Agentforce
AIforce: Salesforce's bet that AI replaces the UI, explained for builders
At Dreamforce 2026, Salesforce framed AIforce as a live interface layer that pushes your data, workflows, permissions, and governance into Claude, Slack, and your own apps. Here is what it actually is under the branding, when to build against it instead of the Lightning UI, and what changes for security.
Marc Benioff walked the Dreamforce keynote through the interface eras: command line, GUI, web, mobile. AI, he said, is the next one, and this time the interface is alive. The product name for that bet is AIforce, announced on September 15. The claim underneath the slogan is that your users stop coming to Salesforce’s screens and start reaching Salesforce from wherever they already work.
Strip the keynote language and AIforce is a positioning layer over plumbing that already shipped. It takes your customer data, workflows, permissions, business logic, and governance controls and projects them into a growing list of entry points: Claude, Slack, custom apps, and Lightning itself. The three packaged surfaces Salesforce is leading with are Claudeforce, Slackforce, and Agentforce Coworker.
My read: AIforce is not a thing you buy or turn on. Think of it as a name for the Headless 360 architecture Salesforce spent 2026 building, and that reframing changes what you plan for. You do not adopt AIforce. You decide, per workflow, whether its right home is a screen your users log into or an agent they talk to somewhere else, then you make sure your permissions hold up when the front door moves.
What AIforce actually is
There is no AIforce SKU, no pricing, and no general-availability date. Salesforce disclosed none of those at launch, which is the first tell that AIforce is a story about existing capabilities, not a new engine.
The engine is the Headless Toolkit: an open architecture that exposes Salesforce as MCP servers, APIs, plugins, agent skills, and developer tooling. It builds on Headless 360, which Salesforce expanded on August 19 to turn every cloud into reusable capabilities any authorized agent can call. The Headless 360 MCP Server reached beta over the summer. The Data 360 MCP Server exposes close to 200 Data 360 APIs the same way.
When the keynote says AIforce projects Salesforce into Claude, the mechanism is the one I wrote about in the Headless 360 MCP post. An MCP client connects to a Salesforce-hosted MCP server, discovers a small set of tools, and invokes governed capabilities behind them. AIforce is the umbrella name for doing that across every surface at once, plus the three prebuilt interfaces so you do not have to wire the common ones yourself.
The clever part of the design is worth restating, because it is what makes “any interface” tractable. Rather than dump thousands of Salesforce features into an agent’s tool list, the Headless 360 server presents an agent four tools, not thousands. Behind those tools sits a library of agent skills that grows without changing the agent’s surface. The agent stays simple. The catalog scales on its own.
The three surfaces, and what each one is for
Claudeforce is the connector that feeds Salesforce data and controls into Anthropic’s Claude, arriving with a set of prebuilt sales skills. I covered its shape and the trust-boundary question in the Claudeforce explainer, and the Claudeforce FAQ answers the pricing and model questions. Reach for it when a seller or service rep lives in a chat assistant and wants CRM actions there.
Slackforce puts CRM context and actions inside Slack, and its Surfaces feature turns a plain-language prompt into a live dashboard or report built from your Salesforce and Slack data. That earned its own post.
Agentforce Coworker is the employee-facing agent that turns up in your search bar and inside the tools your teams use. Its job is internal work, not customer conversations.
The pattern across all three is the same. None of them is a new place to build agents. They are new places for an already-built, already-governed agent to appear.
The decision AIforce forces: screen or surface
For fifteen years the answer to “where does this workflow live” was a Lightning page, a screen flow, or an LWC. AIforce adds a real second option, and the useful work is deciding which one a given workflow deserves.
Build it as a surface, reachable from Claude, Slack, or your own front end, when the user’s center of gravity sits somewhere other than Salesforce. A seller who runs their day in Slack should update the opportunity from Slack, not tab over to a record page. A developer using Cursor should query an account without leaving the editor.
The Agent API and the MCP servers exist for exactly this, so the CRM action happens where the person already is.
Keep it on a Lightning screen when the work is dense, multi-record, and visual: a console for a support agent working ten cases, a guided flow with validation at every step, a report a manager scans every morning. A conversational surface is worse than the screen for anything where seeing twenty fields at once is the point.
A conversational interface is a genuine gain for a single intent expressed in a sentence. It is a downgrade for dense, structured work where the screen shows more than a sentence can say.
The trap is treating AIforce as a mandate to move everything off-screen. Benioff’s framing is a vision statement, not a migration plan. Your users still have jobs where a table beats a chat turn. Move the workflows that fit the surface, and leave the ones that do not.
