Workflow billing
A journey's cost depends on how many runs start, how many steps they execute and where AI is used. Model and metered tool work spends workspace credits; email also draws on the shared sending allowance. Plan these components together before scaling a journey.
What a step costs
Standard internal step execution has no fixed operation fee. Models and metered third-party tools are charged by usage from prepaid credits, and email above the included allowance costs $0.40 per 1,000 recipients.
| Step | Charge |
|---|---|
| Find Records, Set Value, Query, For Each and compiled If/Else, Switch or Delay | No fixed internal step fee. Any paid external service invoked is separate. |
| AI | Model and metered tool usage. |
| Uncompiled If/Else, Switch or Delay | Model usage when the rule is interpreted during a run. Compiling moves interpretation to authoring. |
| Send Email | Recipient sends use the shared allowance; email overage and model-based personalization are separate. |
Charges retain workflow execution context for accounting. In Billing → Usage, an admin sees spend by period, by category (Workflows is one) and by member. The screen does not break it down per workflow, run or step. Use run history to inspect execution volume separately.
Free has a hard cap of 100 recipient sends each month. Premium includes 50,000, plus 10,000 for each paid additional record pack for a full month. Mid-month additions are prorated. Included email does not include AI personalization, and no monthly AI allowance is granted.
Running out of credits
When paid work cannot obtain credits, it waits and checks again. Adding credits can let that waiting work continue; creating a replacement run risks doing the work twice. Premium also allows a limited negative-credit buffer, so a displayed balance at zero is not a precise stop switch.
Credit waits are different from terminal failures such as invalid step data, an unusable sender domain or exhausted email allowance. Topping up does not repair a run already marked Failed. Inspect the run reason first.
Paid model/tool work is recorded for its execution occurrence. Safe recovery reuses settled receipts and recorded effects rather than charging them again. A new run, correction or deliberate new generation can create new paid work.
An unknown model or external tool outcome can place the run in Needs attention, which does not continue automatically. Unresolved AI receipts can retain reserved credits while that outcome is reconciled. A balance change or top-up does not prove whether the action completed, and Terminate run does not refund paid work or reverse an accepted external effect.
Estimate the work that reaches paid steps
Use a planning cohort, then compare it with actual execution and usage. For an illustrative 1,000-signup cohort, suppose your post-delay Query finds 600 people already created a project. A branch that ends their journey leaves 400 people reaching the email and any downstream AI guidance.
- Placing the guidance AI before that branch means it can run for all 1,000 people. Placing it afterward limits that task to the 400 who still need it.
- Adding Adaptyle creates a separate recipient-rendering task for each email requiring generation, even when an earlier AI step already wrote guidance.
- Giving the AI step workspace tools or MCP servers can create several model turns and tool calls per person. The 15-turn model bound is a ceiling, not the expected usage.
- A Find Records selection or For Each list can create child runs, each executing its own later paid steps. Count those occurrences as well as the root cohort.
These counts illustrate where usage arises; they are not predicted customer behavior or a price quote. Model input/output, thinking, tool work and the actual paths determine cost. Preview and authoring usage also spends credits before the workflow runs.
Use the email editor's Personalize → Estimate for its recipient-rendering and sending estimate. It is not a quote for the entire workflow, upstream AI steps, MCP calls or future event branches. Check the email estimate and representative run history.
Concurrent run cap
Each space has a ceiling on runs active at once: 1000 plus 20 per object record. The cap scales with your data, so larger spaces run more in parallel.
A step set to process each record separately creates child runs; For Each creates one per item. Work beyond the active-execution capacity waits for a turn. Durable sleeping delays do not occupy an active execution slot for their whole wait. The cap controls simultaneous execution, not the lifetime total of journeys you may start.
The cap limits pressure on the workspace, but it is not a campaign budget. An indirect loop can still generate continuing work: workflow A fires an event for B, and B updates a record that triggers A again. Design conditions that stop that cycle and review fan-out before publishing.
Record-selection limits are separate from capacity. Find Records can pass at most 1,000 records in one list; separate processing handles larger selections in child runs. For Each accepts at most 1,000 items, and Delete Records rejects a match over 1,000 records. Waiting for a capacity slot does not repair an oversized list or deletion selection.
Control enrollment and execution costs
Stop new runs closes new enrollment while existing runs finish. Pause also parks further execution. Terminate run ends one selected run, its descendants and pending continuations. These controls retain completed effects and incurred usage; none is a refund or campaign budget setting.
After a pause, saved Delay deadlines can already have passed. Resuming or publishing can release waiting execution, including runs on older versions. Inspect those paths and their remaining paid steps before releasing a correction. Publication preserves a separate new-enrollment stop.
Managing usage
Admins on Free and Premium can review usage, buy credits and configure auto reload in Billing & Usage. See AI and usage credits for rates, balances and top-up controls. Buying credits does not raise Free's email allowance.
Related
- Workflows: steps, versioning, and the trigger API.
- Workflow runs: what each run did.