If you are about to put an Agentforce agent in front of customers, there is one number worth working out before anyone writes a line of Apex: how many actions does a typical conversation take?
That number decides which of Salesforce's billing models is cheaper for you. Everything below uses Salesforce's published list prices. The full breakdown, with sources, is in our Agentforce pricing guide.
The two prices you are choosing between
- Flex Credits: $500 per 100,000 credits, so $0.005 a credit. A standard or custom agent action is metered at 20 credits.
- Conversations: $2 per customer-facing conversation session, however many actions happen inside it.
So one action costs 20 x $0.005 = $0.10.
The break-even
cost_per_conversation_on_flex = actions_per_conversation * 0.10
break_even = 2.00 / 0.10 = 20 actions
- Under 20 actions per conversation, Flex Credits are cheaper.
- Over 20, the flat $2 conversation price wins.
That is the whole calculation. The hard part is getting an honest value for actions_per_conversation.
Where the estimate usually goes wrong
1. Counting only the happy path. A test conversation that looks up an order and answers a question might be 4 actions. Real traffic includes retries, clarifying turns and off-topic requests the guardrails have to handle. Those still call actions.
2. Forgetting prompts. Prompts are billed separately from actions: 2, 4 or 16 credits depending on the tier, and chunked in 2,000-token increments. An agent that retrieves too much knowledge pays for it on every turn.
3. Voice. A voice action is 30 credits and every voice minute is another 60, rounded up. A three-minute call with 8 actions is 8 x 30 + 3 x 60 = 420 credits, or $2.10. Voice agents cross the break-even far sooner than chat agents.
4. Sandbox testing. Sandbox runs at a 20% discount, not free. Ten test passes over a big regression suite adds up.
A quick way to get a real number
Before choosing a model, run a pilot on Flex Credits and pull the consumption data per session. Then:
sessions = [...] # actions per session from the pilot
avg = sum(sessions) / len(sessions)
p90 = sorted(sessions)[int(len(sessions) * 0.9)]
print("avg cost on flex:", avg * 0.10)
print("p90 cost on flex:", p90 * 0.10)
print("conversation price:", 2.00)
Look at the p90, not only the average. If a meaningful share of sessions sits above 20 actions, the conversation model caps your exposure.
Also worth knowing
- Pay-per-resolution is also $2, but only charged when the agent resolves the issue on its own. If the customer escalates to a human or leaves negative feedback, there is no charge. For a service team with a good knowledge base this moves the risk to Salesforce.
- Flex Credits expire at the end of the order term with no rollover, so buying a large block "just in case" can waste money.
- The 10,000-token rule you may have read about does not apply to actions. Salesforce's documentation says each standard action is metered as one action regardless of tokens processed.
Summary
Work out actions per conversation from real pilot data, compare the p90 against 20, and pick the model from that. The rest of the pricing detail, including editions, add-ons and negotiation points, is in the full guide on our blog.
If you are also comparing the new Core, Advanced and Max editions, we covered that in Salesforce Core, Advanced and Max explained.
Top comments (0)