👤 For pipe admins
🔐 Available on all plans
🎯 For those who want to eliminate repetitive manual work from the process
Every operation has work that repeats. Reading a request and deciding which team to route it to. Analyzing a document and filling fields with the extracted information. Moving a card to the next phase when the candidate profile matches the job criteria. Tasks that do not require human judgment but consume human time every day.
Pipefy AI Agents were created to take on exactly this work. By the end of this article, you will know which tasks in your process an agent can eliminate, when it makes sense to use an agent instead of a regular automation, and how to structure the configuration before activating anything.
📖 What you will understand here:
Regular automation versus AI Agent: when to use each
A regular automation runs a fixed rule. If the card enters phase X, move it to Y. If the field equals Z, send an email. The logic is exact: the condition is either true or false. It works very well when the process is predictable and the criteria do not change.
An AI Agent handles what the fixed rule cannot: situations that require interpretation.
✅ Reading a request written in natural language and classifying it by urgency.
✅ Comparing a resume with the requirements of a job opening and deciding whether the candidate advances.
✅ Extracting data from a contract in PDF and distributing the information into the correct fields of the card.
The practical distinction is this: automations run instructions. AI Agents make decisions based on context.
In Pipefy, automations and AI Agents work together. The automation triggers the agent at the right moment. The agent does the analysis. The result feeds the next step of the process.
What an AI Agent can do in your process
An agent operates from three elements: a trigger, an instruction written in natural language and the actions it runs when it makes a decision. You do not need to know how to program. The instruction works like guidance for a colleague: the clearer the context, the better the result.
The actions available today are:
- Fill card fields with information extracted or generated by the AI
- Create a new card in any pipe
- Move the card to another phase based on the result of the analysis
- Create a record in a database
- Create a card connected to another process
To make decisions more accurately, the agent can consult two knowledge sources beyond the card itself: documents you attach in the configuration (PDFs of policies, manuals, job descriptions) and data from other pipes or databases.
Three use cases to apply today
The examples below start from real processes. In each one, the starting point is not the feature: it is the time the team spends repeating a task that could be automatic.
1. Ticket triage by urgency and category
An operations team receives tickets by form. The text varies, the format varies, the urgency is rarely stated in a standardized way. The agent reads the description field, classifies the ticket by category and urgency level, and fills the corresponding fields in the card. The analyst opens the card with the context already organized, without needing to read and reclassify each request manually.
2. Data extraction in procurement processes
A purchase request includes the upload of an invoice or commercial proposal. The agent reads the PDF, extracts amount, supplier and due date, and fills the card fields. What took two minutes per request becomes instant, at any volume.
3. Candidate qualification in recruiting
The agent compares the candidate's resume with the job criteria stored in the knowledge base. If the profile is a match, it creates a card in the interviews pipe with name and email filled in and moves the original card to the "Under review" phase. The recruiter only enters the process when the human decision truly matters.
How to structure an AI Agent that actually works
Creating an agent is fast. What defines whether it will save time or generate rework is the quality of the decisions made before saving the configuration. This section covers the four choices that most impact the result.
Choice 1: define the trigger by the right moment, not the easiest one
The trigger determines when the agent acts. The temptation is to choose "when a card is created" because it seems comprehensive. The problem is that agents triggered too early work with incomplete information.
The right question is: at what point in the process is the information the agent needs already available in the card? If the agent needs to read a PDF attached by the requester, the trigger should be the card's arrival in a specific phase, not creation.
Choice 2: give context before giving instruction
The Knowledge section is where you define what the agent knows before it acts. An agent without knowledge operates only with what is in the card. An agent with a well-configured knowledge base makes decisions aligned with the real rules of your process.
Two types of source are available: documents (PDFs of policies, job descriptions, manuals) and data from other pipes or databases. Use documents for static knowledge that rarely changes. Use pipe data for context that updates frequently, such as customer history or product catalog.
Long, generic documents hurt accuracy. If the refund policy has 40 pages, extract the relevant rules into a smaller, specific document. The agent processes up to 30 pages per file.
Choice 3: write the instruction as if it were for a new colleague
The instruction is what the agent will follow. The more specific it is, the more predictable the result. Three practices that make a difference:
Name the fields with the same name that appears in the pipe. If the field is called "Request type", use exactly that name in the instruction. The agent uses the field names as part of its reasoning.
Say what to do in each scenario, not just the main one. If the profile is a match, create the card. If not, fill the "Status" field with "Profile out of scope". Agents without an instruction for the negative path frequently take no action when the main condition is not met.
Use examples inside the instruction when the criterion is subjective. "Classify the urgency as High if the requester mentions a deadline under 24 hours or a risk of immediate financial impact" is far more effective than "classify by urgency".
Each behavior supports up to 3 actions and 10,000 characters of instruction. If the process requires more than 3 actions for one trigger, split it into a second behavior or combine it with a complementary automation.
Choice 4: monitor before you trust
In the first runs, open the cards where the agent acted and check the Activities tab. It shows what the agent did, field by field. If the result is not what you expected, the adjustment is almost always in the instruction, not in the technical configuration.
An agent that runs the wrong action frequently is a sign of an ambiguous instruction, not a limitation of the AI. Refine the prompt before changing the trigger or knowledge.
How AI credits work in practice
Each AI Agent run consumes 2 AI credits. If the agent processes documents (PDF, PNG, JPG), consumption rises to 3 credits per run, due to intelligent document reading.
All plans include 1,000 one-time credits. Additional credits can be purchased in blocks of 500, 1,000, 10,000 or 50,000.
If the trigger condition is not met, the agent does not run and no credit is consumed. Configuring precise filters on the trigger, besides making the agent smarter, protects the balance.
Track consumption at: Admin panel > Usage statistics > AI credits
What an AI Agent does not solve
AI Agents are not suited for mathematical calculations. If the process needs to calculate values, the automated formulas feature is the correct path.
AI Agents do not replace human judgment in high-risk decisions. Pipefy lets you configure a manual review point before the agent runs a critical action, keeping control where it is needed.
A well-configured agent does not make the process more complex. It removes the repetitive part so the team focuses energy on the decisions that need a person.
Before moving on, confirm that you understand:
☐ The difference between a rule-based automation and an AI Agent
☐ Which actions an agent can run inside a pipe
☐ How AI credit consumption works in practice
☐ Which repetitive task in your process an agent could take on today
