Dreamforce 2026 Recap: What Was Announced and What It Means

Dreamforce 2026 was the week Salesforce stopped selling screens. The headline, AIforce, is an interface layer that carries Salesforce data, logic, and permissions to wherever people and agents already work, and around it Salesforce lined up its own CRM reasoning model, Koa, alongside Claude, Gemini, OpenAI, and the Amazon Bedrock range. For a CEO or CTO who was not there, the verdict is simple: very little of it changes your roadmap this quarter, a large share of what looked new was announced months earlier or is still in beta, pilot, or "coming soon", and the decision it forces is not which model to bet on. It is whether your data and permissions are clean enough to be exposed through interfaces you will no longer design yourself.

We spent the week in San Francisco for Dreamforce, and this is the wrap-up we promised in our pre-event questions and our day-one read. It is written for leaders who were not in the room, so it separates what shipped from what was shown, and stage claims from verified facts. Product facts come from Salesforce Newsroom, partner announcements, and Salesforce product pages; keynote quotes are attributed to the outlets that reported them. Everything here is as of 18 September 2026.

What did Salesforce announce at Dreamforce 2026?

The official line was "Become an Agentic Enterprise," and the stack Salesforce drew under it has four layers: Data 360 for data, metadata, and memory; Customer 360 for application logic and what Salesforce calls semantic intelligence; Agentforce for the agents; and, new this year, AIforce as the interface on top. Salesforce listed more than 1,600 sessions ahead of the event. CNBC put attendance at about 50,000, Salesforce's media page says 50,000-plus in person and more than 122,000 watching online, and pre-event estimates from ABC7 and CBS San Francisco were closer to 43,000.

One thing the table cannot show is age. Several items that read as Dreamforce news were announced earlier and demoed this week: Headless 360 in April, Agentforce Contact Center in March, Agentforce in ChatGPT and OpenAI models in Agentforce in October 2025, Gemini in the Agentforce reasoning engine live since October 2025, and Claudeforce itself on 26 August. A demo of a shipped product and the announcement of a pilot deserve very different weight in a planning conversation, and the keynote format does not help you tell them apart.

What does AIforce actually change?

Salesforce describes AIforce as "a live interface layer that brings the full power of Salesforce to wherever people and agents work." In practice it launches inside Claude, Slack, and Coworker, and Salesforce says partner interfaces from Anthropic, AWS, Google, and Microsoft plug in through its Headless Toolkit of MCPs, APIs, plug-ins, and skills. The ability to compose an interface by describing it, rather than having an admin build it, sits at the centre of the story.

The sentence that matters most is quieter. Salesforce says every request runs on existing permissions and business rules, that each agent sees only what the person asking can see, that actions route back through Salesforce, and that there is no new permissions model. It pairs that with Zero Data Retention, so data answers the question at hand and is not kept. That is the right design, and it has a consequence nobody put on a slide: because AIforce reuses your permissions model rather than replacing it, everything wrong with that model becomes visible in more places. Sharing rules that drifted for a decade were tolerable when the only way in was a screen you had to know how to navigate. They are much less tolerable when anyone can ask for anything in plain language.

Marc Benioff framed the shift historically. ITPro reported him describing the interface as alive, tracing an arc from the command line to the GUI, then web and mobile, then the AI interface. He also took on the market's anxiety directly, saying in ITPro's account: "We realize that the SaaSpocalypse isn't about the end of software, but it may be about the end of software that makes humans do all the work." It is a fair line from the company with the largest stake in software surviving.

The open questions are the ones a budget depends on. What will AIforce cost, and in which editions? Which interfaces are generally available rather than in beta? So far Salesforce says only that pricing and packaging are subject to change.

Is the model now just a choice?

This was the week's most important strategic signal, spread across four announcements rather than stated in one.

Koa is Salesforce's first CRM reasoning model, built with NVIDIA by post-training Nemotron 3 Super on a synthetic dataset modelled on close to three decades of CRM deployments. Salesforce says it controls the weights, runs Koa inside its own infrastructure so no customer data crosses the trust boundary during inference, and used no customer data in training. On its own CRM benchmark, Salesforce reports Koa matching or exceeding leading models on CRM actions with three times fewer errors, which is a vendor benchmark and should be read as one. The availability line is precise: select pilot customers now, with "general availability expected winter 2026 in U.S. regions."

