Salesforce announces AIforce: Koa reasoning agent platform, NVIDIA and Siemens partnerships

At the Dreamforce 2026 opening keynote on September 15, 2026, Salesforce:

1. Launched AIforce, a "live interface layer" that brings the full power of the Salesforce platform (data, workflows, business logic, semantics, permissions, security, governance) to wherever people and agents already work — Claude, Slack, or Lightning (FACT: vendor announcement; Salesforce Newsroom story "Salesforce Unveils the Future of Enterprise Software: AIforce", Sep 15, 2026).
   • Launching surfaces: Claudeforce (Salesforce in Claude — a prebuilt MCP server with 37 prebuilt sales skills, built with Anthropic), Slackforce (Slackbot reasoning over Slack + Salesforce, Slackforce Surfaces live interfaces, Slack CRM, Slack Code), and Agentforce Coworker (an AI teammate inside Lightning).
2. Announced Koa, Salesforce's first CRM reasoning model for Agentforce, built on NVIDIA Nemotron 3 Super and developed in deep technical collaboration with NVIDIA (FACT: Salesforce/NVIDIA press release, Sep 15, 2026). Koa was post-trained on a proprietary synthetic dataset modeled on ~27 years of CRM deployments; Salesforce controls the weights and runs inference inside its own trust boundary.
3. Extended the NVIDIA partnership to Missionforce — Nemotron-based models and accelerated computing for government/regulated organizations, including air-gapped and private-cloud deployment; post-trained NVIDIA models power agents for Missionforce Operations.
4. Deepened the Siemens partnership: Agentforce combined with Siemens Teamcenter SLM to put "engineering-grade answers" into sales, service, and customer workflows; Siemens also detailed its own Agentforce deployment (18,000 sellers, 100% of inbound leads engaged across 132 countries) (FACT: joint Salesforce/Siemens announcement, Sep 15, 2026).
5. Context around the announcements: Benioff dismissed the "SaaSpocalypse" narrative as "crazy nonsense," described AIforce as an "interface revolution," and promised ZDR (zero data retention); Jensen Huang joined Benioff on stage and said he wants to "Agentforce every company." Benioff also claimed 7,000 internal Claudeforce users and a $500M pipeline quarter from the Hunter agent (COMPANY CLAIM).

• Access model for enterprise AI: Salesforce stopped positioning its own UI as the destination. AIforce productizes "headless" access (Headless 360 / MCP servers / APIs) into packaged, switch-on surfaces so CRM work can start in Claude, Slack, or a generated UI — a structural inversion of "log in to Salesforce to do work."
• Model strategy: Salesforce moved from renting frontier model reasoning (Claude/ChatGPT via Agentforce's AI gateway) to owning a domain reasoning model (Koa) it controls end-to-end — weights, post-training, and inference inside its trust boundary. This is the first time Salesforce controls the reasoning engine, versus earlier plans around an in-house "Atlas" reasoning engine that were publicly walked back in favor of the NVIDIA partnership (INTERPRETATION, supported by Salesforce Break's retrospective and TechCrunch's report of Govindarajan's explanation that no sovereign open pre-trained base model existed before Nemotron).
• Industrial/OT extension: The Siemens Teamcenter integration moves autonomous agents beyond software workflows into manufacturing engineering data (digital twins, spare parts, upgrade validation), the first large-scale connection of a CRM agent platform to PLM/digital-twin data.
• Partner ecosystem realignment: Salesforce simultaneously deepened ties with NVIDIA (model ownership), Anthropic (Claudeforce), and expanded AWS/Google Cloud integrations (Bedrock model choice, Hyperforce on Google Cloud GA in North America in November 2026, Gemini Enterprise agents) — a multi-model, multi-cloud posture.

Before (pre-Sep 15, 2026): Salesforce sold the Agentforce platform for building agents, but core multi-step reasoning was routed through third-party frontier models via the agentic AI gateway. Users reached Salesforce value through Salesforce UIs (Lightning), Slack CRM beta features, and manually-assembled MCP integrations. Claude was available in Agentforce (through Amazon Bedrock and Anthropic integrations), and the Claudeforce partnership had been announced on Aug 26/27, 2026. NVIDIA's role was compute + models for other workloads; Siemens' Teamcenter data was separate from front-office workflows.

Change (Sep 15, 2026): AIforce becomes the packaging/positioning layer ("live interface layer") with three launch surfaces; Koa gives Salesforce an owned CRM reasoning model; Missionforce gets post-trained NVIDIA models; Siemens Teamcenter SLM is wired into Agentforce with a named deployment at Siemens (engagement + qualification agents, Piper AI SDR, Slackbot onboarding, Regrello-powered Agentforce Operations).

