Can You Trust an AI Agent to Run Your Amazon Ads? What Amazon Shipped, and What It Did Not Require

Until recently, letting software decide something in your ad account meant a bid rule you wrote yourself. Now Amazon ships a conversational assistant inside its own Ads console, and it has opened a door that lets outside software create, edit and delete your campaigns from plain-English instructions.

Whether that is safe depends less on how good the software is than on two things Amazon published: what it required of its own tool, and what it required of everyone else's. Those two answers are not the same.

Amazon Ads Agent: what does it do, and when did it ship

Ads Agent was announced at unBoxed on 11 November 2025. It sits inside the Amazon Ads console and takes instructions in plain English. Pause all campaigns with ROAS under 2. ROAS is return on ad spend, revenue divided by what the ad cost. From that it plans, launches and adjusts campaigns, writes SQL for Amazon Marketing Cloud queries, and moves pacing across hundreds of campaigns. At launch it covered every Amazon Marketing Cloud user across more than 30 countries, plus Multimedia Solutions with Amazon DSP in the US.

Amazon's own product copy carries the line that matters more than anything on that list. Campaigns only launch after you review and approve. Targeting recommendations, in Amazon's wording, still have to be reviewed and applied by the person using it. Amazon's own agent drafts, then waits.

Creative Agent decides the ad creative, not your bids

Different product, and the one people confuse with the first. Amazon Ads Creative Studio launched on 15 October 2024, and its chat layer, Creative Agent, was announced in September 2025, reached US advertisers late that year, and kept rolling out with Full Funnel Campaigns in Q1 2026. It assembles images, video storyboards and ad copy drafts. It does not set bids, budgets or targeting.

The AMC lookback window went from 13 months to 25 months

Announced the same day. Amazon Marketing Cloud is the clean room where you query your own ad and conversion events, and its ad traffic lookback window doubled on 11 November 2025, live in the US and Canada that day, with seven more countries in Q1 2026. History is what any agent reasons from. Twenty five months lets a query compare this Q4 against the last two. Thirteen never did.

What an outside AI agent can change in your ad account

On 2 February 2026 Amazon Ads put its MCP Server into open beta, announced at the IAB Annual Leadership Meeting. MCP is the Model Context Protocol, an open standard from Anthropic that lets an outside tool call a service's functions conversationally rather than through hand-written API calls. Amazon's version connects agents built on Claude, ChatGPT, Gemini or an agency's own stack to the Amazon Advertising APIs.

What Amazon lists as supported: creating, updating or deleting campaigns; running performance and reporting queries; managing account-level settings; and accessing billing and financial data. Read that list slowly. Delete is on it. So is billing.

Eligibility is narrow, limited to Amazon Ads partners holding active API credentials, so this is a lane for agencies and tool vendors rather than a switch a seller flips. That does not put it far from you. It puts it one contract away.

What an agent can read is capped by report and by role

Write access takes the attention. Read access is where the real ceilings sit, and they are uneven enough to change what any tool can conclude.

Search Query Performance, the Brand Analytics report showing real query volume and your share of impressions, clicks, cart adds and purchases at each step, requires Brand Registry plus a Brand Representative role. No Brand Registry, no report, and no amount of reasoning gets around that.

Report history splits by ad type too. Sponsored Products data is retrievable for roughly 13 months through one reporting endpoint. Sponsored Display runs about 6 months through a different one. Ask a tool to compare this year against last on Sponsored Display and it works from half a period, however good the model on top. Both figures come from industry documentation rather than a single Amazon page, so hold the months loosely and the shape firmly.

Why your Amazon ACoS changes a week after you read the report

This is the part that decides whether speed is worth anything.

A click on Amazon stays credited to an ad for a set period afterwards, and that period is the click attribution window. On a seller account, Sponsored Products runs 7 days. Sponsored Brands and Sponsored Display run 14. A purchase made on day six from a click on day one gets written back against day one, after the fact.

So the report you opened this morning is not finished. Amazon's general guidance is to treat the most recent 48 hours as partial, and past that, any day keeps filling in until its window closes. ACoS, advertising cost of sales, is spend divided by ad-attributed sales, and the bottom half of that fraction is still moving. A term that looks wasteful on Monday can read differently by the following Monday without anybody touching it.

Now put the two facts together. A tool with write access can change a bid or delete a campaign in seconds. The data justifying that change has not finished forming. The faster the tool, the further ahead of the truth its decisions run.

That is not a risk agents invented. A rushed operator has always had it. What changed is the pause, because a person opening a report at least stops to look. Seller Central threads from roughly 2024, well before any product named here existed, complain that handing bid control to Amazon eventually damages a campaign. Dated craft worry about automated bidding, not evidence about today's tools.