Beside Koa sat everyone else. Anthropic's Dario Amodei opened the keynote with Benioff, and Anthropic released Salesforce in Claude in beta on all paid Claude plans. Google Cloud announced that its models "continue to be available to power Prompt Builder and Agentforce's Reasoning Engine," and expanded the partnership with Salesforce and Tableau inside Gemini Enterprise, built on MCP, available now, plus Hyperforce on Google Cloud reaching general availability in North America in November 2026. OpenAI's Sam Altman held a separate afternoon fireside with Benioff; the substance of that relationship, Agentforce in ChatGPT and OpenAI models inside Agentforce, dates from an October 2025 announcement.

AWS brought the most concrete package. Per the joint announcement, Agentforce customers can now use the Amazon Bedrock model range plus Anthropic and NVIDIA models, with OpenAI models through Bedrock still "coming soon." Data 360 zero copy now reaches AWS Glue-managed and S3-backed Apache Iceberg tables, Amazon Aurora, Amazon RDS, and SageMaker Lakehouse. Salesforce context is available inside Amazon Quick, and Agentforce Voice with Amazon Connect Customer is due in fall 2026.

Put together: the cloud and the model became choices, and data and permissions became the strategic layer. That is good news for buyers, because it lowers the cost of being wrong about a model. How are Koa and Claude meant to coexist in one org, and who decides which model an agent uses? Is that a per-agent setting, a per-org policy, or a licensing line? What does running three model providers do to your cost model and your audit trail? None of those were answered this week.

Why is Headless 360 the sleeper?

Because it is the plumbing that makes the rest possible, and it was not even announced at Dreamforce. Salesforce introduced Headless 360 in April under the line "No Browser Required," exposing its capabilities as APIs, MCP tools, and CLI commands, and expanded it in August. The Headless 360 MCP Server is in open beta, the Data 360 MCP Server is generally available, and Salesforce says it serves agents in Agentforce, Claude, ChatGPT, Cursor, and others.

AIforce is the product face of that idea; Headless 360 is how it works. Our explainer on what an MCP server is covers the protocol, and our architects' read on headless 360 architecture covers what it changes for Salesforce design. The practical implication is that integration work, not interface work, is where the next two years of Salesforce effort concentrate.

Agentforce Contact Center has been generally available since March as an add-on for Agentforce Service customers in the U.S. and Canada, and appeared this week in sessions rather than announcements. And Salesforce has completed five acquisitions this year, per its own releases: Momentum and Cimulate in March, m3ter in July, Contentful on 1 September, and Fin, formerly Intercom, on 10 September, now part of Salesforce AI Labs. Fin is also the name of the customer agent now in the lineup.

What do the new agents demand before you switch them on?

Salesforce's job-ready agent portfolio, announced on 11 September and carried onto the stage, names seven: Piper for inbound leads, Hunter for outbound pipeline, Carter for shoppers, Casey and Fin for service and customer experience, Marshall for back-office and supply-chain orchestration, and Paige for employee HR and IT requests. Most are generally available; Hunter is in pilot with general availability stated for November 2026. Multi-Agent Orchestration reached general availability in the same cycle, and Agent Optimizer is due in October.

Two more were shown than documented. ITPro reported the keynote featuring a Marketing Cloud agent that builds campaigns, and Salesforce's product page lists a Campaign Agent reaching general availability in Marketing Cloud Next Advanced by October 2026. ITPro also reported a Field Service agent with voice-to-text, which we could not yet find in Salesforce's own documentation, so treat it as shown rather than shipped. Agentforce Voice, Builder, Operations, and Observability were demoed throughout, and all were already available before the event.

The quieter release was for architects. Salesforce's Well-Architected Framework now has five pillars, Trust, Reliability, Operational Excellence, Resource and Cost Optimization, and Fairness, each with an Agentic Enterprise lens. It is the most useful document of the week for anyone actually deploying agents, precisely because it is not a keynote.

