I told a few friends before Dreamforce that the big announcement this year would basically be a change in the shape of the CRM as we have known it since 1999.

I was half joking and mostly guessing, of course, but after Day One I am not so sure it was really a joke.

There were plenty of announcements, as there always are at Dreamforce. AIforce was introduced as a new interface layer for the Salesforce platform, the partnerships with AWS and Google Cloud were expanded, we saw new examples of agents working across enterprise systems, and companies such as Siemens and Adecco were used to show how Agentforce is moving from isolated experiments to large-scale adoption.

All of that is interesting, but what caught my attention was not the individual list of features. It was the direction they collectively point to.

For most of its history, Salesforce has been something you entered. You logged into Salesforce, opened a Lightning application, navigated to a record, clicked a button, launched a Flow, updated an Opportunity or worked inside the Service Console. Even as the platform grew enormously in complexity, the user interface remained the natural boundary of the product. Salesforce was both the underlying platform and the place where you interacted with that platform.

What Dreamforce 2026 is starting to make much more explicit is that those two things no longer need to coincide.

AIforce is probably the clearest signal of this change. The idea is that Salesforce data, metadata, permissions, workflows, business logic and actions should be available to AI interfaces regardless of where those interfaces live. That might be Lightning, but it might also be Slack, Claude, Gemini, Amazon’s AI experiences or something that does not exist yet. MCP, APIs, plugins and skills increasingly become the mechanisms through which Salesforce exposes its capabilities, while Salesforce itself remains responsible for the things enterprises care about most: identity, permissions, trusted data, semantics, governance and execution.

To me, this is the real meaning of the headless strategy.

The interesting idea is not simply that Salesforce can integrate with more AI products. Salesforce has integrated with external systems for decades. The important change is that the user interface itself is becoming optional. You should not necessarily have to go to Salesforce anymore; Salesforce should be able to come to wherever you are already working.

The announcements involving AWS and Google Cloud reinforce this direction. Salesforce context can increasingly travel across ecosystems, while Data 360 continues pushing in the opposite direction at the data layer: the information itself does not always need to be copied into Salesforce either. Both trends point toward a platform whose value is progressively less tied to where the screen is rendered.

I think this will also change the way we design Salesforce solutions. For years, one of the most natural architectural questions has been: what Visualforce or Lightning page should we build, and how should the user navigate through this process? In an increasingly agentic and headless platform, the more interesting question may become: what governed business capabilities, context and actions should we expose, and which human or machine interface should be allowed to use them?

That is a very different mental and operating model.

There is another part of the current Salesforce story that I think deserves some clarification, because the terminology around Agentforce can easily blur concepts that are actually becoming quite distinct: Agentforce, agents and Agentforce Coworker are not the same thing.

The way I currently think about it is that Agentforce is becoming the runtime and orchestration environment for the digital workforce. It is where companies define, configure, govern and execute agents, giving them access to Salesforce data, actions, Flow, Apex, APIs and external systems. Inside that environment, individual agents are the specialized workers. One may qualify leads, another may support customer service, another may prepare an account brief, another may work on pricing, contracts or some very specific industry process. Each can have its own instructions, tools, subagents and actions.

Coworker plays a different role.

Agentforce Coworker is the AI teammate presented directly to the human user. Instead of asking the user to know which specific agent they need, Coworker can become the conversational entry point to the whole system. A user might simply ask, “I have a meeting with ACME in twenty minutes; tell me what I need to know, check the open opportunities, identify any service issues and suggest what I should discuss.” Coworker can reason across the available context, perform actions itself and, crucially, invoke specialized Agentforce agents that the organization has already built.

That makes Coworker, at least from a product perspective, something more interesting than just another agent in the list. I see it more as the colleague through which a human can access and coordinate the digital workforce underneath. Agentforce is the environment in which that workforce exists, the individual agents are its specialists, and Coworker becomes the human-facing layer that can bring those capabilities together.

This distinction becomes even more important when combined with AIforce, because Coworker itself is not necessarily the end of the story. The same capabilities can increasingly surface in multiple interfaces. Lightning may host the experience, but so might Slack or another AI environment. The interaction layer becomes fluid, while the underlying business context and execution remain governed by Salesforce.

If you are questioning “I remember Einstein Cowork, I believe it’s the same thing“, I’m in your team, this is still a bit blurry but only because we have heard most of all marketing messages, this distinction I’m sure will become clearer in the next days/weeks, as the features and capabilities will be revelead.

Then there is Koa, which for me was probably the most interesting announcement of Day One.

