Does the agent execute anything on its own?

No. It builds the proposal and stops. Execution only happens after a click from a person with administrator permission.

Only those with a configured administrator role. Being in the conversation gives nobody authority.

Which actions are available?

Opening tickets and tasks, delivered and running in production. The next ones will be listed when they ship in the product.

AI actions with human approval

Three proposals waiting for you.

This is how the agent delivers work: ready, detailed and stopped. Decide each one below and watch the trail build up alongside.

Talk to an AI expert →See it in the home page simulator

agent queue · awaiting authority3 pending

Inventory agentmonitors coverage and turnover · #a41c

WAITING

Open a restocking ticket SKU-1001 · Runner Pro sneaker · 90 units → Morumbi store coverage 0 days · turnover 95 units/30d

Tax agentchecks invoices and accrual period · #b07e

WAITING

Create a review task 12 invoices with a CFOP mismatch · period 08/2026 suggested owner · tax

Sales agenttracks credit limit and receivables · #c3d9

WAITING

Open a credit review ticket customer 4471 · limit exceeded on 3 orders in a row impact · R$ 84,200 in receivables

Trail · what your decision produced

nothing yet. every click above becomes a line here.

The design choice

There are three possible stances. Two do not work.

Stance 1

The agent executes on its own

Fast and impossible to defend. When something goes wrong, there is nobody in the middle to explain the decision, and the company finds out afterwards.

Stance 2 · ours

The agent builds and stops

It does the whole job, including gathering the context and checking the policy. What it does not do is press the button. That still belongs to whoever has authority.

Stance 3

The agent only answers

Safe and useless. Someone reads the answer and redoes everything by hand, which brings back the human error automation was there to avoid.

Who approves

Human approval is not any human.

Who can approve what is configured by role. Being in the conversation gives nobody authority. It is the same rule that decides what each area sees, applied at the moment of acting. And a rejection is recorded just the same: knowing that someone looked and said no is worth as much as knowing they said yes.

Where it acts today: opening tickets and tasks, delivered and running in production. Just one. We would rather name one that works than list five that are on the roadmap.

Want to see the agent asking permission on your data?

Context Diagnostic: a conversation about a real proposal from your workflow and, if it makes sense, a day with your team.

Talk to an AI expert →