeted for September 17, creation of new GPTs is planned to end September 25, and existing custom GPTs are scheduled for retirement on December 11, 2026. These changes affect enterprise AI workflow architecture, identity provisioning, administration and migration planning.

HCLTech launched HCLTech Pulse, a dedicated business unit aimed at enterprises generating between $500 million and $5 billion in annual revenue. The unit combines AI strategy, data, cybersecurity, platforms, engineering and business-process transformation into an integrated transformation agenda. HCLTech says the target organizations often have enterprise-scale ambitions but historically have needed to combine multiple specialist providers across strategy, implementation and operations. The launch is particularly relevant to the enterprise transformation market because it packages AI together with the technology foundations and process redesign required to move beyond isolated pilots.

Microsoft published evidence from hundreds of internal AI transformation initiatives, arguing that tool deployment alone did not create operating transformation. Microsoft said an AI tool had been licensed to more than 200,000 people, but broad access alone did not deliver the desired impact. After redesigning workflows around specific outcomes, one sales group recorded 9.4% higher revenue per account manager, 20% higher close rates and a 3× increase in adoption of priority AI use cases. Selected supply-chain workflows cut cycle time by up to 75%, while Microsoft deployed more than 100 purpose-built agents across planning, sourcing, fulfillment and logistics. A nine-person engineering team also shipped an initial product release in 35 days. These are unusually useful production metrics because they distinguish AI access from process redesign and measurable business performance.

Vitality and Google expanded Vitality AI into the United States for health plans and employers. Built on Google Cloud and Gemini Enterprise, the system combines Gemini models with Vitality data spanning more than 2,800 health and behavioral dimensions. Vitality reports 99.2% accuracy based on physician review of personalized insights. Members receiving personalized AI recommendations were 3.3× more likely to complete a mental-wellbeing assessment, 2.7× more likely to complete an online health review, and 1.4× more likely to complete colorectal-cancer screening. The deployment is significant because it ties enterprise AI to measurable behavioral and clinical engagement outcomes rather than merely conversational usage.

Google Cloud inaugurated its Singapore Engineering Center, its flagship product-development hub in Southeast Asia, co-located with the region's first Google DeepMind research lab. The center spans AI, AI infrastructure, data, compute, machine learning, networking and storage, with a mandate that includes next-generation agentic cloud infrastructure, moving frontier models into enterprise systems and building autonomous agent-orchestration tooling. Early work includes real-time multilingual AI with Grab and financial agent workflows with DBS. Google is also expanding its regional Forward Deployed Engineer workforce to help customers move implementations from pilot into production. Google Cloud says customers in more than 200 countries and territories use its platform.