Agentforce Pricing: Two Models, One Org
Agentforce consumption pricing comes in two forms that cannot run in the same Salesforce org, so the choice is structural rather than a setting you can change later. The Conversations model is $2 per conversation and covers customer-facing agents only. The Flex Credits model — introduced in May 2025 to replace the original per-conversation-only approach — is $500 per 100,000 credits, where a standard action costs 20 credits (about $0.10) and an Agentforce Voice action costs 30 credits (about $0.15). Flex Credits span every use case: customer agents, employee agents and Voice. Salesforce also bundles a free tier, Salesforce Foundations, giving Enterprise Edition customers and above 200,000 Flex Credits and 250,000 Data Cloud credits at no cost, alongside Agent Builder and Prompt Builder.
The move from a flat conversation price to a credit meter is the same shift reshaping every agent platform in 2026 — the dynamic we set out in AI agent licensing and pricing models and seat-based vs consumption AI pricing. It changes what you are buying from a predictable seat count into a variable bill that tracks how hard your agents work, which is exactly why the contract terms matter more than the headline rate.
| Item | Conversations | Flex Credits |
|---|---|---|
| Unit price | $2 per conversation | $500 per 100,000 credits |
| Standard action | — | 20 credits (~$0.10) |
| Voice action | — | 30 credits (~$0.15) |
| Use cases covered | Customer-facing only | Customer, employee & Voice |
| Free entitlement | Salesforce Foundations | 200,000 credits via Foundations |
Flex Credits or Conversations: Which Wins
The Conversations model wins on simplicity for low-volume, purely customer-facing agents — one predictable price per conversation, easy to forecast and explain to finance. Flex Credits win on flexibility and, past a threshold, on cost: at roughly $0.10 per action they become more economical for most enterprises once monthly volume passes around 50,000 conversations, and they are the only model that covers employee-facing agents and Voice. Because the two cannot coexist in one org, picking wrong means a migration, not a tweak.
The decision turns on your action mix, not just volume. A "conversation" can involve several metered actions under Flex Credits, so a chatty, multi-step agent can consume more in credits than the flat $2 conversation would have cost — while a high-volume, lightweight agent is far cheaper on credits. Model the real number of actions per conversation across your intended agents before you commit; getting this wrong is the most expensive mistake in an Agentforce deal, and it sits inside your wider Salesforce commercial relationship rather than standing alone.
A conversation is one price; an action is another. A multi-step agent can burn several actions per conversation, so Flex Credits can cost more than the $2 flat rate — or far less. Model actions per conversation before you choose, because you cannot run both models in one org.
The Three Buying Structures
Within Flex Credits, Salesforce offers three buying structures, and each shifts risk differently. Pre-purchase buys credits upfront at the contracted rate and draws them down as consumed — good for a known, steady workload. Pay-as-you-go bills monthly in arrears on actual usage with no upfront commitment — the safest choice while consumption is still uncertain, because you never pay for credits you do not use. PreCommit is a volume-based structure with negotiated rates for large deployments, and it carries the best unit price — but it also carries a true-up charge if your actual usage finishes below the committed amount.
That true-up is the trap. PreCommit's discount is real, but it converts your usage forecast into a financial floor: under-consume and you pay for the shortfall anyway, which is the consumption-era version of shelfware. The right approach is to commit only the volume you are genuinely confident you will exceed, and carry the variable layer on pay-as-you-go — a hybrid that captures the PreCommit rate on your baseline without betting the whole forecast. Review the AI contract red flags before signing any PreCommit, because the shortfall mechanics are where buyers get caught.
What to Negotiate in a Consumption Deal
Consumption contracts have different levers from seat deals, and the ones that matter most are not in Salesforce's opening proposal.
Lock the floor and stage the ramp
Lock the per-conversation and per-action rate for the full term so growth in volume cannot quietly reset the unit price upward — rate protection is the single most valuable clause in a consumption deal. Then negotiate a ramp period that excludes internal QA and sandbox traffic from billing: the first 90 days of an Agentforce deployment routinely run 30–60% above forecast because test sessions, internal QA and retry loops all meter. Paying production rates for your own testing is avoidable, but only if you raise it before signature.
Win a true-down and prove the value
Secure a true-down right so that if the first full quarter lands below the committed volume, you can reset the commitment rather than pay for unused credits. Use the strength of Agentforce's own ROI evidence to justify a fair rate, not an inflated commitment — deployments report ROI within 4–6 weeks, around 20% lower total cost of ownership using pre-built workflows, and reference savings that for some customers reach into the tens of millions annually. Those numbers support a confident baseline; they are not a reason to over-commit on day one. Anchor the deal in our AI procurement guide and request a confidential briefing before you sign a PreCommit.
Governing the Spend After Signature
A consumption deal is only as good as the monitoring behind it. Salesforce's Digital Wallet gives real-time, granular insight into credit consumption — and Flex Credits surface more detail there than the Conversations model — so use it from day one rather than discovering overruns on the invoice. Assign a named owner for the credit budget, set internal alert thresholds well below your entitlement, and review consumption weekly through the first quarter while usage patterns settle and feature creep is most likely.
The discipline is the same one we apply to Copilot Studio message pricing and every other metered AI product: govern before the spend, not after the bill. Treat the first quarter as a measurement exercise, feed the real action-per-conversation data back into your commitment level at the next true-up window, and the meter becomes a number you control rather than one that controls your budget.