Amazon says two different things about how much control you keep.
The Ads Agent product page: “Ads Agent summarizes all proposed changes before taking action, leaving you in control of what changes are made. After your review and approval, Ads Agent makes the specified changes.”
Amazon’s own guide to agentic AI: these are systems that “autonomously make decisions, take actions, and adapt to achieve specific goals with minimal human intervention”, operating “without waiting for human approval on every action, though oversight mechanisms remain in place for critical decisions.”
Both pages are Amazon’s. They describe opposite defaults.
That contradiction is the honest starting point for this topic, because it tells you the category is being marketed faster than it’s being defined. Here’s what’s actually verifiable.
What software can genuinely do to your account
The capability question has a precise answer, because Amazon publishes the API surface.
Listings. The Listings Items API lets software create, fully replace, patch or delete a listing. Rate limit: 5 requests per second. Modifiable fields include price, condition, fulfilment channel, quantity, images, shipping dimensions and weight, brand, item name, and parent-child variation relationships. A+ Content is separately editable through its own API.
One documented limit worth knowing: “Only top-level listings item attributes can be patched. Patching nested attributes is not supported.” So a tool cannot surgically change one nested value — it has to replace the whole attribute block, which is where badly-built automation destroys data it never meant to touch.
Campaigns. In December 2025 Amazon consolidated its campaign management API. Its own description of the scale of that change: “Previously, there were nearly 200 separate variations of create, read, update, and delete endpoints that will be reduced to 16 with this launch.” It covers Campaign, Ad Group, Target, Ad and Ad Association across DSP, Sponsored Brands, Sponsored Products, Sponsored Display and Sponsored TV.
Rate limits are “dynamic and based on system load” rather than published, returning 429 with a Retry-After header.
Access is gated. The Ads API requires either Amazon Ads Partner Network approval or direct advertiser status, plus explicit authorisation from the advertiser. SP-API access is scoped by role — Product Listing, Pricing, Inventory, Brand Analytics and so on. A tool cannot do anything you haven’t granted.
So: yes, software can genuinely edit listings and campaigns. That was true before anyone said “agent”.
The actual difference between rule-based and agentic
Here’s the distinction that matters, stripped of marketing.
Rule-based automation is a deterministic IF/THEN, evaluated on a schedule against a metric, with the decision boundary specified in advance by a human. “If ACoS on this target exceeds 40% over 14 days with more than 20 clicks, reduce bid by 10%.” A person decided the threshold, the metric, the action and the size of the action.
Agentic automation gives a language model a goal in natural language and lets it decide, at runtime, which tools to call, in what order, with what values.
Both hit the same endpoints. Both call patchListingsItem and the Campaign Management API, at the same rate limits, under the same role approvals. Nothing new is permitted.
What changed is who composes the call. That’s the entire technical difference, and it’s worth stating plainly because most coverage implies agents have capabilities that rule-based tools lack. They don’t. They have discretion that rule-based tools lack.
Discretion cuts both ways. It means an agent can respond to a situation nobody wrote a rule for. It also means it can take an action nobody would have authorised, for a reason nobody can reconstruct afterwards.
Amazon’s own agents, and what they claim
| Tool | Launched | Availability |
|---|---|---|
| Generative listing creation | Pre-2024, expanded since | Broad |
| Project Amelia | 19 Sep 2024 | US beta |
| Seller Assistant (agentic) | 17 Sep 2025 | US store, no additional cost |
| Ads Agent | 11 Nov 2025 | 20 countries including the UK, no cost |
| Creative Agent | EU5 including UK, 24 Feb 2026 | Within Creative Studio, no cost |
| Amazon Ads MCP Server | Open beta 18 Feb 2026 | Three regional endpoints |
Ads Agent eligibility is worth noting: you need access to Amazon Marketing Cloud or Multimedia Solutions with Amazon DSP. Since AMC became free to all sponsored ads advertisers in September 2025, that gate is lower than it looks.
Seller Assistant‘s autonomy language is careful: it can “reason, plan, and help take action with a seller’s permission”, and “if the seller approves, Seller Assistant implements the solution.”
The MCP server is the most concrete evidence of what’s coming. It exposes campaign management tools to external AI clients — Amazon names Kiro CLI, Claude Code, Claude Desktop and ChatGPT as supported. Amazon’s own documentation shows a configuration example for disabling tools named campaign_management-delete_campaign and campaign_management-delete_ad, which tells you those tools exist and are enabled by default. Amazon’s own example prompts include “Increase my campaign budget to $500 on [campaign_id]” and “Create a Sponsored Products campaign in the US and Canada for [ASIN] with a $20 budget.”
And a disclaimer that deserves attention: “Amazon Ads does not own or operate the LLMs or AI applications that connect through the MCP Server.“
The performance numbers, and who produced them
Amazon publishes figures for its own tools. All self-reported:
- Ads Agent: “65% of beta advertisers saw delivery improvements when using Ads Agent for campaign targeting recommendations, with an average of 18% lower CPM and 16% lower CPA.” Population: US beta advertisers, 2025.
- Performance+: 34% ROAS boost for brands selling on Amazon.