A generally available agent is not a switched-on agent. Each one acts on your data under your permissions, which makes the work before go-live governance work: what the agent may touch, what it may send without a human, and how you will know when it is wrong. Our guide to an AI governance policy covers that shape, and Agentforce pricing covers the other half, since consumption pricing turns agent volume into a budget line. For developers, the thread ran through Agentforce Vibes, which Claude now powers by default, and sessions that paired Vibes with Claude Code.

How should you read the customer stories from the stage?

As claims made by the people telling them, which is how we report them here.

Siemens was the flagship. In a keynote demo, a Salesforce presenter said of Marshall, per ITPro, "Marshall just learned it in 90 minutes," describing the agent learning a supplier-onboarding process and its business rules in an SAP sandbox, against weeks for a new employee. Note what the 90 minutes measures: learning time, not onboarding time. The Siemens release itself is in the future tense: supplier onboarding will go from weeks to days, Piper will build pitches and book qualified partners, and Agentforce Operations will handle the final onboarding steps in SAP.

Adecco's CEO Denis Machuel told the keynote audience that Adecco had placed 20,000 more people this year, according to ITPro's account. That is a year-to-date figure with no stated baseline. Adecco's own release covers an Agentforce Coworker rollout across more than 40 countries, reaching 27,000 employees after pilots in the UK and France.

Live Nation's Venue Agent runs across 120 venue websites, and Salesforce projects it could automate more than 300,000 fan inquiries a year; Live Nation's release with Salesforce states that 95% of Venue Agent questions are answered without a handoff. Salesforce's customer story on SaaStr reports a 72% open rate and a response rate above 10% for an Agentforce outreach agent, figures that predate the event. And Tottenham Hotspur's Ask Spurs fan assistant drew different numbers from different sources: ITPro reported 11 languages, Salesforce's own story eight.

The pattern to watch is consistent. Projections get presented alongside measurements, baselines go missing, and demos speak in the present tense about releases written in the future tense. None of that makes the stories false. It makes them the start of a reference call, not the end of one.

What did the safety debate on one stage settle?

Nothing, which is the useful part. The keynote put three AI chief executives in front of the same audience with visibly different positions. NVIDIA's Jensen Huang, reported by ITPro, said "Safety is an engineering problem," and argued the industry did not need new laws. Anthropic's Dario Amodei, in the same account, called for more investment in safety and for the industry to "set standards for everyone." CNBC reported Sam Altman saying it was imperative for safety and monitoring to come before capabilities. And Benioff, speaking to reporters afterwards rather than on stage, said of AI companies, per CNBC: "And then they should be held accountable."

For a buyer, the lesson is not which of them is right. It is that the vendors supplying your models disagree about how careful to be, so your own governance cannot be outsourced to any of them. The standard you hold agents to has to be your own, written down and enforced in your environment.

What would we do next as a CTO?

Five things, in this order.

Audit data and permissions first. AIforce, Claudeforce, and every agent above inherit your sharing model. Fix drift in profiles, permission sets, and sharing rules before an interface layer starts exposing it to plain-language requests.

Pick one workflow. Choose a single, high-volume process with a clear owner and a measurable outcome, and get one agent working there before anyone proposes a portfolio.

Run evals before granting autonomy. Decide what correct looks like, test against your own data, and keep a human in the loop until the evals say otherwise.

Decide your model policy. With Koa, Claude, Gemini, OpenAI, and Bedrock all on offer, write down who chooses the model for which workload, on what criteria, and where the data may go. That decision is cheaper to make now than to unwind later.

Do not rewrite the UI yet. Composed interfaces are real and early. Keep your Lightning investments working and put the effort into the data and integration work every one of these options depends on.

Where to go from here

If you want the thread that led here, start with Claudeforce explained, then our read of when a Slack dashboard is enough for the Slackforce side of the week. If you are weighing any of this for your own org, Codiot's Agentforce team works through exactly these decisions: the permissions audit, the first workflow, and the evals that decide how far an agent is trusted.