The two 2026 rules that touch an AI agent in your Seller Central account

Amazon AI agent policy, effective 4 March 2026

Section 19 of the Business Solutions Agreement, the Agent Policy, was announced by Amazon's official News_Amazon account on the Seller Central Forums on 17 February 2026 and took effect on 4 March 2026. Any automated software or agent touching Amazon Services has to identify itself clearly as an automated system and comply at all times. If Amazon asks it to stop accessing Amazon Services, it stops.

Two details get missed. Amazon states explicitly that there is no minimum usage threshold, so one repricer counts and so does a profit calculator pulling live data. Acceptance is automatic as well: continuing to sell after 4 March 2026 is acceptance, read or not.

What the policy does not do is govern judgment. It says nothing about data freshness, nothing about review cadence, nothing about the quality of a decision. Quoted as proof that Amazon requires human review of automated decisions, it is being oversold, usually by whoever is selling the automation.

SPP reauthorization by 3 September 2026 caps what a tool can touch

The Seller Central invite process for granting a service provider access was retired on 10 August 2026, and sellers reauthorize service access by 3 September 2026. On reauthorization a service is capped to the Seller Central roles it currently holds approval for on the Solution Provider Portal. Roles outside that set drop, silently. Miss the deadline and access expires entirely.

That ceiling applies to an automated tool exactly as it applies to a person logging in, and it produces a failure worth naming. A tool asked to fix bids across the account can draft the change, request it, and have the request never land because the role is not on file. Nothing errors in a way you would see. The account simply looks unchanged.

Which is the uncomfortable part. A tool blocked from doing anything and a tool that decided to do nothing look identical from outside.

Is Amazon Ads Agent safe? Amazon put an approval step on its own agent

Here is the finding, and it only holds if it is stated exactly.

Amazon's page for Ads Agent says campaigns only launch after you review and approve. That is a documented gate, written into Amazon's own first-party product.

Amazon's page for the MCP Server covers the same class of action for third-party tools. It describes generated campaigns as something that only needs review and approval. It does not require it. No mandatory human gate, and no prohibited-actions list on that page.

Be careful about what that is. Amazon did not forbid anything. Amazon did not require anything of third parties. It built a review step into the tool it shipped itself, and did not write that step into the door it opened for everyone else's. The gap between those two choices is the finding, and anyone reporting it as Amazon banning autonomous agents has broken it.

So the question is not whether Amazon permits a tool to act without review. It plainly does. The question is whether you bought the version with the gate, because that choice now sits with whoever configured your tooling.

Why a fast agent and a slow attribution window are the same problem

One flag is noise. Correlated flags are a finding, and this topic only reads correctly when three separate ceilings get read against each other.

Speed against settling. Write access and the attribution window are one lever seen twice. What makes an agent useful, acting immediately, is what puts it ahead of its own evidence. Slowing a tool down does not close an attribution window, and a longer window does not make a fast decision correct. What can be reconciled is where the review sits, because nothing needs to move before the number underneath it has stopped moving.

Permission against capability. What the MCP Server lets a tool request and what SPP has approved for that tool are different limits, and only the second is enforced inside the account. Ruled out here, and this is the useful part: a silent no-op is not a malfunction in the software, so chasing it as a bug will not find it. The answer is in an approval record, not a log file.

Visibility against reasoning. Search Query Performance needs Brand Registry and a Brand Representative role, so an unbranded or reseller-exposed catalog cannot see it at all, agent or human. Sponsored Display's shorter report history makes a year-on-year comparison there incomplete by construction. That ceiling predates all of this.

Also ruled out: performance figures for Ads Agent. Time-saving and ACoS-improvement percentages circulated on vendor blogs attributed to unnamed beta testers, with no Amazon documentation behind them. They are not in this article.

What you can safely do today with ChatGPT and an exported search term report

With no team and no partner API credentials, you are not running an agent on your account. You are running a chat tool alongside it, which is a different job with a real ceiling.

What works: export the Search Term Report from the Sponsored Ads console under bulk reports, and paste it in with the columns that carry the decision, meaning spend, clicks, orders, sales and ACoS. Ask for patterns rather than actions. Which terms carry spend with no orders. Where conversion falls away against the rest of the account. What a defensible negative list would look like.

Where that stops, it stops hard. A chat tool cannot see your account, your inventory position, your buy box status, your organic rank, or the price a competitor cut on Tuesday. It has no idea which of the rows you pasted are still settling. Every change comes back to you to make by hand, and that manual step is the review gate the MCP Server does not require. Doing it yourself is the version with the gate.

Carry one discipline over: do not act on the last 48 hours, and treat a week of Sponsored Products data as provisional.

If you do hand write access to a tool, specify the check layer before it is connected. Whether an action fired against data still inside its settling window. Whether the change the tool reports making matches the change that actually landed. Whether the roles on file permit what it was asked to do. None of that depends on trusting the tool's own account of itself.