Salesforce and NVIDIA introduced Koa as Salesforce’s first CRM reasoning model, built on NVIDIA Nemotron and post-trained using synthetic scenarios designed around Salesforce business processes. I do not think the interesting part is that Salesforce has created yet another model. The market certainly does not suffer from a shortage of models, and Salesforce itself is very clearly maintaining a multi-model strategy.

What matters is that Salesforce is now trying to own a reasoning layer optimized for the domain in which it has an almost unique structural advantage: CRM and enterprise workflows.

For nearly three decades, Salesforce has accumulated an extraordinary amount of knowledge about how companies model customers, opportunities, service cases, contracts, quotes, approvals, entitlements, workflows and industry-specific processes. Until now, Salesforce essentially owned everything around the reasoning model. It owned the data, the metadata, the permission model, the workflow engine and the actions that could be executed, but the most sophisticated reasoning was still largely delegated to external frontier models.

Koa changes that balance. With Koa, Salesforce controls the weights, the post-training and the inference of a model specifically designed to reason about CRM tasks inside its own trust boundary. Salesforce has published internal benchmark results showing improvements in areas such as action selection, context retrieval and reliability compared with general-purpose models. Those numbers should obviously be treated for what they are — vendor benchmarks that still need broader independent validation — but I do not think the benchmark race is actually the most interesting aspect of Koa. The architecture is.

Salesforce does not need to build the smartest general-purpose model in the world. A CRM reasoning model does not need to write the best novel, solve every mathematical problem or generate an operating system from scratch. It needs to reliably understand situations such as: this customer has this entitlement, this opportunity is in this stage, this policy requires this approval, this user has these permissions, and therefore these actions must happen in this sequence.

That is a much narrower problem, but in an enterprise context it can also be an extremely valuable one.

A specialized model that deeply understands enterprise processes could potentially offer better latency, more predictable economics, stronger governance and fewer errors for many CRM workloads than simply sending every problem to the largest frontier model available. Koa does not need to replace Claude, Gemini, GPT or any other general-purpose model. It simply needs to reach the point where, for a growing class of enterprise problems, Salesforce can decide that it does not need to rely on them.

That is why I think Koa may eventually turn out to be the most important announcement of Dreamforce 2026.

When I put all of these pieces together, I see an architecture starting to emerge that is quite different from the Salesforce we have traditionally known. Data 360 provides access to enterprise data and context. Customer 360 provides the business model, the metadata, the relationships, the permissions and the processes. Koa and other models provide reasoning. Agentforce provides the runtime in which specialized agents operate. Coworker provides a human-facing colleague capable of working across that digital workforce. AIforce then makes those same capabilities available through an increasingly open set of interfaces.

Lightning, Slack, Claude, Gemini or whatever comes next can all become different windows into the same underlying system.

This is obviously my interpretation of where Salesforce is heading, not an official Salesforce architecture diagram, but I think the direction is becoming increasingly visible.

For years, we have thought about Salesforce primarily as a collection of business applications: Sales Cloud, Service Cloud, Commerce Cloud, the Industry Clouds and everything around them, all sharing the same platform. My impression is that the Salesforce of the next few years may progressively become something different. The real product could increasingly be the invisible layer underneath those applications: enterprise data, semantics, identity, permissions, reasoning, agent orchestration and trusted actions.

The screen through which you access all of this may become secondary.

Sometimes it will still be Lightning, of course. Sometimes it will be Slack. Sometimes it may be Claude, Gemini or some future AI workspace. In other cases, the interaction may happen entirely between agents, without a conventional interface at all.

This is why I think the headless strategy is one of the most important signals coming out of Dreamforce this year, and why Koa matters far more than simply adding another model name to the Salesforce ecosystem.

AIforce suggests that Salesforce no longer wants the Salesforce interface to be the necessary boundary of Salesforce. Koa suggests that Salesforce may not want someone else’s intelligence to be a necessary dependency either.

If those two trajectories continue, Salesforce is not simply adding AI features to the CRM we have known since 1999.

It is changing the shape of the CRM itself.

And, more broadly, it may be changing the shape of Salesforce from an application company into something much closer to an enterprise AI runtime: a place where business data, semantics, permissions, reasoning, agents and actions are coordinated, while the actual user experience can live almost anywhere.

Maybe telling my friends that Salesforce was about to “change shape” was not such a bad guess after all.

And while Salesforce is busy changing shape, I have also been working on a little surprise for ORGanizer for Salesforce.

The timing, I have to say, feels pretty good.

More on that soon.

Want to see the whole thing in context? Watch the Dreamforce 2026 Day One keynote here: Dreamforce 2026 Day One Keynote