Agentforce: what an agent costs when you pay ten cents an action

Salesforce sells Agentforce by the action. One action costs 20 Flex Credits, credits go for $500 per hundred thousand, so an action is ten cents. Before May 2025 the same product was billed at $2 per conversation, which is the number most people still have in their heads.

That change is the most interesting thing about this product, and it has nothing to do with the AI. Per-action pricing is an argument about what an agent is: not a seat, not a conversation, but a sequence of small pieces of work that either happen or don’t.

What you are buying

Agentforce is Salesforce’s agent layer: agents that answer service cases, qualify inbound leads, prep sellers and act on records, built on the data already sitting in the org. It runs against Data 360 (the artist formerly known as Data Cloud) for anything that needs context beyond a single object.

Salesforce’s own numbers from the Q1 FY27 release on 27 May 2026: $1.2 billion in Agentforce ARR, up 205 per cent year on year; $3.4 billion combined with Data 360; 3.8 billion “Agentic Work Units” delivered to date across Agentforce and Slack; 28.6 trillion tokens processed. More than half of new bookings came from existing customers.

“Agentic Work Unit” is a unit Salesforce defined for itself, and it appears in no accounting standard. Cumulative-to-date numbers in a category that is eighteen months old will always be large, because the denominator is short. Read them as a growth signal, not as a measure of work done.

The arithmetic nobody does before the meeting

An action is a single step: a retrieval, a generated response, a tool call, a record update. A support conversation typically burns several. Do the sum with your own volumes and the number stops being scary.

Scenario	Actions/month	Flex Credits	Cost
500 service conversations, 4 actions each	2,000	40,000	$200
2,000 service conversations, 4 actions each	8,000	160,000	$800
1,000 lead qualifications, 6 actions each	6,000	120,000	$600

At the old $2 per conversation, the first row cost $1,000. At ten cents an action it costs $200, as long as the agent takes four steps and not forty. That caveat is the whole risk in the model: a badly scoped agent that retries, re-reads and re-asks does not fail loudly, it just gets expensive quietly. Consumption pricing moves the cost of sloppy design from your patience to your invoice.

Salesforce also hands out 100,000 Flex Credits at no cost through Foundations for customers on Enterprise Edition or above. That is 5,000 actions to find out whether the thing works, which is a genuinely fair offer and more than most vendors put on the table.

$0.10 per action, $500 per 100,000 credits. The previous model charged $2 per conversation, a 20× difference in the unit, in the customer’s favour, for anything that resolves in under twenty steps.

Salesforce, Agentforce flexible pricing announcement, 15 May 2025

Where the money actually goes

The credits are not the expensive part. Three things underneath them are.

The edition. Agentforce assumes a Salesforce org, and the useful tiers assume Enterprise or above. If you are not already on Salesforce, the honest price of an Agentforce project starts at the CRM licences for your whole revenue team and goes up from there.

Data 360. Agents that only see one object give thin answers. The ones that impress in the demo are reading across service history, orders and product data, and that unification is a separate product with a separate bill.

The build. Somebody has to write the topics, the instructions and the guardrails, connect the actions, and then keep them working when the underlying process changes. This is the line item that gets estimated at “a few weeks” and behaves like any other integration projectDE (auf Deutsch).

None of this is unique to Salesforce. It is the standard shape of an agent programme, and it is why the per-action price is the wrong number to negotiate. Nobody has ever been ruined by ten cents.

The DACH question

Two things come up in every German evaluation, and only one of them has a clean answer.

Where the data sits is answerable: Salesforce runs EU infrastructure, and residency belongs in the order form rather than in a marketing claim. Ask which components are covered, name them individually, and get the sub-processor list for the model providers, because the difference between storage and processing residency is exactly where these conversations go wrong.

The harder one is the works council. An agent that reads service cases and generates responses is also, incidentally, a system that produces a record of how fast individual employees work. That makes a rollout a co-determination topic under § 87 BetrVG before it is a procurement topic, and starting that conversation late is the single most common reason a technically finished rollout sits unused for a quarter.

Who should buy it

Companies already running Salesforce as their system of record, with a support or inbound volume in the thousands per month, and someone internal who owns the CRM. For them Agentforce is the shortest path from “we should automate this” to something in production, and the consumption pricing means a pilot costs a rounding error.

Companies not on Salesforce should not buy Salesforce in order to get Agentforce. The agent layer is not the scarce part, and orchestration on top of the tools you already run gets you there without moving your CRM, which is the comparison worth making first.

And any company whose process is not written down anywhere should expect the discovery phase to be the project. An agent needs a decision rule for every branch a human currently handles by knowing things. Writing those rules down is the work; the agent is what happens afterwards.

If you want that written down without buying a platform first, thirty minutes with us will tell you which parts are actually automatable.

Frequently asked questions