👤 For people configuring AI Agents and monitoring operations
🔐 Concepts apply across plans, subject to the resources and credits available in your account. Pipe admins configure agents; company admins and super admins access usage statistics.
🎯 Estimate consumption and evaluate changes against cost and quality criteria
Finance receives reimbursement requests every day and needs to check receipts and company policy before sending them for approval. Adding an AI Agent raises another decision: how many credits should the team plan for each month, and how can it tell whether that consumption is producing the expected result?
By the end of this article, you will be able to estimate consumption for each behavior, choose a model to match the task, and compare credit use with the quality of the review. The reimbursement example gives you a way to apply that reasoning to a process at your company.
In the training session IA sem surpresas, delivered in Portuguese, I used expense triage to connect configuration, testing, and consumption monitoring.
Keep learning with Pipefy training
Join upcoming sessions and bring your questions to our specialists. Explore the training page in Portuguese.
What you will learn:
What causes credits to be consumed
A behavior brings together the event that starts the work, any applicable conditions, and the agent’s instructions. To forecast consumption, identify when it actually needs to run: for every new request, for example, or only for requests that meet a triage criterion.
AI Credits are consumed when an AI execution runs. A trigger without an execution does not consume AI Credits. Counting every card in the pipe can therefore overestimate the volume for a behavior that handles only some requests.
Calculate consumption for each behavior
The model is selected per behavior. If triage and the manager’s summary are separate behaviors, estimate the volume and cost of each and add the results. One card can involve more than one AI execution.
How to choose a model and skills
Choose the capability needed to deliver an acceptable result. Classifying structured data calls for a different assessment from reading receipts and applying exceptions. The figures below are the base consumption per behavior execution, before additional charges.
| Model | AI Credits | Selection criterion |
|---|---|---|
| Lite | 1 | Short, predictable tasks with little ambiguity |
| Standard | 2 | Operational analysis involving documents and conditions |
| Pro | 3 | Complex work requiring deeper analysis |
Skills add capabilities to a behavior and may add consumption when used. Document analysis adds 1 credit; web search and calculations or analysis add 0.5 credits each. Check the model and consumption reference before finalizing your estimate.
In the session’s reimbursement example, the agent needed to read the receipt and compare the expense with company policy. Standard with document analysis was the configuration used to explain an estimate of 3 credits per execution. That figure depends on this combination and on the skill being used.
In practice, test the same sample with the options suited to the task. Check the amount, date, classification, and expected routing. A model that uses fewer credits but requires more corrections may not meet the process requirements.
How to estimate monthly consumption
Use this relationship: expected behavior executions during the month × credits per execution. Then add the behaviors and agents included in your process. A forecast in credits does not, by itself, tell you the monetary cost under your contract.
| Triage scenario | Calculation | Monthly forecast |
|---|---|---|
| 500 requests reviewed at 3 credits each | 500 × 3 | 1,500 credits |
| 200 requests meet the criterion, at 3 credits each | 200 × 3 | 600 credits |
The first scenario reproduces the calculation presented in the session. The second shows how to revise it when only some requests require analysis. Use the expected execution volume, including any further executions of the behavior on the same card.
If another behavior summarizes the request for the manager, include it in the forecast. Record your assumptions so that you can later explain whether a difference came from volume, configuration, or execution count.
How to monitor cost and quality
Use two complementary views. In the Admin Panel, the AI Credits tab under Usage Stats helps admins and super admins monitor the period’s total. Filters for pipe, tag, and status help identify where consumption is concentrated.
To understand an individual case, review the agent’s logs. The status, card, and triggered behavior help you investigate what happened. Compare the recorded result with what Finance expected for that request, including routing for human review when required.
Where cost visibility is available in your account, use the test estimate and the consumption shown in the execution summary to compare configurations. The Community explains cost visibility before and after execution. This view is being rolled out gradually.
Agree on a recurring review with the person responsible for credits. If the total rises, first check whether request volume, executions per request, or configuration cost increased. This helps you adjust consumption while keeping necessary analysis in place.
Questions answered during the training
Does testing a behavior show what it would cost?
Yes, in accounts where cost visibility is available. The test shows an estimate and does not consume AI Credits. Use the same cases to compare configurations before activating the behavior for operational use.
Do testing, activation, or unmet trigger conditions consume credits?
Testing does not consume AI Credits, and activating an agent alone is not an execution. Without an execution, there is no consumption. One qualification matters: a partial error can still consume credits when part of the work was completed. Do not assume that every failure means there is no charge.
How should I choose a model for different documents and rules?
Assess complexity, input quality, and the analysis required. Test common cases and exceptions against the correct policy. PDFs have a limit of 30 pages per file; choosing a more capable model does not remove input limits. Check the documented formats and limits, in Portuguese. Every model still requires quality checks.
How do I know an optimization worked and avoid exceeding the forecast?
Compare the same cases before and after the change: credits per execution, review accuracy, and the need for human checking. Documentation provides for email alerts to admins and super admins starting at 80% of the contracted limit. A consumption alert is not a configured cap for an individual agent. Coordinate monitoring with your account administrator and see usage limit notifications, in Portuguese.
What to review in your next behavior
Choose a behavior the team already understands. Record its trigger, which requests require analysis, the model, the skills used, and the expected result. Estimate monthly consumption and compare a sample of results before expanding use. Monitor the first few weeks to revise your assumptions against operational data.
Before moving on, confirm that you understand:
☐ What makes each behavior run and which volume belongs in the forecast.
☐ How the model and skills used contribute to consumption.
☐ Why one card can generate more than one execution.
☐ How to compare quality with credits and who monitors the account limit.
Follow Pipefy on social media
Keep up with content, news, and tips on processes, automation, and AI: LinkedIn | Instagram Brasil | Instagram Global | YouTube | X
With your estimate ready, the next step is to inspect the outcome of real cases. Read AI Agent logs and tracing, in Portuguese, and use one request to connect the recorded decision with the expected result.

