Where Loopfour uses AI

Understand the specific places Loopfour uses AI, such as chat and the AI Agent step, and the review and approval controls around each one.

Loopfour runs deterministic, auditable workflows. It is not black-box AI: a workflow follows the steps you set, the same way each time. AI is used in five specific places. Each one has controls, and this page lists them.

The five places

Where	What AI does	Your control
Chat, on Home	Builds and changes workflows, explains runs, builds charts, answers questions	You approve before chat runs, deploys or deletes anything
The AI Agent step	Reads a document or text and pulls out or sorts information	You write its instructions and the shape of its answer
The browser step	Follows plain-language instructions in a web portal	You set the instruction and a limit on its actions
Failed-run explanations	Explains in plain words why a run failed	It explains; it does not change your workflows or systems
Approval request emails	Writes the subject and wording of the email sent to an approver	The approvers, the buttons and the decision are not written by AI

Chat

Chat builds a workflow from a template or from your description, changes an existing one, explains a run, and builds tables and charts. It reads these docs to answer how-to questions. See Chat.

Controls around chat:

It asks before it acts. Before chat runs, deploys or deletes a workflow, disconnects a system, or removes a chart from a dashboard, Studio shows Approve and Deny buttons. Typing "yes" is not enough; only the button counts.
New workflows start as drafts. A workflow chat builds is a Draft. It does not run until someone deploys it.
You see what it does. While chat works, it shows each action it takes, such as "Creating workflow" or "Checking integrations".
It uses your connections. Chat cannot connect a system for you. It shows a Connect button, and you sign in yourself.

The AI Agent step

The AI Agent step is one step in a workflow, under AI in the editor's block list. It reads text or a document, such as a vendor bill, and returns the fields you asked for, such as vendor, amount and due date. It can also sort items into groups you name. See Read, extract and classify with AI.

Controls around the AI Agent step:

You set the instructions and the answer's shape. The step's settings hold the instructions and the fields you expect back.
Answers that do not fit are rejected. When you define the fields you expect, Loopfour checks the answer against them. If the answer does not match, the step fails instead of passing bad data on.
It starts from the most predictable setting. The step's temperature setting starts at 0, which keeps answers as consistent as the model allows.
It can only use the tools you list. If you let the step look things up in a connected system, it can use only those tools, with a cap on how many rounds it takes.
You put checks after it. Follow it with a Condition step, for example "amount over $10,000", and an approval step so a person signs off before anything is posted.

TriggerA bill arrives by email
StepAI Agent reads the bill
Returns vendor, amount and due date in the fields you defined.
DecisionOver your approval limit?
ApprovalController approves
OutputBill created in your ledger

The browser step

The browser step, listed under Browser Automation in the editor, works in a web portal that has no direct connection, such as a customer's payment portal. You give it plain-language instructions, and AI decides which buttons and fields to use. See Work in a web portal.

Controls around the browser step:

You write the instruction for each action, such as "download the latest statement".
You cap the number of actions it may take when it works through a longer task.
You can make it check first. The Observe Before Acting setting has it find exactly one action and check it before it acts.

Failed-run explanations

When a run fails, the run page can explain what went wrong and who can fix it. An explanation written by AI carries the label Explained by Loopfour AI. Otherwise the explanation comes from fixed rules.

Controls around failed-run explanations:

It only explains. It does not retry, change or undo anything.
Secrets are removed first. Passwords, keys and similar values are taken out of the run data before it is sent for an explanation.
It says what already happened. The explanation lists what the run already changed in your systems before it failed, so you know whether a retry could repeat it.

See Fix and retry a failed run.

Approval request emails

When an approval step asks someone by email, AI drafts the email's subject and wording from the request, so the approver sees what they are signing off.

Controls around approval emails:

AI writes the wording only. Who is asked, the Approve and Reject buttons, and what happens after the decision are set by the approval step, not by AI.
A plain email is the fallback. If the AI draft cannot be made, Loopfour sends a plain email built from the request instead.
Fixed wording is possible. Your developer can give the step its own email text, and then no AI draft is made.

Where AI is not used

The other steps follow their settings without AI. That includes Condition, Loop, Parallel, Wait and Transform, reading text from a PDF, the approval decision itself, and the steps that read from or write to your connected systems. A Code step runs the code your developer wrote.

For security and data handling, see Security and Legal.