After (announced trajectory): CRM work can begin in Claude, Slack, or a described-to-be-generated UI, with every action routing back through Salesforce permissions and audit. Enterprises can select Koa as a model provider org-wide or per agent; regulated customers can run NVIDIA Nemotron models in air-gapped/private-cloud Missionforce environments; manufacturers can get engineering-grade answers in seconds in sales/service workflows; and Salesforce plans additional surfaces and pricing (AIforce Max Edition, SDK) after Dreamforce.

• AIforce / Headless 360: AIforce sits atop the "Agentic Enterprise" stack — Data 360 (harmonized/federated data, metadata, memory), Customer 360 (application + semantic intelligence: business logic, processes, permissions, actions), Agentforce (deployable agents + builder) — and now AIforce as the interface layer exposing that foundation through MCP servers, APIs, plug-ins, skills, and CLI tools (the "Headless Toolkit"). Requests are grounded in Salesforce data; every request runs on existing permissions/business rules; every action routes back through Salesforce (FACT per vendor; architecture described in Salesforce Newsroom and diginomica).
• Claudeforce / Salesforce in Claude: A prebuilt MCP server plugin inside Claude with 37 prebuilt sales skills (meeting prep, deal-health review, pipeline review, onboarding dashboards). Auth and permissions are configured once by an admin; Claude sees only what the user is authorized to see; write actions enforce Salesforce business rules; paired with Zero Data Retention + monitoring that flags serious misuse to the customer without Anthropic reviewing underlying data first (COMPANY CLAIM; some specifics e.g. Enterprise Frontier Safeguards per Salesforce Break).
• Slackforce: Slackbot reasons across Slack conversational context plus Salesforce context; Slackforce Surfaces generate live, interactive interfaces (dashboards/decks/calculators) that keep pulling live data; Slack CRM lets users create accounts, log call notes, and update records by prompt; Slack Code is the agentic coding surface.
• Agentforce Coworker: An AI teammate embedded in Lightning that reasons across accounts/activity/history, calls the customer's already-deployed specialized Agentforce agents, inherits existing permissions, and activates with a push button (no migration, no new permissions model).
• Koa model: Post-trained from NVIDIA Nemotron 3 Super (a 120-billion-parameter, open, US-trained model with published training-data provenance) using SFT + reinforcement learning with GRPO, via NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel. Training corpus: synthetic scenarios simulating reasoning, tool use, and decision-making across 14+ industries (manufacturing, financial services, healthcare, travel, etc.) modeled on ~27 years of CRM deployment experience; no customer data used in training (COMPANY CLAIM — consistent across Salesforce and repeated by TechCrunch/SiliconANGLE based on executive briefings). Served at temperature 0 inside a dedicated serving harness with trust/safety controls; selectable as a managed LLM in the Data Cloud generative models catalogue, as an org-wide model provider in Setup, or per agent/sub-agent in Agentforce Builder.
• Koa benchmark (CRM Bench): Salesforce reports Koa matches or exceeds leading models on CRM actions with 3x fewer errors; 11% more precise at calling the right action; 2.1x greater reliability; 15% better at remembering long-conversation context (COMPANY CLAIM — Salesforce's own benchmark; methodology not independently published).
• Siemens integration: Agentforce + Teamcenter SLM connects engineering/digital-twin data to agent workflows: service technicians identify correct spare parts by serial number before site visits; sales reps quote only technically valid/manufacturable upgrades; customers find and order parts without engineering involvement. Siemens runs an engagement agent and a qualification agent in Sales Cloud for inbound leads (public-key-secured lead identification), with Piper (AI SDR) and Slackbot + Agentforce Operations (Regrello) automating partner onboarding.

• Enterprise agent front-end war: AIforce is a direct answer to the "SaaSpocalypse" thesis — that agents will hollow out application UIs. Salesforce's counter-position: the system of record + governance is the moat, and it will meet agents anywhere (Claude, Slack, Gemini Enterprise, Amazon Quick). This reframes the value conversation from seats to governed API/MCP consumption (InformationWeek raised the pricing/predictability question explicitly).
• A software vendor owning a reasoning model: Koa marks a major enterprise-software vendor taking model ownership (weights, post-training, inference) rather than renting frontier APIs — with cost ("tokenomics"), data-provenance, and trust-boundary arguments. TechCrunch framed Koa as "everything the AI labs should fear": an open-weight, domain-tuned, token-efficient alternative to Claude/ChatGPT for CRM work.
• Agents crossing into industrial/OT data: The Siemens Teamcenter integration extends autonomous agents from software workflows into manufacturing engineering — spare parts, quoting validity, digital twins — the "industrial AI operating system" narrative, and a concrete aftermarket economics case (company-cited industry analysis: aftermarket grows ~6x faster with ~4x margin).
• Multi-model pragmatism: Salesforce is simultaneously owning one model (Koa), partnering with Anthropic on the flagship interface (Claudeforce), and deepening AWS/Google Cloud — signaling that the enterprise agent market will be multi-model and governed, not single-vendor.

