What Salesforce Agentforce costs, and which unit you are actually buying

The short version

• Salesforce lists Agentforce at $2 per conversation, and that rate was still on its pricing page in November 2026.
• Flex Credits bill per action, and one customer request can spend several actions.
• Per-user editions make the bill predictable by charging for seats whether the agents do work or not.
• You cannot forecast an agent's cost until you know how many actions an average request consumes.
• Comparing Agentforce with outcome-priced competitors is a unit conversion before it is a price comparison.

How much does Salesforce Agentforce cost?

Three ways, and they do not compare cleanly. There is a per-conversation rate, listed at $2. There is Flex Credits, a consumption meter that bills per action, with the rate published on the Agentforce pricing page. And there are per-user editions, whose list price sits in the pricing summary of Salesforce's own flexible pricing announcement. The question is never 'what does it cost', it is 'which unit are we buying'.

A conversation is a unit your business already recognises. You know roughly how many support conversations you handle in a month, and you can multiply. An action is an internal unit. It is one step the agent takes: a lookup, a record update, a generated reply. One conversation can spend a single action or a dozen, and nobody in the room knows which until the agent runs against real traffic.

The per-user editions look safest because they bill like the software you already buy. They move the risk rather than removing it. You pay per seat whether the agents do work or not, and the seat count grows with headcount rather than with value delivered. If your agent resolves a high share of contacts, per seat is cheap. If it does not, you are paying a subscription for a feature nobody uses.

What is a Flex Credit, and what does one request actually spend?

A Flex Credit is one action, and an action is one step inside the agent's work, not one message from the customer. That distinction is the whole forecasting problem. Salesforce's own worked example on the pricing page assumes two actions per request and 20 requests a day, which comes to 40 Flex Credits a day, run across a 30-day month. That is 1,200 credits a month before anything goes wrong.

Two actions per request is a modest assumption. It describes an agent that looks one thing up and answers. Real deployments rarely stay there. An agent that checks an order, checks a policy, asks a clarifying question, checks again and then writes a case note has spent five actions on one request, and that conversation may still end with a handover to a person.

So treat the vendor's example as a shape, not a forecast. Run your own version: actions per request, requests per day, days per month. Then run it again with the actions figure doubled, because that is what happens when the agent gets the extra tools it needs to be useful. If the doubled number is unaffordable, you have learned something important before signing rather than after.

Which meter are you buying?

Which Agentforce pricing model is right for your team?

Match the meter to the shape of your traffic. If volume is predictable and conversations are short, per conversation is the easiest number to defend to finance, because it maps onto a figure you already track. If the agent does heavy internal work on a small number of high-value requests, per action can be cheaper, and it rewards you for making the agent efficient. If agent use is spread thinly across a large licensed team, a seat-based edition can beat both.

The honest answer for most buyers is that you do not know yet. Agent behaviour under live traffic is the input to this decision, and a pilot is how you get it. Ask for a short consumption arrangement first, measure actions per conversation, then negotiate the longer commitment with a real number in your hand.

Be careful about committing to an annual seat count during the enthusiasm phase. Seats are easy to add and hard to remove.

What pushes an Agentforce bill above the estimate?

Nothing dramatic. The bill rises because of ordinary engineering decisions that each look reasonable on the day they are made.

• •Retries. A failed tool call that the agent runs again is a second action, and the customer sees one answer.
• •Added tools. Every integration you connect makes the agent more useful and more expensive per request.
• •Clarifying questions. An agent that checks before acting spends more actions than one that guesses, and checking is usually the better behaviour.
• •Handovers. A conversation that ends with a human still consumed its actions, so you pay the meter and the salary.
• •Test traffic. Automated evaluation runs hit the same meter unless you have arranged otherwise.
• •Knowledge gaps. When retrieval returns nothing useful, the agent tends to try harder, which means more actions for a worse answer.

Is Agentforce any good, and how do you review it honestly?

It depends on whether you are already deep in Salesforce, and on what share of your conversations an agent can finish without a person. Those two variables decide most of the outcome. Everything else is detail.

Public reviews are hard to use here. Ratings tell you about onboarding and support experience, which is worth reading, but they rarely tell you the actions per conversation figure for a company shaped like yours. Salesforce does publish production telemetry from its own help site deployment on a public page, which is more transparency than most vendors offer. Read it as what it is: the vendor running its own product, on its own documentation, with its own data in good order. Your knowledge base is probably messier than that.

The review that matters is the one you run. Take a hundred real conversations from last month, including the awkward ones. Count how many an agent could have finished end to end with the systems it would actually be allowed to touch. That number, times your contact volume, times the published rate, is your business case. If you cannot reach that number, you are not ready to sign a multi-year deal.

How does Agentforce pricing compare with its competitors?

Mostly it does not compare directly, because the competitors sell different units. Intercom's Fin is priced per billable outcome, charged at most once per conversation, with a 50-outcome monthly minimum when it sits on a helpdesk other than Intercom's own. That is a resolution-style meter. Flex Credits are a work-style meter. Comparing the two quotes is a unit conversion first and a price comparison second.

Do the conversion with your own numbers. Outcome pricing tends to be better value when the agent resolves a high share of conversations, because you only pay for the wins. Action pricing tends to be better when conversations are short and the agent does little work per contact. If you cannot estimate your resolution rate yet, outcome pricing is the safer starting point, because the vendor carries more of the uncertainty.

The third option is building on a model API directly, which removes the per-conversation markup and replaces it with engineering cost and operational responsibility. It is cheaper per unit at volume and more expensive at small scale. It only makes sense when the agent is close to your core business rather than a bolt-on to a CRM you already pay for.

Do you need Salesforce itself for Agentforce to be worth buying?

In practice, yes. Agentforce earns its price by sitting on your CRM records, your objects and your permission model. Strip that away and you are buying a general agent platform in a Salesforce-shaped wrapper, which is not the thing it is good at.

If you are not on Salesforce today, add the CRM subscription to your total before you compare anything. Salesforce publishes a list price for its small-business Starter Suite, and an entry-level CRM subscription across a team is a real line item next to the agent spend. Changing CRM in order to buy an agent is rarely the cheapest route to the outcome you want. Say that plainly before anyone starts drawing a migration plan.

If you are already on Salesforce, the calculation flips. The data is there, the permissions are there, and much of the integration work you would otherwise pay an engineering team for is already done. That is the case where the per-conversation or per-action premium is easiest to justify.

What should you ask Salesforce before you sign?

Pricing questions are answerable. Ask them in writing, and ask for answers specific to your configuration rather than to the published page.

1. What counts as one action in our configuration, with our tools connected?
2. What happens to the meter when an action fails and the agent retries?
3. Does a conversation handed to a human still consume its full charge?
4. Is test and evaluation traffic billed at the same rate as production?
5. What happens to unused credits at the end of the term: do they roll over, or expire?
6. What is the price at renewal if our volume doubles, and is that in the contract or in a conversation?