Should I Let AI Change My Amazon Bids?

Someone on a tactical gear maker’s team asked me a hard question at an AI lunch-and-learn I gave for their company. If the best way to use AI is to let agents work across our computers, what about the sensitive information on those computers? Where are the boundaries?

That’s governance, I told them. Then I told them how I pay vendors.

I tell my agent: pay this vendor $1,000 from my bank account.

It opens a browser window I can watch and goes to the bank. It reaches the login screen.

And it stops.

The password never reaches the AI. It isn’t typed by the agent, and it isn’t saved anywhere. I sign in myself.

Once I’m in, the agent does the tedious part. It finds the vendor, enters the amount, fills in the details. Then it reaches the confirmation screen, the last one before the money moves.

And it stops again.

The money hasn’t moved. The agent is waiting for me.

I read the screen. The vendor. The amount. The account.

I say, “Go for it.”

Only then does it click.

Someone in the room said, “So you had to go in and build those guardrails yourself.”

Exactly.

I learned that the hard way

I didn’t start there. When I first gave agents real access, I did what most people do. I wrote the rules down in plain English, in the instructions the agent reads every session: draft emails, never send them. Never post to Slack for me.

My agents sent emails anyway, when I’d told them to only draft. They posted messages to my team I never asked for.

The rules were right there. They got overridden in the middle of a task.

So I moved them out of the instructions and into code that blocks the action outright. The agent can ask to send, and the answer is no until I say yes. That held.

That’s the difference I wish someone had told me sooner. A rule the AI reads is a request. A rule the AI can’t get past is a guardrail.

I used to think the question was whether the AI was good enough to trust. It isn’t. The question is where it’s forced to stop.

The same question, with your Amazon account

For an Amazon seller, the version of that question is: should I let AI change my bids, budgets or listings?

It’s not hypothetical. Our MCP Server connects AI tools like Claude to a seller’s Amazon data, and it can make live advertising changes: bids, daily budgets, turning campaigns and keywords on or off, adding negative keywords.

I do it on my own store. In one session, the AI audited the ad account, found the search terms burning money with no orders, and added 83 negative keywords across the campaigns that were wasting it. Work I’d been putting off for weeks took one sitting.

But the change list had a judgment call buried in it. One brand name was a zero-order search term inside one set of campaigns, and one of the better-converting terms across the rest of the account. Block it everywhere and you cut off sales. Block it only where it was wasting money and you keep them. That’s not a call I want made without anyone reading it.

That’s the whole job of a guardrail. The AI does the hours of looking. A person reads the change before it touches money.

Six guardrails before AI touches your account

1. Start read-only. Let the AI analyze your account and recommend changes for a few weeks before it can make any. You’ll learn where it’s wrong while being wrong costs nothing. Most of the value is in the analysis anyway.
2. If it’s a fixed rule, make it a rule. “Raise the budget 20% during Prime Day” or “pause any keyword that spends $30 with zero orders” doesn’t need AI. Amazon Ads has budget rules and rule-based bidding built in, and a rule is cheaper, predictable and easy to audit. Save the AI for judgment: the brand term that wastes money in one campaign and converts in the rest.
3. Stop at the review screen. Have the AI write out every change before it makes one: what, the old value, the new value, and why. Then you say go. Asking for the change authorizes the prep, not the click.
4. Put the limits where the AI can’t argue with them. Telling the AI “never raise a budget over $100” is a request. Settings the AI can’t override are a guardrail. At the time of writing, Claude lets you set each connected tool to Always allow, Needs approval, or Blocked. A sane start: let it read freely, and make every tool that changes something need your approval.
5. The AI never gets the password. Never paste your Seller Central login into a chat. Connect through a sign-in built for it, so the AI gets access without ever holding the keys. When the AI has to work a website the way you would, like Seller Central or your bank, use a browser you sign in to yourself. That’s how my vendor payments work. The Playwright MCP I use is free and open source: it opens a real Chrome window you can watch, works with your password manager, and masks anything that looks like a password before it reaches the AI. Give every person their own login through Seller Central’s User Permissions, never a shared master login.
6. Widen slowly, and narrowly. Trust is earned one task at a time. After months of approving every transfer myself, I let my agent move money between two of my own business accounts each week without asking. Only that pair of accounts, only the full balance, and only after it checks five conditions first. Everything else still stops and waits.

One warning on the flip side. If your AI asks permission for everything, you’ll stop reading the prompts, and a prompt you click through without reading isn’t a guardrail. Let the reads run. Save the stops for changes. Some AI tools can have a second AI review each action and block the dangerous ones, or save what you’ve already approved so it stops asking three times for the same thing. Use them, so that when it does stop, you look.

If you’re an agency

Everything above gets bigger when the accounts aren’t yours.

An AI with write access to ten client accounts can repeat one bad instruction ten times before anyone notices. So:

• Work one client account per session. Never give the AI write access across every client at once.
• Get it in writing. Agree with each client on what the AI may change, how far, and what needs their sign-off.
• Keep the change list. Every change the AI makes, with the before and after. It’s your audit trail, and the best report you can hand a client.
• Stay read-only on any account without that agreement.
• Ask clients to add you through User Permissions. It protects you as much as them.

One more thing for business owners and agencies alike. When the approved tools lag behind, people don’t stop using AI. They use their personal accounts, and paste in data you’d never approve. Give your team a safe, sanctioned way to use it, or they’ll build their own.

The screen where it stops

The people in that room didn’t need a better AI. They needed to know where it would stop.

So do you, before it touches your bids.

I’ve never trusted an AI with my money. I’ve never needed to.

I trust the screen where it has to stop and wait for me.