Dreamforce 2026 Day One: AIforce, Koa, and What Changes
The interface itself was the announcement. AIforce, unveiled at the 15 September keynote, is a live interface layer that puts Salesforce data, logic, and permissions wherever people and agents already work, with a new CRM reasoning model called Koa underneath and Claudeforce alongside it. The honest thesis: for most customers nothing changes this quarter. What changes is a decision you now have to make deliberately, which is how much of your Salesforce experience you are willing to let AI generate on the fly, and on whose data it runs when it does.
We are at Dreamforce this week, and this is the day-one read written the morning after the keynote. The fuller takeaways come after the event, once the sessions and the documentation have caught up with the stage. Product facts come from Salesforce Newsroom and keynote framing from the outlets that covered it live, as published on 15 and 16 September 2026.
What did Salesforce announce on day one?
AIforce is the headline, described by Salesforce as "a live interface layer that brings the full power of Salesforce to wherever people and agents work." Salesforce places it on top of an architecture it calls the Agentic Enterprise: Data 360 for harmonized and federated data, Customer 360 for business logic and permissions, Agentforce for the agents themselves, and AIforce as the interface over all of it. The stated point is that people can ask questions, update records, and trigger workflows without opening the traditional UI, inside Slack, Claude, or Coworker.
Koa is Salesforce's first CRM reasoning model, built with NVIDIA by post-training Nemotron 3 Super on a proprietary synthetic dataset modelled on close to three decades of CRM deployments. Salesforce says it controls the weights and runs Koa inside its own infrastructure, that no customer data crosses the trust boundary during inference, and that no customer data was used in training. On its own CRM benchmark, Salesforce reports Koa matching or exceeding leading models on CRM actions with three times fewer errors. That is a vendor benchmark, and worth reading as one.
Claudeforce appeared on the main stage with Anthropic's Dario Amodei alongside Marc Benioff, and the AIforce announcement states that Salesforce in Claude is "available to all customers in beta." Slack was positioned as central to this, with Slackforce as its agent-focused refresh.
On framing, TechRadar reported Benioff pushing back on the idea that AI ends software: "This is not about the end of software, but it may be about the end of software that makes humans do all the work." NVIDIA's Jensen Huang, per the same coverage, called the end of software "nonsense." Take both as what they are, which is two vendors with a considerable interest in software continuing to exist.
What is AIforce, in practical terms?
Strip the branding and AIforce is a governed access layer with a rendering trick on top. The governed part is the substance: Salesforce says every request runs on existing permissions and business rules, with Zero Data Retention, so data answers the question at hand and is not retained. The rendering part is what gets demoed, which is describing an interface in words and having it composed live rather than built by an admin.
The mechanism underneath is worth understanding, because it is neither new nor Salesforce-specific. Salesforce points to an open Headless Toolkit of MCPs, APIs, plug-ins, and skills. That is the same structural shift our architects wrote about in headless 360 architecture, and the same one behind Claudeforce: the system of record stops being a place you visit and becomes something that renders wherever you are, reached through a described interface rather than a built one.
If you have been treating MCP as a curiosity, this is the week it stopped being one. The practical consequence is that your permissions model and your business logic become the product surface, because they are what everything else now runs through. Estates where permissions have drifted for a decade are about to find that out in public.
What is Koa, and does a CRM-specific model matter?
The case for a domain-specific model is reasonable on its face. A model post-trained on how CRM workflows actually run should handle multistep tool use in that domain better than a general model reasoning from scratch, and Salesforce running the weights on its own infrastructure is a genuine answer to the data-residency objection that stalls a lot of AI projects.
What was not answered is most of what a buyer needs. Which clouds and editions get Koa, and at what cost? How does it sit next to using Claude through Claudeforce, given both were on the same stage on the same day, and is that a choice customers make per agent, per org, or not at all? What does the benchmark measure, and does three times fewer errors on Salesforce's own CRM benchmark survive contact with your data model and your unusual objects? Salesforce named pilot customers including Formula 1, UChicago Medicine, and Xero, which tells you it is real and early, not that it is ready for your estate.
Those are open questions rather than criticisms; the answers usually arrive in the release notes.
What the new agents mean for sales, marketing, commerce, and service teams
The job-ready agent portfolio was announced on 11 September, just ahead of the event, and carried onto the keynote stage. The named agents and their stated availability: Piper works inbound leads across websites and inboxes for B2B teams, generally available now. Hunter runs outbound pipeline from research through outreach over weeks, in pilot now with general availability stated for November 2026. Carter helps shoppers discover and compare products with in-chat checkout, generally available. Casey resolves service issues across voice, SMS, WhatsApp, and web chat, and Fin handles more complex customer experience workflows, both generally available. Marshall orchestrates back-office and supply-chain processes with deterministic execution, generally available.
Siemens is the flagship deployment, and its announcement with Salesforce describes Piper as an AI SDR agent building a personalised pitch on the spot and booking qualified partners with a partner account manager, with partner onboarding compressing from weeks to days. That is Siemens' and Salesforce's characterisation of their own project, not a result we have verified. The same applies to Adecco, whose CEO Denis Machuel joined Benioff on stage as the company announced an Agentforce Coworker rollout across more than 40 countries.
A generally available agent is not a switched-on agent. Each of these acts on your data under your permissions, which means the work before go-live is governance work: what the agent may touch, what it may send without a human, and how you will know when it is wrong. Our AI governance policy guide covers the shape of that, and Agentforce pricing covers the other half, which is that consumption-priced agents make volume a budget question rather than a technical one. Run the evals before the rollout, not after the first bad send.
What we are watching for the rest of the week
Four things. Whether the Claudeforce beta, now stated as available to all customers, behaves like a beta or a preview once we have hands on it. Whether Slackforce and Surfaces hold up in a real workspace, which we started on in our read of when a Slack dashboard is enough. What AIforce actually costs and which editions get it, since the announcement carries the usual note that pricing and packaging are subject to change. And how Koa and Claude are meant to coexist, which we will ask in every session that will take it.
The Benioff and Altman fireside ran on Tuesday afternoon. We are not going to characterise it from second-hand notes, so it goes into the full takeaways once we have reviewed it properly.
What to do with all this
Nothing this quarter, for most teams, and that is a legitimate answer. Two exceptions are worth acting on now: fix a drifted permissions model before an interface layer starts exposing it in new places, and if you are already piloting agents, get evals and governance in place ahead of the next wave.
We are on the ground at Dreamforce through the end of the week, and we will publish the full takeaways after the event, measured against the questions we came in with. If you want help working out what any of this means for your own estate rather than in the abstract, Codiot's Salesforce development and Agentforce teams do exactly that.
