For most of the last two decades we described Salesforce as a system of record for relationship data. Then, as it absorbed more of the working day, as a system of interaction for more of your workflow: the place the work happened, not just where you logged it afterward.
Headless complicates that.
You are still interacting with Salesforce. You're just interfacing with it indirectly, through an MCP or API by way of your agent(s), from inside Slack or Teams or Claude. So technically it is still a system of interaction. But if the interface isn't Salesforce's own anymore, then the interface was never the thing you were paying for.
Which leaves a real question: what is Salesforce for, once you stop looking at it?
It is not the model. Salesforce doesn't offer an enterprise model that surpasses the capabilities of anything you would get directly through OpenAI or through Claude, and Copilot has enough good capabilities that it is difficult to rationalize why you would use one of Salesforce's models instead. Across the rest of your stack, there are plenty of systems that do their one thing better than Salesforce. If the last two years were a capability race, the honest advice to most enterprises would be to buy the best model, wire it up yourself, and leave the CRM as a database. But it's not that simple.
One of the terms I found myself using at Dreamforce this year was this: what does Salesforce's role become in an enterprise organization now? And what's left, once the interface and the model are off the table, is the thing almost nothing else in your stack has. Your narrative.
Salesforce becomes the system that owns what your processes are, why they're relevant, and how they relate to one another, both sequentially and semantically. If there is one critical or differentiating thing Salesforce can offer your organization, it is an understanding of how and why you do work.
What a system of narrative actually knows
It can understand who you are as a persona within your organization. It knows what you mean when you ask questions about your data. It knows what comes next when you log a task: you create an opportunity, you want to advance that opportunity, and what does that look like in your industry, for you, as a sales representative?
A general-purpose AI will answer your question. It won't know whether it answered the right one, because it doesn't know any of that about you beyond what you put in the prompt, and it certainly won't know how to guide you to and through the next logical step in your work.
What Salesforce still has the capability to offer you is the intuition. Helping you do your job better. Completing things in the pipeline, moving work forward, making suggestions proactively about how to move your work forward. Very few other systems within your organization are as well positioned to service that narrative, and to almost complete the sentence for you as the operations user who is talking to Claude.
Your organization spent years teaching Salesforce all of it, mostly without meaning to. That is the asset. Not the storage, not the interface, and not the model.
Where this breaks today
Everywhere I hear people talking about where they are struggling with Agentforce (and soon Claudeforce) adoption, it comes back to the same thing. It is all well and good what you saw on stage. But how do I break up everything I have built into a UI today into bite-sized, AI-consumable processes?
Everybody watches the demo and says: that was beautiful, but my work is just more convoluted than that. They are right.
Take a five-step screen flow you have in your org right now. It can't be consumed in Claude today. How does artificial intelligence know how to expose that experience to you? How many disparate parts does it become? And how does it know that the narrative still needs to execute as if you were in the UI?
Multiply that by years of customizations, permissions, flows and packages, and that is the actual constraint on adoption. Not model quality. Not licensing.
And underneath all of it sits an elephant in the room that is still not being addressed. Data 360 underpins so much of these agentic capabilities, and it is still not being adopted well. Moreover, it is often not implemented well. It is a deeply involved enterprise in and of itself before you even get to AI Data 360 configuration. We have been pitching that narrative in our blogs for months and it is still not surface level in the conversations.
Business leaders are seeing the AI gallery that's on display. They are not being pulled aside to hear that getting there requires this massive trek.
Organizations can see the green valley over there. They can see that it's lush and it's fertile and it's got a gorgeous river running through it. They just can't fly over the canyon. Digital Mass can build the bridge.
The shift your organization has to make
None of this is a licensing decision, and none of it is a release you wait for. It is a change in what your organization treats as the work.
From building the screen to describing the process. For twenty years the craft was making the interface good: the right fields in the right order, the flow that walks somebody through it. All of that assumed a human would be looking at it. When the consumer is an agent, the value moves from the screen to the description of what the screen was for. The flow is the artifact you have. The articulation is the artifact you need.
From tribal knowledge to written narrative. Most of what your organization knows about its own process isn't in your org at all. It's in the person who knows what to do when the order comes in wrong. An agent cannot inherit that. It can only inherit what somebody wrote down, which means every undocumented exception is a piece of your narrative your agents will never have.
From buying capability to curating context. The question used to be which product to license. Now it becomes how much of your own process you are willing to make explicit. That is not a procurement exercise and it is less hands on keyboard than people expect. It is business analyst work intersecting with senior technical resources: pressing rewind on what you already built, retranslating it into business language so the agent understands the workflow, and organizing that menu, if you will, of workflows your agents should be able to consume.
So what to do next? Start with the ten workflows that matter most. Write down what each one actually does, in business language, including the parts that live only in somebody's head. If you can't describe it to a new hire, an agent can't execute it either.
The companies that get the most out of the next two years won't be the ones with the best model. They'll be the ones whose narrative is written down.