Wiring a surface: what the plumbing looks like
If you want a non-Salesforce client to reach your org, the concrete step is connecting an MCP client to a Salesforce-hosted MCP server. The client config is ordinary MCP, the same shape any MCP client uses:
{
"mcpServers": {
"salesforce-headless360": {
"url": "https://<your-domain>.my.salesforce.com/mcp/headless360",
"transport": "http",
"authorization": "oauth2"
}
}
}The values are illustrative, check the Headless 360 MCP reference for the exact endpoint and auth flow for your org, but the structure is the point. The client authenticates as a real user through OAuth, discovers the handful of tools the server advertises, and calls them. Everything the user is not allowed to see or do stays invisible, because the call runs under their identity, not a superuser’s.
That is the difference between this and an old-style integration user with a wide profile. The MCP path is designed to carry the running user’s permissions, which is the only reason “Salesforce from any interface” is safe to say out loud. The MCP vs. APIs guide Salesforce published lays out when to reach for which, and it is worth reading before you pick.
The governance bill comes due at the boundary
Here is the part your security team will care about more than the keynote did. When Salesforce was only reachable through its own UI, the UI was the enforcement point. Sharing rules, field-level security, and profile restrictions all sat behind a login page you controlled. AIforce moves the front door, and the enforcement has to move with it.
Two things decide whether that is safe. The first is identity: every surface has to run under the real user’s permissions, not a service account with broad access wired in to make the demo work. A service account behind an MCP tool is the fastest way to leak data an individual user should never see. The Data 360 governance post is about exactly this: an agent must only see what the person it acts for is allowed to see.
The second is the audit trail. An action taken from Claude or Slack still needs to land in your logs as that user’s action, from that surface. When agents you did not build start calling your org, you also need to know they are who they claim to be, the problem Web Bot Auth and delegated identity are meant to solve.
AIforce also widens the surface you have to govern for sprawl. Every new interface is another place an agent can act, and agent sprawl is already the new shadow IT. The AI Control Plane framing, governing the agents you did not build and cannot see, is the muscle you need before you turn on three new front doors.
What to do before your users ask for it
Someone on your team saw the keynote and will ask to put the CRM in Slack or Claude by Friday. The useful preparation is not waiting for an AIforce license, because there is not one to wait for. It is getting the foundation right so the surfaces are safe when you open them.
Confirm your permission model is real, not aspirational. Run the org health scorecard or your own audit and find every profile with more access than the job needs, because those become data leaks the moment a wide-open user reaches the org through a chat window. Then pick one workflow, one that genuinely fits a conversational surface, and wire it through the MCP server end to end, under a normal user’s permissions, with the audit trail checked. One workflow proven properly teaches you more than a fleet of half-governed ones.
The interface revolution Benioff described may or may not play out the way the slides said. What is already true, today, is that your org is reachable from outside its own UI, and the work of making that safe is the same work it always was: least privilege, real identity, and an audit trail that survives the trip.
Understanding the basics
What is Salesforce AIforce?
AIforce is the name Salesforce introduced at Dreamforce 2026 for a live interface layer that projects your Salesforce data, workflows, permissions, and governance into other interfaces such as Claude, Slack, and custom apps. It is built on the Headless 360 architecture and its Headless Toolkit of MCP servers, APIs, and agent skills, rather than being a separate product with its own engine.
How much does AIforce cost?
Salesforce did not disclose pricing, a general-availability date, or regional availability for AIforce at launch. The underlying capabilities, such as the Headless 360 and Data 360 MCP servers, are delivered through the platform, and agent consumption is billed through Agentforce Flex Credits.
Is AIforce the same as Agentforce?
No. Agentforce is the platform for building and running agents. AIforce is the layer that lets an already-built, already-governed agent or capability appear in interfaces outside Salesforce. The three packaged surfaces at launch are Claudeforce, Slackforce, and Agentforce Coworker.
Does AIforce mean I should stop building Lightning pages?
No. A conversational surface fits a single intent stated in a sentence. Dense, multi-record, visual work is still better on a Lightning screen, a console, or a guided flow. AIforce adds a second option for where a workflow lives, it does not retire the first.
If you are planning for an interface layer
The safe path to “Salesforce from any interface” runs through least privilege and real user identity, not a new license. If you want a second pair of eyes on whether your permission model and audit trail hold up before you open a new front door, talk to us, or start with the Agentforce readiness assessment.
