How it works
How an AI agent carries out a task
Despite the name, an AI agent does practical work: it carries out the authorized steps of a task and asks for human intervention when it encounters a decision reserved for a person, an exception, or a boundary the company has set.
- 01Agent
The request comes in
The task can arrive by message, form, or email, sometimes with incomplete information. The agent identifies the request and what information is missing.
In this scenarioA customer emails asking for a proposal for 40 licenses of the Enterprise plan but does not provide payment terms.
A store sends a WhatsApp message: "I need 200 boxes of item 4471 by Thursday." The message arrives without a tax ID or delivery address.
Every Monday, leadership expects the weekly sales report. The data sits in three different systems.
- 02Agent
Consults the authorized sources
The agent accesses only the sources the company has authorized, according to the permissions on record for that process.
In this scenarioThe customer's history in the CRM, the current price in the 2026 price list, and the terms of the previous contract.
The customer database, which identifies the store by phone number; inventory at the distribution center; and the delivery route schedule.
Billing data in the ERP, the sales pipeline in the CRM, and the spreadsheet of targets by region.
- 03Agent
Applies the company's rules
Deadlines, maximum discount, inventory policy, who can approve what. The company's rules are documented to govern execution and the handling of exceptions.
In this scenarioDiscounts of up to 10% need no approval. Anything above that requires the manager's approval. Standard payment terms are 30 days.
Orders above R$20,000 go through credit review. 48-hour delivery applies only to routes in the capital city. Minimum stock: 50 boxes.
Compare each region against its target. Highlight any that fell below 90%. Exclude returns from the total.
- 04Agents
Drafts the document and another agent checks it
The execution agent prepares the document and links each piece of information to its source. A verifier agent then recalculates the figures, checks the sources and rules, and flags anything that needs attention. Anything it cannot confirm is clearly marked.
In this scenarioA proposal for 40 licenses at R$1,200 each, with a 10% discount and a total of R$43,200. The payment terms are marked as pending.
The order is prepared with the store identified, 200 boxes, and delivery on Wednesday's route. The system flags that it requires credit review.
A report with charts and the source for each figure. A discrepancy of R$38,000 between the ERP and the CRM in the South region is flagged.
- 05Person
A person reviews and decides
Here, the agent stops because this scenario requires human approval. The person responsible reviews it, corrects what is needed, and sets the final terms.
In this scenarioMarina, the sales manager, raises the discount to 12% and sets 45-day payment terms.
Paulo, a credit analyst, checks the store's history and clears the order with payment upfront.
Ana, the controller, checks the discrepancy, confirms it was a canceled order, and adjusts the number.
- 06Record
Delivers and records
After approval, the agent completes the authorized actions. The decision, who made it, and the time of approval are all recorded, making it possible to review the execution later.
In this scenarioThe proposal goes to the customer at 2:32 PM. The three sources, the rule applied, and Marina's approval are all on record.
The order is sent to billing, and the store receives confirmation. The record shows who approved the credit and why.
The report reaches leadership at 8 AM. The adjustment and its approver are both on record.
What changes from one project to another
The steps vary by process, but every project must address the same questions. The process and its risks determine how many steps are needed and where review is required. The available sources and permissions determine what the agent can consult. Responsibilities determine who decides and what gets recorded.