How Do You Forecast Agentforce Flex Credits When Overages Bill in Arrears?

Consumption pricing does not remove budget risk. It moves the risk from the signature date to the close of every month.

Flex Credits are the better meter. They are also the meter that can invoice you after the spend has already happened, on a balance that expires if you do not burn it. That combination is where the Proof Gap opens: the pilot case was built on in-month dashboards, and the finance record arrives weeks later.

This page is the forecast-and-close model. It is not a Claudeforce two-meter page and it does not invent a Salesforce list price. Every dollar figure below is labeled with the source that published it. Your order form beats any blog math, including ours.

The buyer question, stated cleanly

Director of RevOps, revenue finance partner, VP Sales Ops, CRO sponsor at 100+ headcount:

How do we forecast monthly Agentforce Flex Credit burn, set a threshold that stops the pilot, and close the books when the overage invoice lands after the period it belongs to?

Useful answer: model actions, not conversations. Cap the cohort, not the org. Name one owner for the Digital Wallet review. Write the kill threshold before you scale, because the platform will not enforce one for you.

What Flex Credits are, and what they replaced

Salesforce Ben's pain-point explainer describes the old Conversations SKU as billing roughly two dollars per conversation, which punished long or repeated interactions regardless of what the agent actually did. Flex Credits move billing to the action level. Source: Salesforce Ben.

The same explainer publishes the list math Salesforce uses: 20 Flex Credits per action, which works out to about ten cents per action, sold in packs of 100,000 credits for $500, with Agentforce Foundations including 100,000 Flex Credits. Those figures are Salesforce Ben's, not ours. Source: Salesforce Ben.

Aquiva's Flex Credits explainer frames the same entry point as a 100,000 credit minimum at $500, with a standard action at 20 credits (about ten cents) and a voice action at 30 credits (about fifteen cents), citing the Flex Credits Rate Card updated 21 April 2026. Aquiva also notes sandbox usage is billed at 80% of the production rate, which it puts at 16 and 24 credits respectively. Source: Aquiva.

Two more mechanics from Aquiva that change the model more than the unit price does. An org picks Flex Credits or Conversations, not both. And bring-your-own-LLM is still billed per action on the Salesforce side, on top of whatever your LLM provider charges separately. Source: Aquiva.

If a figure is not in one of those two sources, it is not on this page. Ask your AE to put the rate card version and your contracted rate on the order form.

Three failure modes that create the Proof Gap

1. No rollover, so unburned budget is a write-off

Aquiva reports that unused Flex Credits expire at the end of the subscription term and that Salesforce permits no rollover. Source: Aquiva.

The failure is not the expiry. It is the incentive it creates in month eleven. Teams that under-burn start finding uses for credits to avoid an optics hit at renewal. That is Optimization Theater with a meter attached. Activity rises, attribution does not.

2. Arrears invoicing against an in-month dashboard

Salesforce Ben states there are no overage penalties: usage past your balance is billed at the contracted rate. Aquiva adds that overages are billed monthly in arrears at the agreed rate. Sources: Salesforce Ben and Aquiva.

No penalty is good news for procurement and bad news for the close. The month you overspend is the month you report as a success. The correction arrives in the next cycle, against a period the team has already moved on from. If nobody reconciles the invoice back to the Digital Wallet reading for that period, the cost per outcome in your board deck is wrong by construction.

3. Alerts without a kill switch, plus token-split actions

Salesforce Ben notes that Digital Wallet gives real-time tracking and overage alerts, but explicitly has no kill switch to freeze agents when a threshold is hit. The same explainer notes a 10,000 token per action cap, which means a long prompt can be counted as multiple actions. Source: Salesforce Ben.

Read those two together. Your unit forecast assumes one action per task. A verbose prompt template shipped by an enthusiastic admin turns one task into two or three billable actions. The alert fires. Nothing stops. The bill arrives next month.

The monthly forecast model

Run this before the pilot scales, not after the first invoice.

1. Count actions, not conversations. For each agent use case, estimate completed tasks per week and actions per task. Multiply by the credit rate on your rate card, not a blog average. Sandbox work goes in a separate line at the sandbox rate: see the sandbox rate decoder for the published pre-production multiplier.
2. Add a token-split factor. Sample twenty real runs and record actual actions consumed per completed task. Use the measured ratio, not the design assumption.
3. Cap by cohort. Give the pilot cohort a monthly credit ceiling in your own model even though the platform will not enforce it. One cohort, one number, one owner.
4. Review Digital Wallet weekly, same day, same person. RevOps owns the reading. Log burn to date, run rate, and projected month-end against the ceiling.
5. Write the kill threshold before scale. Example shape: if projected month-end burn exceeds the cohort ceiling by more than a set percentage for two consecutive weeks, agents are disabled by the admin until the use case is re-scoped. Because there is no kill switch, that action is a human on a calendar.
6. Separate BYO LLM cost. Salesforce actions and your model provider's bill are two lines with two owners.

The output you want is not a credit number. It is cost per completed revenue outcome, by use case, with the pre-AI baseline next to it. Credits are the input. Attribution is the claim.

Close-the-books checklist when the arrears invoice lands

1. Reconcile the invoice line to the Digital Wallet reading for the same period. Any variance you cannot explain is a question for the AE that week, not at renewal.
2. Restate cost per outcome for that period using the invoiced amount, not the in-month estimate. Update the board number if it moved.
3. Attribute the variance to a cause: volume above plan, token-split actions, sandbox usage, or scope creep into a use case nobody approved.
4. Check burn against term expiry. If you are tracking far under, do not invent usage. Re-scope the commitment at renewal instead.
5. If the invoice kills the pilot business case, open a Reversal Ledger entry with the date, the use case, the invoiced cost, and the decision. Undoing a deployment for a documented cost reason is a receipt, not an embarrassment.

The discipline is boring and it is the whole point. A pilot that cannot survive its own invoice was never a pilot. It was a demo with a budget code.

What this is not

This is not an anti-Agentforce page. Action-based billing is a real improvement over per-conversation pricing, and Digital Wallet visibility is more than most vendors ship.

This is not the Claudeforce two-meter decoder. That covers a different problem: two contracts and two vendors. Read it at Do We Need Two Contracts to Buy Claudeforce.

This is not a price list. Every figure here carries the URL of the secondary explainer that published it. Confirm your rate card version and contracted rate on the order form before anything reaches a budget slide.

Digital Wallet governance, including the alert thresholds and the written kill rule, is covered separately at Does Digital Wallet Stop Agentforce When Credits Run Out.

Related: Claudeforce two contracts, two meters · Claudeforce wait or buy · AI pricing from seats to outcomes · The Reversal Ledger · AI SDR kill criteria · Agentforce and Data 360 share one credit pool