• Work that previously required logging into Salesforce (or asking engineering) can now be done by prompting in tools employees already use — with the same permissions, business rules, and audit trail.
• Enterprises can run CRM agents on a reasoning model whose weights Salesforce controls and whose inference never leaves the Salesforce trust boundary — relevant to regulated industries (financial services, healthcare, government, accounting).
• Government/regulated organizations can deploy post-trained NVIDIA Nemotron models inside Missionforce on air-gapped networks and private clouds (Missionforce Operations GA now; post-trained models Oct 2026).
• Manufacturers can compress industrial processes "from weeks to hours" by putting engineering-grade answers (spare parts, upgrade validity) directly in front of service techs, sellers, and customers.
• Users can generate their own UIs ("describe it, get a live interface") rather than navigate fixed layouts — the "everyone gets their own UI" claim.

• What does AIforce actually cost per surface, per MCP call, and per seat? (No pricing published; InformationWeek flags CIO predictability concerns.)
• Can Salesforce prove governance parity — validation, approvals, audit — when actions originate in Claude or Slack rather than Lightning?
• Will Koa's benchmark claims survive independent evaluation (e.g., third-party or MLCommons-style testing)?
• How will CRM Bench and "3x fewer errors" hold up in production at scale vs. frontier models on non-CRM reasoning?
• Does Koa's token-efficiency advantage persist as frontier models (Claude, GPT) improve and lower prices?
• When will Slackforce surfaces/skills reach GA, and which Slack/plan tiers include them?
• Will the "drop Agentforce branding" reports (The Information via TNW) materialize — and how does AIforce naming settle vs. Agentforce/Claudeforce/Slackforce?
• How does the Siemens Teamcenter integration generalize outside Siemens — connector availability, data governance, and certification for other PLM ecosystems?

• Immediate (late Sep–Oct 2026): Salesforce in Claude open beta expansion; Slackforce documentation/rollout expectations; Missionforce post-trained NVIDIA models to select customers in October 2026; Koa customer pilots begin (1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, Xero).
• Winter 2026: Koa GA in US regions; hours-to-automate industrial processes maturing with Siemens; Hyperforce on Google Cloud GA in North America in November, phased 2027 expansion; AIforce pricing, AIforce Max Edition, and SDK expected to be detailed post-Dreamforce.
• Watch items: whether "Agentforce" branding is retired per The Information's report; whether AIforce surfaces extend to service/marketing/commerce skills; how OpenAI (ChatGPT Ads/Agents API) and Microsoft (Copilot Studio) counter-position; third-party benchmark attempts on CRM Bench claims.
• Signals to track: Fortune or industry reporting on Koa production metrics; whether other PLM vendors (PTC, Dassault) follow Siemens; enterprise pricing adoption curves for MCP-consumption billing.

Dreamforce 2026 was Salesforce's declaration that the application UI is no longer the product — governed data and business logic are, and agents will meet users wherever they work. The headline trio — AIforce (interface layer), Koa (owned CRM reasoning model on NVIDIA Nemotron), and the Siemens Teamcenter integration (agents reaching into industrial engineering) — turns three trends into one thesis: the enterprise agent economy will be multi-model, headless, and governed, and the system of record plus trust boundary is where the value collects. The weakest link is evidence: benchmark superiority, adoption numbers, and industrial "weeks-to-hours" compression are all vendor claims awaiting independent verification. For the AI-news audience, the story's weight is strategic more than technical today — but the reference architecture (open base model + domain RL + synthetic simulation + trust-boundary serving) is the part engineers should actually study, because it is the template every vertical SaaS vendor is about to copy.

Evidence labels used

• FACT: announcement existence, dating, product names, partnership scope, availability milestones (verified across primary sources and multiple independent outlets).
• COMPANY CLAIM: Koa benchmark results (CRM Bench, 3x fewer errors, 2.1x reliability, 15% context), zero-customer-data training methodology details, adoption metrics (100,000 Coworker users, 7,000 Claudeforce users, $500M Hunter pipeline, 5M Casey conversations, Siemens 100% lead engagement, aftermarket 6x/4x economics), ZDR enforcement.
• INDEPENDENT EVIDENCE: TechCrunch executive interviews (Govindarajan on Qwen provenance; NVIDIA VP on token efficiency), Computer Weekly reporting of pre-briefings and enterprise-product commentary, InformationWeek CIO-facing pricing analysis, UC Today governance-risk analysis, diginomica definitional skepticism.
• INTERPRETATION: the "SaaSpocalypse answer" framing; model-ownership as frontier-lab threat; AIforce as multi-model hedge.
• PREDICTION: Koa GA winter 2026 as scheduled; pricing surfaced within 2–3 quarters; competitors launch headless agent-interface equivalents within 12 months; benchmark claims face third-party scrutiny by mid-2027.