The Pipes Are Solved. The Judgment Isn't.
Salesforce spent 2026 solving a problem that looked hard in January. As of the Dreamforce keynote on September 15, the plumbing that connects AI models to Salesforce data is largely built, and Salesforce is no longer betting on one model to run through it.
That is the real story of this week, and it is worth separating from the part that is still genuinely unsolved.
What got announced
The keynote launched AIforce, described as a live interface layer sitting above Agentforce, Data 360 and Customer 360, aimed at letting users break free of traditional UI-based approaches. It launched alongside Claudeforce, Slackforce and Agentforce Coworker.
Salesforce also introduced Koa, its first CRM reasoning model, built on NVIDIA Nemotron. An expanded Google Cloud partnership connects Agentforce and Gemini Enterprise over MCP. An expanded AWS collaboration covers Amazon Quick, Bedrock model choice and Agentforce Voice. Sam Altman appeared on the main stage with Marc Benioff.
Claude, OpenAI, Gemini and Bedrock model selection, in a single morning. Whatever ambiguity remained about Salesforce's model strategy is gone. It is multi-model, deliberately, and the model layer is something Salesforce now shops around rather than commits to.
A note on names, because the coverage is messy. Claudeforce was announced on August 26, 2026 alongside Salesforce's Q2 FY27 earnings, and Salesforce in Claude is the plugin that came out of it, shipping 37 prebuilt sales skills for work like meeting prep, deal health review and pipeline review. At least one recap this week found an auto-transcript rendering Claudeforce as "CloudForce," and two products circulating in secondhand coverage have no published release at all. If you are writing about this, check the names against a primary source.
The pipes were laid before the partnership
None of this connectivity arrived this week. Headless 360, announced in April 2026, exposed platform capabilities as APIs, MCP tools and CLI commands rather than screens. Hosted MCP Servers went to beta in July. AIforce is the layer that brings business data and workflows to agents through exactly those MCP servers, APIs and CLI tools.
The partnerships put scale and brand names behind pipes that were already there.
This matters because a lot of commentary frames AI-plus-Salesforce as a connectivity problem still waiting to be solved. On that front it is largely solved, and it got solved faster than most people expected.
The part that did not get solved
I have been driving Salesforce work from Claude Code for a while now. Managed package development, metadata work, integration builds. Not as a demo. As how the work actually gets done.
What I keep running into has nothing to do with whether the model can write Apex. It writes fine Apex, and so do the others.
The problem is context.
An org is a decade of accumulated decisions. Fields nobody uses. A validation rule encoding a policy from a reorg three years ago. And every mature org has a handful of automations that only work in a particular order, with nothing anywhere recording what that order is or why.
Some of that lives in metadata, and metadata is now reachable. But a meaningful share of it does not live anywhere at all. It lives in the heads of people who were in the room, and some of those people left the company years ago.
A pipe can hand a model every field definition, every flow and every validation rule in your org. It cannot tell the model which of those were deliberate, which were mistakes nobody cleaned up, and which one will break billing if you touch it.
Builder tooling exists. That is not the same thing.
It would be wrong to say the builder side is untouched. Agentforce Builder lets teams build, test and deploy agents in one place, and Agentforce Vibes brings vibe coding to enterprise app development. Both shipped with Agentforce 360, and there were deep-dive sessions on Agentforce Builder running across all three days of Dreamforce this week.
So generation is covered, and getting better.
What none of that supplies is judgment about a specific org. Generation tools produce code. Retrieval tools produce metadata. Neither produces the institutional memory of why your org is shaped the way it is, and that memory is the input that determines whether generated code is safe to deploy.
It is also worth being clear about who the current wave is aimed at. Salesforce in Claude ships sales skills. Those are revenue-team workflows, built for people who use Salesforce rather than people who build it. That is a real product and a sensible first target. It is not an implementation tool, and it is not trying to be.
What to actually do
Three things are worth attention over the next two quarters.
Get your metadata reachable. Headless 360 and Hosted MCP Servers are real infrastructure and there is no reason to build your own version. Learn what they expose.
Start writing down what is not in metadata. The ordering dependencies. The reasons behind the strange validation rules. The fields that look dead but are not. This has always been good practice and it has never had a clearer payoff, because it is the one input the tooling cannot generate for itself.
Be skeptical of ROI claims, including from the keynote stage. There is still little public data on what agents cost to run in production at scale, and that number will matter more than capability demos.
The pipes are solved. The judgment isn't. That is not a complaint about the technology. It is where the remaining work is.