Dreamforce 2026: Specialized Models and Agents Where Work Already Happens

Dreamforce 2026: Specialized Models and Agents Where Work Already Happens

2026-09-15

The interface war is over. Context won.

For two years, enterprise AI demos lived in a chat box that knew nothing about your pipeline or your approvals. Dreamforce 2026 made that look old.

On September 15, Salesforce announced moves that all point the same way: put specialized models and agents where people already work—Slack, Claude, Lightning, Amazon Quick—grounded in CRM context, permissions, and a trust boundary you can explain to a regulator.

For founders: vendors are racing to put agents in the tools your team opens every morning. You still need judgment on data boundaries, model choice, and what stays in-house.

Koa: a CRM reasoning model, not another generic brain

Salesforce and NVIDIA announced Koa, Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron.

Salesforce says Koa was post-trained on Nemotron 3 Super with a proprietary synthetic dataset modeled on nearly three decades of CRM deployments—processes, workflows, operational policies. On Salesforce’s CRM benchmark (tasks like updating an opportunity, routing a case, scheduling a follow-up), the company says Koa matches or exceeds leading model performance on CRM actions with three times fewer errors.

Trust details worth keeping:

  • Salesforce states no customer data was used to train Koa; training used synthetic scenarios across more than 14 industries.
  • Salesforce controls the model weights and runs post-training and inference inside its own trust boundary.

Koa is in select customer pilots now, with general availability expected winter 2026 in U.S. regions. For government and regulated orgs, the same announcement brings NVIDIA models and accelerated computing into Missionforce—private clouds, air-gapped networks, and environments that never touch public infrastructure. Missionforce Operations is generally available now in U.S. regions; post-trained NVIDIA models for select customers arrive October 2026.

Translation: generic frontier models are table stakes; domain-tuned reasoning inside a trust boundary is the enterprise product.

AIforce: Salesforce comes to Claude, Slack, and Coworker

If Koa is the specialty brain, AIforce is the interface move—a live layer that brings Salesforce data, workflows, business logic, permissions, and governance to any AI interface. People do not have to open Salesforce; Salesforce comes to them. Salesforce also positions Zero Data Retention so business data answers the question at hand and is not retained by the model provider.

AIforce launches with three surfaces:

  • Claudeforce — Salesforce in Claude via a prebuilt MCP server; 37 prebuilt sales skills at launch; Development plug-in for Claude Code with 40+ skills. Available to customers in beta.
  • Slackforce — live Salesforce context in Slack, Slackbot with governed action, Slack CRM updates from prompts, and Slack Code for multiplayer agentic development.
  • Agentforce Coworker — an AI teammate inside Lightning under existing permissions. Salesforce says 100,000 users activated Coworker within its first 35 days.

Marc Benioff called it an “interface revolution”: model intelligence plus the context customers already built into Salesforce. For operators, the useful read is simpler. The battleground moved from “best chatbot” to “best place to inject trusted business context.”

AWS + Salesforce: agents in the daily toolkit

The AWS–Salesforce expansion repeats the pattern: CRM context and agents inside tools teams already use, without forcing a migration.

Operational highlights:

  • Salesforce data and skills natively in Amazon Quick (available now).
  • AWS frontier agents in Slack, starting with AWS DevOps Agent (now); Security, FinOps, and Partner Central agents planned for Fall 2026.
  • Agentforce model choice through Amazon Bedrock, with Anthropic and NVIDIA models called out and OpenAI models coming soon—plus Bedrock Zero Data Retention and Salesforce’s Trust Layer. Salesforce also cites HIPAA, PCI, SOC 2, and ISO 42001 compliance and audit trails for regulated industries.
  • Data 360 zero-copy expansion to more AWS data services so agents ground on enterprise data without duplication.
  • Agentforce Voice with Amazon Connect Customer via Agent2Agent (A2A)—Fall 2026.

The majors are competing on where the agent lives and which model you pick without moving the lake.

What to decide before pilots multiply

Draw the data boundary on paper. Which systems may leave your trust boundary? Which stay Salesforce-hosted, AWS-hosted, or air-gapped (Missionforce-style)?

Choose models by job, not brand. Koa’s CRM specialization is interesting for Agentforce workflows. Bedrock choice matters when a use case needs a different frontier model.

Prefer surfaces your team already lives in—then govern them. Slack + Claude + Lightning drops adoption friction and creates shadow workflows overnight. Permissions must travel with the person and the agent.

Keep something in-house: agent inventory, approval gates for writes, evaluation harnesses, a kill switch. Platforms provide rails. You still need a driver.

Pilot with kill-criteria. One workflow, clear success metrics, a rollback—then expand.

Closing

Dreamforce’s September 15 story is coherent: specialized CRM reasoning (Koa), Salesforce context in the interfaces people already use (AIforce), and deeper AWS plumbing for model choice and zero-copy data. Regulated teams get a Missionforce path; everyone else gets agents in Slack and Claude whether they planned for it or not.

What does not change is the founder’s job. Decide where data may go. Decide which model fits which workflow. Decide what stays under your roof.

Yellow Coop helps with that call—fractional CTO judgment, AI solutions sized to your stack, and tech projects that respect data boundaries. The interface layer got interesting. Your operating model still has to be boring enough to trust.

Sources

  1. Announcing Koa: Salesforce’s First CRM Reasoning Model, Built on NVIDIA Nemotron — Salesforce, Sep 15, 2026
  2. Salesforce Unveils AIforce — Salesforce, Sep 15, 2026
  3. AWS and Salesforce Expand Collaboration for Enterprise AI — Salesforce, Sep 15, 2026