Salesforce Agentforce Pricing and Consumption
Agentforce is consumption-priced, not per seat. This page breaks down the $2 per conversation rate, the Flex Credit action model, what a billable conversation actually is, and the cost at realistic enterprise volume.
Salesforce Agentforce is priced on consumption, not per seat: $2.00 per conversation on the standard list, with Flex Credits at $0.10 per action serving as the alternative most enterprises move to once monthly volume passes roughly 50,000 conversations. The shift from per-user to per-conversation pricing is the most important commercial change Salesforce has made since the move to the Customer 360 bundle, because it converts a predictable seat cost into a variable cost line that scales with how often the agent runs. This page explains both meters, what counts as a billable event, and how the bill behaves at real enterprise volume.
Two meters: conversations and actions
Agentforce bills on one of two units. The first is the conversation, defined by Salesforce as a coherent session between an end user and an agent, with a list rate of $2.00 each. The second is the Flex Credit, a prepaid unit consumed by agent actions at roughly $0.10 per action, sold in blocks of 100,000 credits for $1,000. A single conversation usually consumes several actions, so the two meters are not interchangeable at face value. The conversation meter is simpler to forecast; the Flex Credit meter is cheaper at high volume and more honest about what the agent is actually doing.
The right meter depends on how chatty your agents are. A deflection bot that resolves a question in two or three actions per session is cheaper on conversations. A multi-step agent that queries records, calls an external system, and writes back data can burn ten or more actions per session, which makes the per-action Flex model the lower bill. Model both before you commit. Salesforce sales teams default to the conversation meter because it is easier to forecast high, and the forecast becomes the commitment.
| Pricing component | Unit | List rate | Notes |
|---|---|---|---|
| Agentforce conversation | Per conversation | $2.00 | One end-user session, regardless of action count |
| Flex Credits | Per action | $0.10 | Sold in 100,000-credit blocks at $1,000 |
| Agentforce 1 Editions | Per user per month | From $550 | Bundles platform, Data Cloud allotment, and a credit pool |
| Included Data Cloud credits | Allotment | Bundled | Agent grounding consumes Data Cloud credits separately |
What counts as a billable conversation
The definition of a conversation is where consumption pricing turns into a cost-control problem. Salesforce counts a conversation as a session, but the boundaries of a session are softer than most buyers assume. A user who returns to an abandoned chat, a session that times out and restarts, and an agent that hands off to a second agent can each generate additional billable conversations. Without a contractual definition that you control, the meter runs faster than the demo suggested.
The fix is to pin the definition in the order form, not the marketing page. Specify the timeout window that closes a session, state whether agent-to-agent handoffs are one conversation or two, and require a usage dashboard with daily granularity so finance can see the meter move. The same discipline applies to Data Cloud grounding, because every retrieval the agent runs against your unified data also draws down credits. The interaction between the two meters is covered in our Data Cloud pricing breakdown.
Consumption scope creep: The first 90 days of an Agentforce deployment routinely run 30 to 60 percent above the forecast because test traffic, internal QA sessions, and retry loops all meter. Negotiate a ramp period where internal and sandbox traffic is excluded from billing, and a true-down right if the first full quarter lands below the committed volume.
Cost at realistic enterprise volume
The two meters cross over at a predictable point. Below roughly 50,000 conversations per month, the conversation meter is simpler and competitive. Above that, the Flex Credit model pulls ahead for any agent that resolves sessions in fewer than twenty actions. The table models a mid-size support deployment handling 100,000 sessions per month at an average of eight actions per session.
| Monthly volume | Conversation meter | Flex Credit meter (8 actions) | Lower cost |
|---|---|---|---|
| 10,000 sessions | $20,000 | $8,000 | Flex Credits |
| 50,000 sessions | $100,000 | $40,000 | Flex Credits |
| 100,000 sessions | $200,000 | $80,000 | Flex Credits |
| 100,000 sessions (3 actions each) | $200,000 | $30,000 | Flex Credits |
The pattern is consistent: for most real agents, the Flex Credit meter is materially cheaper, often by half or more. Salesforce leads with the conversation meter precisely because it produces the higher number. The exception is a heavy multi-step agent calling many external systems, where action counts climb past twenty per session and the conversation meter caps the per-session cost. Measure your action-per-session ratio in a paid pilot before you choose, then size the commitment to the meter that wins.
Agentforce inside the Einstein 1 bundle
Agentforce also ships inside the higher Salesforce editions as an allotment of bundled credits rather than a separate line. Agentforce 1 Editions start near $550 per user per month and fold in the platform, a Data Cloud allotment, and a starting pool of conversation or Flex capacity. The bundle looks efficient until you compare the included allotment against real volume, at which point most enterprises find the allotment covers a fraction of demand and the overage reverts to the standard consumption rate. Treat the bundled allotment as a starting balance, not a cap, and model the overage explicitly. The broader edition math sits in our Enterprise versus Unlimited comparison.