What this decides: whether your agent is a drafter or a spender

Strip out the product names and the dates, and one budget decision is left.

A tool with write access is either drafting or spending. Amazon chose drafting for its own agent and wrote that choice into the product. On the door it opened to everyone else, the choice belongs to whoever set the tool up, and most sellers will not find out which version they bought until a campaign is already gone.

So the money question is not whether to use one. It is what an unreviewed action is worth against what a reviewed one costs. A review step costs latency, and latency has a price in a category where the auction moves daily. An unreviewed step costs whatever the worst single action can spend or delete before anyone opens the console. Price both, then choose. That is a budget decision, not a preference about technology.

The seller who never prices it has not avoided the decision. They made it by default, in whichever direction their vendor happened to ship.

If you want to know what an automated tool is actually permitted to touch in your account, and which of your reports it can even see, that is one of the reads a full account audit pulls apart. You keep the findings either way.

Frequently asked questions

Can I trust AI to run my Amazon ads?

Trust the setup rather than the software. Amazon's own agent drafts campaigns and waits for approval, because Amazon wrote that step into the product. Outside tools connecting through the Amazon Ads MCP Server are not required to include the same step, so whether yours has one depends on who configured it. Ask the vendor directly whether an action can execute without a human approving it, and ask what happens when the underlying data is still inside its attribution window.

Is Amazon Ads Agent safe?

It ships with a documented control. Amazon's own copy says campaigns only launch after you review and approve, and targeting recommendations have to be reviewed and applied by the user. That makes it a drafting tool by design. It went live at unBoxed on 11 November 2025 across Amazon Marketing Cloud users in more than 30 countries, plus Multimedia Solutions with Amazon DSP in the US.

What is the Amazon MCP server?

An infrastructure layer Amazon Ads put into open beta on 2 February 2026, built on the Model Context Protocol, an open standard that lets an outside tool call a service's functions conversationally. It supports creating, updating or deleting campaigns, running performance and reporting queries, managing account-level settings, and accessing billing and financial data. Eligibility is limited to Amazon Ads partners with active API credentials.

What do I need to do about the Amazon AI agent policy in March 2026?

Section 19 of the Business Solutions Agreement took effect on 4 March 2026 and applies to any automated tool touching Amazon Services, with no minimum usage threshold. Continuing to sell after that date is acceptance. Practically, inventory what automated software touches your account and confirm each vendor identifies its tool as automated and can stop access on request. The policy governs identification and compliance, not the quality of the decisions a tool makes.

How many days until Amazon PPC data is accurate?

Treat the most recent 48 hours as partial, per Amazon's general guidance. Beyond that, Sponsored Products runs a 7-day click attribution window on seller accounts, and Sponsored Brands and Sponsored Display run 14 days, so sales keep getting written back against earlier days until those windows close. A week of Sponsored Products numbers is provisional, and any decision made on top of it inherits that.

Sources

Primary (Amazon's own pages and posts)

• unBoxed 2025: introducing Ads Agent, Amazon Ads, 11 November 2025. Launch date, capability list, availability, and the review-and-approve line.

• Expanded ad traffic lookback window, Amazon Ads, 11 November 2025. AMC 13 to 25 months, and the country rollout.

• unBoxed: AI Creative Studio, Amazon Ads. Creative Studio launch, 15 October 2024.

• Amazon Ads agentic AI creative tool, Amazon Ads, September 2025. Creative Agent announcement and scope.

• Introducing the Amazon Ads MCP Server, Amazon Ads, 2 February 2026. Open beta date, supported actions, partner eligibility, and the review-and-approval wording.

• Business Solutions Agreement update, Agent Policy, official News_Amazon post, Seller Central Forums, announced 17 February 2026, effective 4 March 2026.

• Amazon Ads API policies, Amazon Ads. Eligibility rules for API access.

Secondary (trade press and vendor documentation)

• Amazon Ads MCP Server moves to open beta, MediaPost, 2 February 2026. Corroborates the beta date and venue.

• Amazon's new AI agent rules, PPC Land. Corroborates the Agent Policy dates and scope.

• Amazon BSA Agent Policy 2026, SellerSprite. Corroborating overview.

• Amazon Seller Central authorisation process 2026, EcomRanker. SPP reauthorization deadline and the approved-role ceiling.

• Amazon ads data challenges 2026, Improvado. Report retention windows by ad type.

• Amazon attribution window, bidX. Click attribution windows by ad type and account type.

• Amazon advertising attribution, definitions and timing, Openbridge. Corroborates the same windows independently.

• Amazon Search Query Performance report guide, GigaBrands. Brand Registry and Brand Representative requirement.

• Rule-based bidding vs dynamic bids, Seller Central Forums thread, posts dated roughly 2024. Used only as pre-2026 craft sentiment.