Salesforce, Inc. is a U.S. enterprise-software company, the largest vendor of customer-relationship-management (CRM) software, founded in 1999 by Marc Benioff, Parker Harris, Dave Moellenhoff, and Frank Dominguez and headquartered in San Francisco. In AI policy it matters chiefly as a distribution channel: its Agentforce platform embeds AI agents directly into the enterprise systems large organizations already run, making Salesforce one of the principal routes by which agentic AI reaches non-technical businesses. Its chief executive, Marc Benioff, is among the more frequently quoted executives on enterprise AI deployment and its labor effects.
Agentforce and enterprise AI
Agentforce is Salesforce's platform for building and deploying AI agents that act within its CRM and connected business applications — handling customer service, sales, and marketing workflows on top of the company's Data 360 data layer. Salesforce positions the product line as a move from AI as an assistant to AI as an autonomous actor in enterprise workflows, and its 2026 reporting describes AI agents shifting from experimental pilots toward routine operational use, with financial services, retail, and hospitality among the earlier-adopting sectors (Source: cloudwars.com; salesforce.com). Independent coverage notes continued caution among some buyers about cost, reliability, and governance of agentic systems (Source: salesforce.com). As a vendor whose own marketing has at times described AI agents as substitutes for headcount, Salesforce sits within the broader AI labor disruption debate.
Salesforce also faces competitive pressure from AI coding tools at the low end of its market: reporting on July 6, 2026 described small firms, including Atlanta-based Greenleaf Management, replacing Salesforce with custom applications built using Claude, Lovable, and Replit, cutting software costs by 40–80% (Source: theinformation.com). See Anthropic, Replit.
Koa and in-house models
On September 15, 2026, at its Dreamforce conference, Salesforce and NVIDIA jointly announced Koa, which Salesforce describes as its first CRM reasoning model for Agentforce. Koa was built by post-training NVIDIA's open-weight Nemotron 3 Super model on synthetic scenarios modelled on enterprise workflows in more than 14 industries; Salesforce states that no customer data was used in training, that it controls the weights, and that it runs post-training and inference within its own trust boundary. At announcement the model was in use internally, including in an employee agent in Slack, and available to select pilot customers in Agentforce (1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero), with general availability expected in winter 2026 in U.S. regions. The same announcement brought Nemotron-based models and NVIDIA accelerated computing into Missionforce, Salesforce's line for government and regulated organizations, including private clouds and air-gapped networks, with post-trained NVIDIA models for Missionforce Operations stated as available to select customers in October 2026 (Source: salesforce.com).
A technical report by 25 Salesforce Agentforce & AI Research authors, posted to arXiv the day before the announcement, describes training Koa with reinforcement learning on tasks generated from Agent Script specifications, the declarative language in which Agentforce agents are written. It reports that Koa improves on its base model and exceeds GPT-4.1 on Tau2Bench, the Berkeley Function Calling Leaderboard benchmark and a Salesforce CRM benchmark, while "remaining below the strongest frontier models"; on the CRM benchmark it scores below Claude Opus 4.8 and GPT-5.5 (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)). The press release states that Salesforce applied supervised fine-tuning and reinforcement learning, whereas the report states that Koa is built on an RL-only checkpoint. Salesforce's earlier function-calling models, the xLAM-2 series, appear on the Berkeley Function Calling Leaderboard, and the benchmark's authors reported that an earlier xLAM-7B model showed higher perplexity on their crowd-sourced test set than on the older public set (Source: gorilla.cs.berkeley.edu; Source: proceedings.mlr.press).
Pricing for AI agents
Salesforce began in 2026 to let customers choose how they pay for Agentforce, including custom contracts priced on how much the AI grows revenue by helping salespeople close deals or cuts costs by automating customer service. Benioff said on an investor call on August 26, 2026 that "customers want to buy and want to price in different ways," and described the aim as moving beyond per-task billing: "We want to be able to say, 'No, we improve revenue by this much, so give us $2 because we made you $20 or we made you $40.'" Salesforce had said the previous week that sales growth for Agentforce and a data-management service more than tripled year over year, after which its stock rose about 23% (Source: theinformation.com).
The shift is not confined to Salesforce. OpenAI has begun offering some large customers the option of paying only when its AI completes tasks such as handling customer support interactions, according to a person with direct knowledge; an OpenAI spokesperson declined to comment. Cognition is promising enterprise customers up to $10 million in credits if it fails to deliver engineering results worth at least what they pay. Stripe has published guidelines for outcome-based pricing warning that unless attribution rules are explicit, customers may dispute whether an outcome belongs to the vendor (Source: theinformation.com). See Inference Economics and Token Pricing.
Labor positioning
Benioff's public framing of AI's employment effects shifted over 2025–2026. After earlier statements describing Agentforce-driven headcount substitution as a benefit to enterprise customers, on April 27, 2026 he announced that Salesforce would hire 1,000 new graduates during the year and rejected "AI jobpocalypse" rhetoric, emphasizing entry-level hiring rather than cuts (Source: fortune.com). Whether the reversal reflects a durable change in position or a public-relations correction is treated on the Marc Benioff page as dependent on subsequent hiring data and Agentforce deployment metrics.
Relationships
- deploys-in: Financial Services — AI Deployment, Retail — AI Deployment, Legal Services — AI Deployment — Agentforce agents deployed across CRM-using sectors.
- related: Marc Benioff, Agentic AI, AI Labor Disruption, Professional Services AI Adoption
- related: Inference Economics and Token Pricing — outcome-based pricing as an alternative to per-token and per-task billing.
- related: Salesforce Koa — Salesforce's CRM reasoning model, built on NVIDIA Nemotron; Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report) — its technical report
- related: Nvidia & TSMC — AI Compute Infrastructure — base-model provider and co-announcer of Koa; partner for Missionforce
- related: Berkeley Function Calling Leaderboard (BFCL), τ-bench — public tool-use benchmarks on which Salesforce reports Koa's results