AI chat on a dashboard vs an assistant that can act on your ads

What is the difference between a chat box on a dashboard and an assistant that acts?

A chat box on a dashboard answers questions about what the dashboard already shows. An assistant that acts is an AI model you already use, such as Claude or ChatGPT, connected to your ad account so it can read the data and, if you allow it, change the account. Xnurta MCP connects Claude, ChatGPT and other MCP clients to your Xnurta data. That is the second kind.

The gap is not intelligence. Both can write a decent summary of last week. The gap is what happens after you say "do it."

What makes an assistant that acts safe to run on a live account?

Five things: separate write permissions, a preview before money moves, an honest rule about what cannot be undone, a log of what happened, and a way to say no. If a tool that can change your account cannot show you all five, keep it read only. The table is a checklist you can hold any vendor to, and the right-hand column shows how Xnurta MCP answers each line.

How does Xnurta MCP split write access?

It splits writing into three kinds, and each kind has its own permission scope. The kinds are managed group management, ad-entity editing, and Sponsored Products campaign creation. A scope that edits ads does not necessarily create campaigns.

That separation is the first thing to look for in any tool that acts. One broad "write" switch means a token meant for a small bid tweak can also build new campaigns. Separate scopes mean you grant what the job needs and nothing else.

Atomic means all or nothing. If one campaign in a request fails validation, none of them are created, so you never end up with a half-built account.

What does the preview show before money, bids or archive change?

It shows which objects will change and from what value to what value, and nothing runs until you confirm. This applies to anything involving money, bids or archiving. The assistant does not get to skip it.

Here is what that looks like in practice. You ask for a budget raise. The assistant lists each campaign, its current daily budget and the new one. You read the list. You say yes, or you decline. Only then does it submit.

You can also ask for the same discipline on every kind of write. Tell the assistant to list the objects and the exact values it plans to change, and to wait for your confirmation. That works for managed group changes, new keywords and negatives too.

In ad account [account name], raise the daily budget on campaigns [campaign A], [campaign B] and [campaign C] by 10%. Before you change anything, show each campaign with its current budget and the new budget. Do not submit until I confirm.

Why does archiving need its own rule?

Because archiving cannot be undone, and in Xnurta MCP there is no way back. Deleting an AI managed group archives it, and that is irreversible. Campaigns cannot be deleted at all. They can only be archived, and that is irreversible too.

So the rule is simple: if you want spending to stop for now, pause. Archive only when you mean it. When you create campaigns, you choose the state they start in, so create them paused if you want to review before they spend.

Where do you see who changed what in the account?

In the operation log, which you can query from the same chat. It holds human and AI actions on campaigns, ad groups, targets, bids, budgets and managed groups. You can filter by operator (the AI, a human, an automation rule or one person's email), by action type, by entity type and by time window, and the log reaches back about 15 months. Reading it is read only.

A dashboard with a chat box might show you a number that moved. A log answers the follow-up: who moved it, and when? Xnurta ships an official skill for this, Amazon Ads Operation Log Query, and the full walk-through is in Who changed my Amazon campaign?.

In ad account [account name], who changed campaign [campaign name or ID]'s budget, and when? List all human actions from the last [30] days.

How do you stay in control once it can act?

You control it at four points: access, scope, the preview and the pause button. None of them depends on trusting the AI to behave.

1. Access. Write access is off until it is turned on for your account. Reading never changes anything, and queries on performance, settings and the log are read only.
2. Scope. If you connect with a token, create it with the scope you need. A scope that edits ads is a different thing from one that creates campaigns.
3. The preview. Read every one. It is the point of the whole design.
4. Pause, not archive. Pausing is how you stop something and still get it back.

Your access also follows your own Xnurta login. You see only the ad accounts and permissions that login carries, main accounts and sub-accounts included. Read-only vs read-write MCP for Amazon advertisers explains which kind of access to start with.

Which one should you choose?

Choose by the job. If you only want answers, either kind can do it, and with Xnurta MCP the reading side is read only, so starting there does not change your account. If you want the assistant to carry out the work you currently do by hand in the console, you need the second kind, and then the five checks above decide whether you can trust it.

A practical way to start: connect, ask questions for a week, then turn on writing for one narrow job, such as budget changes on a few campaigns, and read every preview. The endpoint for Claude, ChatGPT and other MCP clients is https://mcp.xnurta.com/mcp, and Connect Claude or ChatGPT to Xnurta MCP has the steps. For how this compares with working through an API or a dashboard alone, see MCP vs API vs dashboard for retail media teams.