Is the agent safe? Approvals, limits and what it won't do
When the agent asks before it changes your account, in chat and in routines: the approval card and its four buttons, the reminder above the message box, how much one approval covers, where changes are recorded, and what the agent refuses.
TL;DR: In chat, the agent shows an approval card with the exact change before changing anything, until you allow that kind of change. Routines follow Settings › Agent permissions: on Pro, a routine makes any change you have not set yourself, though the page shows Needs approval; choose Needs approval to make it propose. Every Amazon Ads change you make through the agent is recorded in your Action log.
In chat, the agent asks before it changes anything, until you tell it otherwise. Before a change — a bid, a budget, a pause, a new campaign, a goal, a skill or a routine — it shows an approval card with exactly what it will do, and waits for you. Lookups — metrics, campaign lists, keyword data, profit checks — run straight away, unless you have changed that in Agent permissions.
The approval card
The card lays out the change, then offers four buttons:
- Allow once makes this change. Next time, the agent asks again.
- Always allow makes this change and stops asking for this kind of change.
- Block once stops this change. Nothing is applied, and the agent says so and offers to adjust it.
- Always block stops this change and turns this kind of change off.
The two Always buttons save your choice in Settings › Agent permissions. They apply to a kind of change, not one campaign: Always allow on a bid change covers every bid change after it. Your routines follow the same setting. You can set it back at any time — see Agent permissions.
When an approval is waiting
While a card waits, a banner above the message box reads Approval pending and names the change, with Approve and Decline buttons, so you can answer without scrolling back to the card.
Sending a new message declines the waiting change. Nothing is applied. If you still want it, ask again and the agent shows a fresh card.
How much one approval covers
One card covers the whole batch, so you approve 200 bid changes once, not 200 times. Each card has a ceiling:
| One card can | Up to |
|---|---|
| Create campaigns | 20 |
| Pause, enable, rename or re-budget campaigns, or move them between portfolios | 50 |
| Change placement adjustments | 50 campaigns |
| Add keywords, product targets or negatives | 100 |
| Add product ads | 100 |
| Archive campaigns, or archive targets | 100 |
| Change bids | 1,000 targets |
A bigger job takes more than one card. If one item in a batch fails, the rest still go through, and the agent tells you which failed.
Where every change is recorded
Every change the agent makes in your Amazon Ads account is recorded in the Action log: when, on which seller, what was sent and what Amazon replied. That holds whether you asked in the app, a routine made it, or you asked through Claude or ChatGPT. Each person sees their own changes; a teammate's are in theirs. The ad engine's own bidding is not in the Action log; its work shows on Goals and on Today. See The Action log: changes you made through the agent.
When Amazon refuses a change
If Amazon rejects a change you approved, the agent tells you in one plain sentence what was blocked — for example, "That bid is below Amazon's minimum for this marketplace" — and asks what you would like to do. The Action log keeps Amazon's full reply.
What the agent won't do
- Say a change happened when it didn't. If you declined it, blocked it, or Amazon refused it, the agent tells you nothing changed.
- Make up numbers. If it does not have the data, it says so.
- Bring back an archived campaign or target. Archiving is permanent on Amazon. To stop something for now, ask for a pause.
- Change its own permissions. Only you can, in Settings or with a card's Always buttons.
- Show internal IDs. You see product names and ASINs.
- Answer outside Amazon Ads and Autron. It politely declines.
Your Agent permissions cover the agent in the app and your routines, and nothing else. Email, Slack and WhatsApp handle changes their own way, so an action you have blocked can still be offered there for your approval. Each one's article explains how — Email your agent, for example, covers approving by reply. In Claude or ChatGPT, that app decides whether to ask you first. To stop a reply part-way, see Using the chat.
The short version
| What the agent is doing | What happens |
|---|---|
| Looking something up | Runs straight away, unless you have changed that in Agent permissions |
| Any change in chat: campaigns, bids, targets, goals, skills, routines | Approval card first |
| A change you set to Always allow | Runs without a card |
| A change you set to Blocked | Not made in the app or your routines; in chat the agent points you to Agent permissions |
| A change in a routine | Made if allowed, proposed if Needs approval, skipped if Blocked. On Pro, a change you have not set is made — see Agent permissions |