- Brand+: 71% more detail page views, 64% more purchases.
- AI listing tools: 900,000+ sellers using them, with sellers accepting AI content with minimal edits around 90% of the time.
Take them as vendor claims from a vendor with unusually good data. Then note what’s missing.
There is no independent, controlled study comparing AI-agent-managed against manually-managed Amazon advertising. Not one. Every number in circulation comes from Amazon or from a software vendor selling an agent. If that seems like a small point, consider that the entire category is being sold on an efficiency claim nobody outside the sellers of it has tested.
The compliance change nobody sent you an email about
This is the part with real consequences, and it’s had almost no coverage.
Amazon updated its Business Solutions Agreement effective 4 March 2026, introducing a new Agent Policy. It adds “Agent” as a defined term and sets three requirements. AI agents must:
- Clearly identify themselves as automated systems
- Comply with the new Agent Policy at all times
- Cease access if Amazon requests
Amazon’s framing: “We’ll add requirements for the use of automated software or AI agents to access Amazon Services, and we may restrict their access in certain instances.”
The same update added restrictions on using Amazon materials to train AI or develop machine learning systems, strengthened anti-reverse-engineering language, and added a new Section 20 on dispute resolution.
Acceptance is passive: “Your continued use of Selling Services after the effective date constitutes acceptance of these changes.”
So if you run any automated tooling against your Amazon account, you have already accepted this policy, and the tooling needs to comply. That’s worth an email to your vendors.
One thing to ignore. Several blog posts describe this policy in far more detail — 12-month audit log retention, mandatory SP-API registration for every automated action, documented human authorisation checkpoints above defined thresholds. None of that appears in Amazon’s announcement or in the trade press reporting on it. It’s invented detail. Don’t build a compliance process around it.
Questions worth asking a vendor
If you’re evaluating anything calling itself an AI agent:
“Which endpoints do you write to, and under which SP-API or Ads API roles?” A vendor who can’t answer precisely isn’t doing what they claim, or hasn’t thought about your risk.
“Does it act by default, or propose by default?” Amazon’s own two pages disagree on this for Amazon’s own product. Assume nothing.
“How do I reconstruct why it did something three weeks ago?” With rule-based automation, the rule is the explanation. With an agent, you need a log of the reasoning, not just the action.
“How do you comply with the Agent Policy effective 4 March 2026?” A vendor who hasn’t heard of it hasn’t read the agreement governing their own access.
“What happens when two of your automations write to the same entity at once?” Not hypothetical — Amazon’s DSP bid adjustments API now returns a specific concurrent-modification error for exactly this collision. If you run more than one tool, they will contend.
Where this genuinely helps
I’m not arguing against the category. The useful cases are real and specific.
The best example is cross-functional reasoning that no rule anticipates. A sales velocity spike is unambiguously good to a bid-management rule, which raises bids to capture it. A system reading the same spike against inventory levels and supplier lead time may conclude a stockout is coming inside two weeks, and that stocking out will cost more organic rank than the extra orders are worth, so it throttles spend instead.
No rule-based tool does that, because the rule would have had to be written by someone who imagined that exact interaction between advertising and supply chain. It’s the same class of cross-domain reasoning that a proper account audit does by hand, and the same reasoning a marketplace expansion decision needs.
That’s the real promise. It’s also the reason the oversight question matters more here than with rule-based tools: a system capable of reasoning its way to a good decision nobody specified is equally capable of reasoning its way to a bad one.
FAQ
What is the difference between an AI agent and a rule-based PPC tool? Both call the same Amazon API endpoints under the same permissions. A rule-based tool executes IF/THEN logic a human specified in advance. An agent is given a goal in natural language and decides at runtime which calls to make and with what values. The difference is discretion, not capability.
Can AI tools actually change my Amazon listings? Yes, through the Listings Items API, at 5 requests per second, if you’ve granted the relevant role. Nested attributes cannot be patched individually, so tools replace whole attribute blocks, which is how poorly-built automation overwrites data unintentionally.
What is Amazon’s Agent Policy? A Business Solutions Agreement update effective 4 March 2026 requiring AI agents to identify themselves as automated systems, comply with the policy, and cease access on Amazon’s request. Continued use of selling services after that date constitutes acceptance.
Is Amazon Ads Agent available in the UK? Yes. It launched on 11 November 2025 across 20 countries including the UK, at no cost, requiring access to Amazon Marketing Cloud or Multimedia Solutions with Amazon DSP.
Do AI agents outperform manual Amazon PPC management? No independent study has tested this. Amazon reports that 65% of Ads Agent beta advertisers saw delivery improvements, with average CPM down 18% and CPA down 16%, but that is Amazon’s own data on Amazon’s own product.
Sami Sultan runs Ecom Enable, an Amazon growth partner for DTC brands. API capabilities, product launch dates and the Agent Policy verified against Amazon developer and Seller Central documentation on 27 August 2026.
Sources: Amazon SP-API Listings Items and roles; Amazon Ads campaign management API; Amazon Ads MCP Server; Ads Agent; Amazon on agentic AI; Seller Assistant; BSA update and Agent Policy, March 2026.