The short answer: Agentforce is metered on consumption rather than on seats, through Flex Credits. Actions that invoke a model consume credits, and a conversation consumes as many as its design requires. Which means conversation design is a cost decision: a chatty agent that calls three actions to answer a question costs several times what a direct one costs. Do not take pricing figures from an article. Read the current pricing page and your own contract.
What consumes credits
The unit being paid for is work done by the platform on your behalf, not people logged in. In practice that includes the reasoning the agent does to pick a topic and an action, the generation of the response text, and depending on the feature, other generative work such as prompt template runs.
The specific rates, the list of which operations draw credits, and how different actions are weighted all change. They are published, and they are the sort of detail that is wrong within a quarter, so treat the current pricing documentation as the only source worth quoting.
The structural point is stable, though, and it is the one worth planning around: cost scales with conversation volume multiplied by turns per conversation multiplied by actions per turn. Two of those three are entirely in your control.
Why conversation design drives the bill
Consider two agents answering the same question, “has invoice 4471 been paid”.
The first has one topic and one action that takes an invoice number and returns status, amount and date. One turn, one action, one generated response.
The second has the same job spread across three actions: find the account, list its invoices, get status for one. The agent must chain them, each step is a reasoning step plus an action call, and the conversation runs four or five turns because the agent asks clarifying questions between steps. Same answer, several times the consumption, and a worse experience.
Designs that quietly multiply cost:
- Granular actions that must be chained. Every chain step is reasoning you pay for. Prefer one action that does the job.
- Confirmation habits. “Would you like me to look that up?” doubles the turn count for no information.
- Long instructions. Instruction text is read on every turn, in every conversation.
- Oversized grounding. Returning fifty records so the model can pick one is paying to send data you could have filtered.
- Retries. An agent that fails and tries again consumes twice for a worse outcome, and poor action descriptions cause exactly this.
Estimating before you build
The estimate does not need to be precise. It needs to be the right order of magnitude, and you can do it on paper.
- Count the conversations. Use the real volume of whatever the agent replaces: cases per month, chats per week, internal queries per day. Not an aspiration.
- Estimate turns per conversation. Walk through five real examples and count. Most teams guess two and find six.
- Estimate actions per turn from your design. Chained designs average more than one.
- Multiply, then apply the current published rate.
- Double it. First designs are chattier than intended, and pilots include a lot of people poking at it.
Then compare against the thing it replaces, measured the same way. An agent handling a case that would otherwise take an agent twenty minutes has an obvious margin. An agent answering a question that a report already answers, for free, does not.
Controlling it once live
Watch consumption from day one of the pilot, not from the first invoice. Establish a per-conversation average early, because that number is what tells you whether an instruction change made the agent more expensive.
Deploy narrowly first. A pilot with one team gives you a real per-conversation figure to multiply, and a figure from your own org beats any benchmark.
The cheapest optimisation is usually consolidating actions. The second cheapest is deleting instruction text nobody can write a test for.
Frequently asked
How much does a conversation cost? It depends on the design and the current rate card, and anyone quoting a firm number without seeing your agent is guessing. Build one, measure it, multiply.
Do test runs consume credits? Generative work is generative work. Assume testing consumes, budget for it, and prefer sandboxes where your agreement allows.
Is there a cap? Consumption controls and alerting exist and have been changing. Find out what your org supports before a pilot, not after a surprise month.
Is it cheaper than a person? For high volume and low complexity, often. For low volume and high complexity, frequently not, and that is the case people avoid calculating.
