Dreamforce, one week later - how Salesforce gave up the screen

This year Salesforce used Dreamforce to give up the screen and double down on everything underneath it, which makes sense given where enterprise applications are going. Salesforce is betting that value is shifting from the user interface to the metadata, and that models will commoditize while context, governance, and orchestration become more valuable. At the same time, the growing roster of pre-built agents also signals a pragmatic shift from 'anyone can build an agent' to delivering more test, production-ready AI. Giving up the screen Among other announcements, the big headline was AIforce, a new user interface layer for accessing Salesforce that enables users to access Salesforce via Anthropic Claude, Slack, Agentforce Coworker, or any other conversational AI interface. This means users can access all the data workflows, business logic, and semantics in Salesforce from their preferred user interface (UI) while preserving Salesforce permissions, security, and governance. In practical terms, Salesforce is saying users shouldn’t have to open Salesforce to use Salesforce – which most users will welcome. Snarkiness aside, this reflects where enterprise software is headed: As conversational AI becomes a primary interface, the value of traditional user interface (UI) declines, and the value shifts toward the data, workflows, business rules, and context – the famous Salesforce metadata. Speaking of conversational interfaces... Salesforce is positioning Slack as the human-to-agent interaction layer. Many of the Claude sales capabilities Salesforce touted in the Claudeforce announcement are running on Slack, not Agentforce. Slackforce Surfaces enable a user to describe what they need and have Slackbot build a live dashboard, report, deck, or calculator from Salesforce to run in Slack. This opens up Slack’s opportunity to more fully compete with other generative AI tools. Slack has the advantage of already being in the collaboration flow of work – an important differentiator for Slack, and one that makes its strategic importance at Salesforce even greater. Getting back in the model business Salesforce also announced Koa, a new CRM reasoning engine built on NVIDIA Nemotron. The model is purpose-built and trained on a proprietary synthetic CRM data set, meaning it is grounded in the scenarios Agentforce agents perform across typical CRM workflows such as generating leads and qualifying opportunities. This is important for a couple of reasons: Basically, Salesforce has taken an open model and built CRM intelligence directly into it, with Salesforce controlling the weights and running it within its own infrastructure, so no customer data ever leaves Salesforce – an important consideration for AI FOMU (fear of messing up). Because it is CRM and business operations-specific, Koa is smaller and cheaper to run than a broad generative model like those of Anthropic or OpenAI, meaning customers will likely get better accuracy at a lower cost. This isn’t Salesforce trying to beat OpenAI, Anthropic, or Google at the frontier-model game; it is Salesforce recognizing that enterprise AI economic will increasingly factor specialized models optimized for specific workloads. It’s all about the ecosystem Salesforce also announced expansions of the partnerships with AWS and Google to enable customers to access Salesforce data and context from Amazon Quick and Google Gemini Enterprise, and to bring more Salesforce customers to AWS and Google Cloud. So, Salesforce is simultaneously building its own model and making it easier to use everyone else’s. This is aligned with what Salesforce has been saying for some time: it’s not the model that’s important, and models will likely be commoditized – it’s everything around the model, like data and context, that are important. The Trusted Enterprise AI Harness makes that argument explicit, positioning context, agency, actions, governance, security, and model management as a control plane around enterprise agents. The shift from build to buy On the pre-built agent front, Salesforce announced the expansion of its Agentforce agent portfolio with Casey for customer service, Paige for IT and human resources (HR), Carter for commerce, Hunter for outbound sales (with the new long-horizon runtime), Marshall for supply chain, Piper for inbound pipeline management, and Fin (through the acquisition) for Customer Experience (CX). The jury is still out on whether or not cutesy names for enterprise AI agents is a good idea (my vote is no and Stuart Lauchlan has made his position clear!), particularly if their names don’t make clear their function - nine out of 10 customers I spoke to didn’t know what Paige did or why ~~she~~ it was named that. However, it’s a sign that Salesforce has further retreated from its previous 'it’s so easy to build an agent, even a caveman can do it' talk track and recognized that if it wants to get enterprises from interesting AI pilots to broad production it’s going to have to do more than throw forward-deployed engineers at accounts – it has to deliver more pre-built, pre-tested, de-risked AI agents to the market. New vision, new questions For customers, however, Dreamforce also raised a new set of architectural and economic questions. CEO Marc Benioff touted a future where everyone would have their own personalized UI. This has to be giving a lot of design thinkers hives, never mind the Salesforce admins and support teams that will have to support them and respond to a whole series of new support requests when a user’s self-created app doesn’t render correctly or show the correct data. My money is on admins locking down a few interfaces choices to keep the support complexity manageable. If Salesforce becomes headless, Slack becomes an operating environment, Claude or Gemini becomes the interface, Koa handles some reasoning, Data 360 provides context and Agentforce executes actions, what exactly are customers buying, and how should they measure its value? Salesforce’s strategy potentially increases the value of its underlying platform, provided the right data and workflow hygiene are there (which is an open question for many customers). It also further erodes per-user models, which is why owning the harness and the control plane is so important: whoever owns control, permissions, and actions determines which agent and model gets called when and how often – ultimately becoming the arbiter of who gets the biggest piece of the token pie. That’s why governance is such an important part of what Salesforce still needs to prove out. Salesforce says every agent can only see what its permissions allow it to see, and that only works if permissions are current and correct today (and Salesforce permissions are consistent with, say, Slack or other agent permissions). These are more human issues than tech issues; Salesforce needs to beef up its training, documentation, and best practices to help bring admins along and help them identify and rectify any gaps or messiness in their permissions, data, and workflows before AI agents expose them at speed and scale. My take Ultimately, Salesforce – like everyone else – is racing to own the control pane in a headless world where few companies will have a single-vendor agent strategy. It’s also racing to own and monetize that layer before per-user pricing erodes further. However, the headless Salesforce vision creates new headaches around pricing, permissions, data hygiene, support, and governance – and right now, there’s a knowledge gap. Traditionally, Salesforce would paint the vision, Trailhead would supply the tactical technical training, and the Trailblazer Community would fill in the “here’s how you actually do it” that admins needed to be successful deploying new features or capabilities. Today, the “how-to” piece is missing, and it’s desperately needed if Salesforce expects its AIforce vision to become reality for customers. Check out diginomica's complete Dreamforce coverage here.