Watch an AI agent read a form, extract structured data, apply simple rules, and take a real action. No rocket science required.
Human view
What a visitor submits
scanning…
scanning…
scanning…
Agent view
What the agent sees and does
agent online
1 Read2 Extract3 Decide4 Act
Waiting for a form submission…
{"name":"Alex Rivera","email":"alex@example.com","topic":"I need a refund for my last order"}
Rules the agent evaluates
IF topic containsTHENRoute to billing team
IF topic containsTHENSend support article
OTHERWISE→Log to CRM
Waiting for a run…
Agent log
Every run the agent has performed
No runs yet. Press Run agent to see the pipeline in action.
Substring rules choose local display actions
Read the explanation
The example topic refund and help contains both configured keywords. The evaluator lowercases the topic and walks rules in order, selecting the first substring match. It therefore picks rule one, refund, and the local billing display. At one hundred pixels per match or selection the bars show two hundred for two matching rules and one hundred for one selected action. This is deterministic string routing, not a model judging the request. The completion panel says Action executed, but the action only renders a card and saves a local run. It does not contact billing, send email, or write a real CRM. Before running, the page requires nonempty name, email, and topic strings. Three populated fields pass that check, while clearing topic leaves two and shows an error. At eighty pixels per populated field the bars shrink from two hundred forty to one hundred sixty. This function does not validate the email syntax or prove the named person owns that address. With valid-looking local input, the pipeline animates field collection, JSON display, rule checking, and a local result. Exporting the run log records these inputs and the selected rule, not an external delivery receipt. If neither of the two keyword rules matches, the evaluator chooses its one default, Log to CRM. At one hundred pixels per rule or action the bars show two hundred and one hundred. This still creates a local result card rather than a CRM write. A blank keyword does not disable a rule: the source substitutes refund for the first and help for the second. Substring matching also means an unrelated larger word can contain a keyword. Inspect the exported matched-rule identifier, keyword, and action to understand the actual branch, and keep the local demonstration separate from live agent behavior.