Salesforce moved the reasoning into Claude — and the meter that would bill it is three blank lines on its own rate card

What was announced

Salesforce’s framing at Dreamforce was blunt enough that its own trade press ran it as the headline: AI replaces the UI. In the company’s words, customers “don’t have to come to Salesforce to get work done” — Salesforce “comes to them,” in Claude, in Slack, in Lightning, wherever they already are.

Mechanically, AIforce is an interface layer exposing Salesforce data, workflow and business logic to external agents through MCP servers, APIs, plugins and skills. It reasons across Salesforce data, logic and conversations at once, runs inside existing permissions and security models, and carries zero data retention so prompts are not retained by model providers. It launched with three pieces:

• Claudeforce — a prebuilt MCP server inside Claude carrying 37 prebuilt sales skills across prospecting and pipeline hygiene, plus a Salesforce Development Plugin for Claude Code with more than 40 skills and access to a GitHub skills library. In beta to all customers, piloted by Deloitte, GitLab and Legora. Tableau analytics and service, marketing and commerce skills are promised later.
• Slackforce — Salesforce context inside Slack through interactive Surfaces, a Slackbot reasoning across both systems, natural-language record management via Slack CRM, and Slack Code, the team development environment we covered when it launched in August.
• Agentforce Coworker — an AI teammate inside Lightning, which Salesforce says reached 100,000 activated users within 35 days of launch.

None of it is a new architecture so much as the payoff of an existing one. Headless 360, introduced at TrailblazerDX earlier this year, already exposed Data 360, Customer 360, Agentforce and Slack through APIs and more than 60 MCP tools. One Salesforce executive put the relationship plainly: the earlier work “was like the raw materials,” and AIforce “is like the chocolate chip cookie that comes out of the oven.”

The meter follows the reasoning

Here is the part that matters to anyone with a budget.

Salesforce’s agent pricing is consumption-based and denominated in Flex Credits: $500 per 100,000 credits, or half a cent a credit. A standard Agentforce action costs 20 credits — about $0.10 in production and $0.08 in sandbox. A voice action runs 30 credits, roughly $0.15. A basic prompt is about a cent. That meter is the one we wrote about on 12 September, when Salesforce shipped agents capable of pursuing goals across days and weeks onto a per-action bill, with the AI Control Plane that governs cost not arriving until early fiscal 2028.

AIforce changes which side of that meter the work happens on. When Claude pulls records through the hosted MCP server and reasons over them itself, Salesforce is a data source and Anthropic is the compute. Basic MCP access is included with Enterprise Edition and above and is not gated behind an Agentforce licence or add-on. The reads and writes count against your org’s existing daily API allowance, not against Flex Credits. Flex Credits are drawn when you invoke an Agentforce action, run a Data 360 query, or use Prompt Builder — that is, when Salesforce does the thinking.

So the same question, answered two ways, lands on two different invoices. A read-heavy analytical session bills to Anthropic in tokens. An action-heavy automation bills to Salesforce in credits, and to Anthropic as well. The 37 prebuilt skills shipping inside Claude were built to be useful, not to be cheap, and nothing in the announcement tells a buyer which of them cross the line.

Three blank lines on the rate card

This is where the story stops being an architecture observation and becomes a procurement one.

Salesforce’s published Flex Credits Rate Card, dated 17 June 2026, carries usage types for Headless Platform Interaction, Record Operation and Process Invocation — CRUD operations through Salesforce Hosted MCP Servers, and Salesforce processes including Flow and Apex invoked through MCP. Every one of them lists its credit multiplier as TBA.

Salesforce Ben’s reading of the accompanying documentation is that both new usage types are “currently unmetered,” and that Salesforce “has committed to 30 days’ notice before a multiplier is added and metering begins, so these operations could carry a cost down the line.”

Read that against Tuesday’s announcement. Salesforce has just made third-party agents reading and writing through MCP the centrepiece of its product strategy — and the line items that would bill exactly that traffic are named, published, and blank. A blank multiplier on a rate card is not an omission. It is a reservation. Salesforce has built the meter, wired it, labelled it, and left the rate unset, while telling customers in a footnote how much warning they will get before it starts spinning.

There is nothing improper here. Salesforce disclosed the usage types, disclosed that they are unmetered, and disclosed the notice period. That is more transparency than most vendors offer for a price they have not decided. The problem is one of asymmetry: the disclosure lives in a rate card PDF and an admin guide, while the thing it governs is the flagship announcement of the company’s biggest event of the year.

The pattern this belongs to

Unpriced-then-priced is not a Salesforce quirk. It has been the dominant shape of AI pricing all year, and this site has now tracked enough instances to call it a pattern rather than a coincidence.

GitHub Copilot’s promotional credits expired and included usage dropped. OpenAI’s Sol price cut arrived with a promotional clock attached. Gemini 3.8 Flash’s headline rate is introductory through 31 December, doubling on 1 January. The GSA’s OneGov deal runs on dollar-a-year pilot pricing with no spend cap behind it. OpenAI’s Agents API ships the harness free and meters the sandbox hours. Sakana’s orchestration tokens made the effective price diverge from the list price.

The common structure: the capability is free or cheap while adoption is the goal, the meter is built before the rate is set, and the disclosure is technically complete and practically invisible. Anthropic’s own compaction billing change last week is the same species — a documented change to what you are charged, published where the people being charged do not read.

AIforce is a larger instance than most, because it is not a rate on a model. It is a rate on the act of an external agent touching your system of record, which is the thing every enterprise AI architecture built this year ends up doing.

What to do with this

Pilot Claudeforce. The engineering is good — prebuilt MCP rather than bespoke glue, zero data retention, existing permissions enforced rather than a parallel access path. Those choices remove the objections that normally stall this class of integration, and the capability is real.

But instrument it while it is free, because that is the only window in which measuring is cheap. Count MCP reads, writes and process invocations per user per week and keep the baseline. Get the unmetered status and the 30-day notice confirmed in writing by your account team rather than inferred from a PDF. And when you evaluate a skill, ask where it puts the reasoning, because a skill that reads into Claude and a skill that fires an Agentforce action look identical in the output and diverge in the bill.

If you are choosing between agent platforms rather than extending Salesforce — our agent platform rankings and the Sierra versus Decagon comparison cover the customer-facing end of the market, where per-resolution pricing is more settled than per-action pricing is here. And if the Claude side of this is new to you, the Claude and Claude Code reviews cover what the model and the harness actually do before Salesforce’s skills are layered on top.

Thirty days’ notice is enough time to receive a bill. It is not enough time to re-architect a workflow that ten thousand people have already started using.

Questions or corrections? Email Pick Right. Want the full list? See all news.