Negotiation points that hold the rate
Consumption contracts have their own negotiation points, distinct from seat deals. The first is the per-unit floor: lock the per-conversation and per-action rate for the full term so growth in volume does not reset the price. The second is rollover: unused Flex Credits should carry forward at least one quarter rather than expiring monthly. The third is the true-down, the right to reset the committed volume downward at the anniversary if actual usage came in low. The fourth is excluding non-production traffic from the meter entirely.
The forecast trap: Salesforce sizes the commitment off an aspirational deflection rate, often 60 to 70 percent of contacts. Real first-year deflection lands closer to 25 to 40 percent. Committing to the vendor forecast locks you into volume you will not use, and consumption commitments are use-it-or-lose-it. Anchor the commitment to a conservative deflection assumption and buy incremental capacity as adoption proves out.
Agentforce versus per-seat bot licensing
Agentforce is a departure from how Salesforce licensed automation before it. The earlier Einstein Bots were tied to Service Cloud seats and a session allowance, so cost was bounded by the support team's license count. Agentforce decouples cost from headcount entirely, which is the point of an autonomous agent that works without a human in the loop. The commercial consequence is that an Agentforce deployment can grow its bill without adding a single user, because the meter runs on agent activity rather than on people. Teams that budget Agentforce like a seat product are repeatedly surprised by the second-quarter invoice.
This also changes who owns the cost. A per-seat product is a fixed line that procurement signs once a year. A consumption product is a variable line that moves daily with traffic, which pulls finance into a monitoring role it did not have before. The organizations that control Agentforce cost well are the ones that assign a named owner to the consumption dashboard and review it monthly, the same discipline mature cloud teams apply to infrastructure spend.
Budgeting for a variable AI line
A consumption line needs guardrails that a seat line does not. The first guardrail is a spend cap or alert threshold set in the admin console, so a runaway agent or a traffic spike triggers a notification before it triggers an overage. The second is a monthly reconciliation that compares actual conversations and actions against the committed volume, so finance can see whether the commitment is tracking high or low with enough time to act at the anniversary. The third is a forecast that is rebuilt each quarter from real data rather than carried forward from the original sales model.
The reconciliation matters because consumption commitments are use-it-or-lose-it. Money committed to conversations that never happened is simply gone, and it does not roll into the next year unless the contract says so. Tracking the burn rate against the commitment monthly is what turns a surprise true-up into a planned adjustment. The same monitoring habit applies across every consumption product in the estate, including the Data Cloud credits the agent draws down.
Running a paid pilot before the commitment
The most reliable way to size an Agentforce commitment is a paid pilot scoped to one or two real use cases. The pilot answers the three questions that decide the bill: what is the average action count per session, what deflection rate does the agent actually achieve, and what is the cost per resolved case once both meters are counted. None of these can be known from the demo, because the demo runs on clean data and ideal questions. Real traffic is messier and meters faster.
Pilot data also resets the negotiation. Instead of accepting the vendor's deflection forecast, the buyer brings measured numbers to the table and sizes the commitment to them. A pilot that shows 35 percent deflection and eight actions per session produces a commitment grounded in evidence, not optimism, and it gives the buyer a credible basis to push back on an inflated volume estimate.
| Pilot metric | Why it matters | Drives |
|---|---|---|
| Actions per session | Decides which meter is cheaper | Conversation vs Flex Credit choice |
| Deflection rate | Sets realistic volume | Commitment sizing |
| Cost per resolved case | Compares to human cost | The business case and ROI |
Common Agentforce pricing questions
Does Agentforce replace Service Cloud seats?
Not directly. Agentforce handles sessions a human would otherwise handle, but agents still escalate to people, and those people still need seats. The realistic model is a blended one where the agent deflects a share of volume and human seats handle the rest, so the seat count falls gradually rather than disappearing.
Are internal, employee-facing agents billed the same way?
Yes. Internal agents that answer HR or IT questions meter on the same conversation or Flex Credit basis as customer-facing ones. Internal use cases often run high volume at low complexity, which usually favors the Flex Credit meter.
What happens to committed volume the agent does not use?
Unused committed volume is generally forfeited at the period boundary unless rollover is negotiated into the contract. This is why conservative sizing and a true-down right matter more on consumption deals than on seat deals.
Where this fits
Agentforce is one line in a Salesforce estate that increasingly mixes per-seat, per-credit, and per-conversation pricing, and the meters interact. For the full picture of how the pieces price together, start with the complete Salesforce licensing guide, then review the Einstein add-on pricing that often sits alongside Agentforce. When the agent grounds on unified data, the Data Cloud credit model drives a parallel bill. For help sizing the commitment and protecting the per-unit rate, see our Salesforce advisory practice and our SaaS license optimization service. A full estate review starts with the software licensing advisory team.