Skip to content
Engineering Square
Salesforce

Dreamforce 2026 decoded: Koa, AIforce, and the renames that matter

What Salesforce actually announced at Dreamforce 2026 — the Koa reasoning model, the AIforce interface layer, Salesforce in Claude, and a product-name map — and what an enterprise buyer should do about each.

Published September 30, 2026Read 7 minBy Engineering Square

Dreamforce 2026 (September 15–17) delivered four announcements that change how enterprises should plan their Salesforce roadmap: Koa, Salesforce's first CRM reasoning model, built with NVIDIA and now in pilot; AIforce, an interface layer that puts Salesforce inside Claude, Slack, and every Lightning search bar; Salesforce in Claude, a beta that brings accounts, opportunities, and pipeline into Claude under each user's own permissions; and a round of product renames that reverses much of the 2026 'Agentforce everything' naming. This is a working guide to what each one is, what it costs or changes, and what to do about it — written from the implementation side, not the keynote side.

Koa: a CRM-tuned reasoning model, and a new cost curve

Koa is Salesforce's first reasoning model built specifically for CRM work. It was post-trained with NVIDIA on the Nemotron model family using synthetic CRM data — Salesforce states no customer data was used in training — and it runs inference inside Salesforce's own trust boundary rather than calling out to a third-party API. Salesforce's published claim is that Koa matches or exceeds leading general models on CRM tasks with roughly three times fewer errors. It is in pilot now, with early customers including Formula 1 and UChicago Medicine, and general availability is expected in winter 2026 in US regions.

The architectural idea matters more than the benchmark. Inside Agentforce, an AI gateway routes each piece of work to the cheapest model that can do it reliably: routine, well-scoped CRM actions go to Koa, and frontier models — including Claude — are reserved for genuine judgment calls. If your agents run on consumption pricing, that routing decision is now a budget line. Designing agent workflows so the routine path stays routine is where the savings actually come from.

AIforce: Salesforce where people already work

AIforce is not a product you buy; it is the umbrella for the surfaces that take Salesforce out of its own tabs. It groups four things:

  • Salesforce in Claude (beta, announced September 15) — accounts, opportunities, and pipeline inside Claude, with 37 prebuilt sales skills. Claude signs in as each user, sees only what that user's Salesforce permissions allow, and by default asks for approval before writing any change.
  • Agentforce Coworker (generally available since May 2026) — the AI teammate in every Lightning search bar, successor to Agentforce Assistant and Einstein Copilot.
  • Slackforce — the Salesforce-in-Slack surface, following the same pattern.
  • Headless Toolkit — the developer path to the same capabilities without the Salesforce UI (the architecture formerly marketed as Headless 360).
The strategic read: Salesforce stopped insisting you come to its UI. The system of record stays; the surface becomes wherever the work is.

The renames: what reverted, what stayed

Salesforce dropped the Agentforce prefix from most product names in September 2026, reversing much of the earlier rebrand. If your internal documentation, procurement templates, or statements of work still use the mid-2026 names, this is the correction map:

  • Agentforce Sales → Sales Cloud (restored)
  • Agentforce Service → Service Cloud (restored)
  • Agentforce Marketing → Marketing Cloud (restored as the umbrella)
  • Headless 360 Platform → Salesforce Platform, with the headless architecture now called the Headless Toolkit
  • Data 360 keeps its name (renamed from Data Cloud earlier in 2026)
  • Agentforce 360 keeps its name — the only product retaining the Agentforce prefix, alongside Agentforce as the brand for the agent platform itself

What to actually do

  1. 01Correct the names in your contracts, runbooks, and internal enablement before renewal season — mismatched product names in procurement documents create real friction in legal review.
  2. 02If you run Agentforce on consumption pricing, ask your implementation partner to model the Koa routing split before GA. The economics of an agent designed for gateway routing differ materially from one that sends everything to a frontier model.
  3. 03Treat Salesforce in Claude as a governance decision, not a feature flag. The permission inheritance model is sound, but opt-in is org-wide through an admin request on AgentExchange — pilot it with a defined cohort first.
  4. 04Re-check anything you wrote down about Salesforce's AI stack before September 17. Dreamforce 2026 made a measurable share of it stale — this page included, which is why it carries a dateModified.

We implement Salesforce for enterprises — Agentforce 360, Data 360, and the AppExchange path — and we run agentic systems in production inside our own company, including a company-wide MCP orchestration server. When we say the gateway-routing split changes agent economics, it is because we design for the same split in our own systems.

Have a project like this?Book a working session
Newsletter

Field notes from the bench

Occasional writing on how we build — analytics, launch discipline, AI engineering. No fluff.

No tracking · Unsubscribe anytime · ~Monthly