What Agentforce really costs: licences, credits and implementation

Flex Credits are a prepaid pool: you buy credits and every action an agent performs draws from it. An action is a single step, such as looking up a record, calling an API or generating an answer. Conversation pricing ignores actions and charges a flat rate per conversation, however long it runs. Per-user licences are the classic model and make sense when the agent is a tool for your own staff rather than a front door for customers.

The arithmetic is easy and worth doing before you commit. Twenty actions at roughly $0.10 each equals $2.00, which is exactly the price of one conversation. So if your average conversation takes fewer than twenty actions, credits cost less; if it takes more, the flat conversation price wins.

The trap is that nobody knows their number in advance. A support agent that answers from a knowledge base and closes might use five or six actions. One that checks an order, calls an external system, updates a record and drafts an email can pass twenty without anyone noticing. This is the single most useful thing to measure in a pilot, and it is measurable within days.

Prices here are Salesforce list prices verified in September 2026, before VAT and before any discount negotiated on your contract. Salesforce changes them more often than you would expect, so check the official pricing page before signing anything.

The licence buys you the engine. Someone still has to decide what the agent does, connect it to your data, write the instructions, restrict what it is allowed to touch, test it against real conversations and set the rules for handing over to a human.

I quote a first Agentforce use case from EUR 1,400 at a fixed price: use-case design, agent build, guardrails, test conversations before go-live and a review after the first weeks of real traffic. That is a defined scope with a defined price, and it is deliberately small, because the point of the first project is to find out whether the second one is worth doing.

An agent answers from what it can read. Duplicated accounts, half-empty fields, three competing knowledge articles saying different things, statuses that mean whatever each team decided they mean: all of it comes straight back out of the agent, phrased confidently.

In practice, on a mid-sized org, cleaning and structuring what the agent needs is often the largest slice of the first project, and the one with the longest life. That work keeps paying off in reports and automations long after the agent question is settled.

Before asking what an agent costs, it is worth asking whether you need an agent. Three different tools solve three different problems, and two of them have no per-conversation cost at all.

The rule of thumb is short: if the request arrives phrased differently every time and the answer depends on data sitting in several places, you need an agent. If the path is always the same and only the values change, a Flow does the same job with no consumption and no chance of answering badly.

If you already run an Einstein Bot, replacing it is not automatic: it pays off when the share of conversations ending in "let me pass you to an agent" stays high, because that is exactly where an agent adds value. If the bot closes almost everything on its own, that money works harder somewhere else.

Volume decides. Below roughly a hundred repetitive interactions a month, the cheapest answer is almost never an agent: a well-built Flow, a decent knowledge base or a better email template will cost less and break less.

The other disqualifier is variability. Agents earn their keep on processes that repeat with small variations, such as first-level support, lead qualification or appointment booking. On decisions that are different every time, they add cost and risk without removing work.

On 26 August 2026 Salesforce and Anthropic announced Claudeforce: Claude becomes the default reasoning model in Agentforce Vibes and Coworker, stays selectable in Agent Builder, and Salesforce moves into Claude through a plugin with prebuilt sales skills. For the cost question it changes little: the three pricing models above still apply, and Claude inference is consumption-based and contracted separately. What changes is the quality of reasoning you get for the same credits, which is worth measuring on a pilot before assuming anything.

One use case, chosen for having a number attached to it: tickets deflected, minutes saved per case, leads qualified without a human. Build it, measure it against that number for a few weeks, then decide whether to scale. If the number does not move, you have spent the price of a small project instead of the price of a programme.

If you want a figure for your own case, describe the process you have in mind and you get a written quote within one business day, including an estimate of the credits it will burn at your expected volume.