Amazon Just Gave AI Agents the Keys to Your Ad Account (And Honestly, They Might Drive Better Than You)

Okay, I need to talk about something that happened at the IAB Annual Leadership Meeting this month that I cannot stop thinking about. And no, it's not the catering (though conference catering is its own kind of violence against sandwiches). It's about Amazon quietly rolling out a piece of infrastructure that could fundamentally change who — or what — manages your ad campaigns.

Amazon Ads just launched the open beta of its Model Context Protocol server. If your eyes glazed over at "Model Context Protocol," stay with me, because behind that deeply unsexy name is something that should have every seller, agency, and PPC manager paying very close attention. Or updating their LinkedIn profiles. Possibly both.

What Is MCP and Why Should You Care?

Model Context Protocol is an open standard originally developed by Anthropic (the folks behind Claude — hi, colleagues of my ghostwriter). At its core, MCP is basically a universal translator between AI agents and software systems. Instead of building a custom integration every time you want an AI tool to talk to Amazon's advertising APIs — which historically required a developer, several energy drinks, and a tolerance for documentation that reads like it was written by a committee of robots who'd never met a human — the MCP server provides one standardised connection point.

The practical upshot? You can now tell an AI agent something like "create a Sponsored Products campaign for this ASIN with a $5,000 monthly budget" in plain English, and the agent handles the entire sequence of API calls. Campaign creation, ad groups, targeting, budget allocation — the whole thing, without a human clicking through seventeen screens.

Paula Despins, Amazon's VP of Ads Measurement, described it as giving agents an "instruction manual" instead of dumping the entire filing cabinet on their desk. Which is a nice analogy, and also exactly how I wish someone had onboarded me at my first job instead of handing me a 200-page process document and saying "you'll figure it out." (Reader, I did not figure it out.)

When AI Agents Go Rogue (In the Most Boring Way Possible)

Here's where it gets genuinely fascinating — and by "fascinating" I mean "the kind of thing that makes you laugh nervously and then immediately check your ad account."

Amazon revealed what happened during internal testing when they let AI agents loose on their raw API documentation without the MCP guardrails. In one test, an agent tasked with generating a simple path-to-conversion report decided that the standard reporting tool was beneath it. Instead, it wrote its own code and ran it against three years of historical data in Amazon Marketing Cloud. Did it produce a usable result? Yes. Did it cost significantly more compute and time than necessary? Also yes.

It's like asking an intern to grab you a coffee and they come back three hours later having sourced single-origin beans from a micro-roastery in Guatemala, roasted them in-house, and built a pour-over stand from reclaimed wood. Impressive? Technically. What you wanted? Absolutely not.

In a separate test, an agent defaulted to a deprecated API version from 2023. Which is the AI equivalent of showing up to a 2026 meeting with a fax machine and genuine confidence.

These aren't funny edge cases (okay, they're a little funny). They're characteristic behaviours of large language models in complex environments — over-thoroughness when given too much freedom, and a genuine inability to distinguish between current documentation and legacy stuff that should've been deleted years ago. The MCP server essentially puts bumper rails on the bowling lane so the AI stops throwing gutter balls into your ad budget.

What This Actually Means for Sellers

For those of us running multi-ASIN catalogues across multiple Amazon markets (and quietly losing the will to live while doing it), this is a meaningful shift. Tasks that currently require an analyst who speaks fluent SQL, or a developer coordinating multiple API calls while you hover anxiously behind their desk, now become accessible through conversational prompts.

The MCP server works with Claude, ChatGPT, Gemini, and custom-built agents — so which AI tool you're using matters less than whether it can connect to the protocol. It's a bit like how it doesn't matter if you drive a Toyota or a BMW, as long as both can use the same motorway. (The BMW will tailgate you regardless, but that's a different newsletter.)

Amazon has also built in pre-packaged "tools" — templated workflows for the stuff you do constantly. One constructs a complete campaign from a minimal prompt. Another migrates high-performing keywords across campaigns. A third handles geographic expansion, so if you're already running campaigns in the US and Canada, a single prompt can replicate them for a new market. As someone who has manually duplicated campaign structures across markets while questioning every life choice that led me to that moment, this feels almost offensively convenient.

And here's the bit that smaller sellers and lean agencies will care about most: the server auto-updates as Amazon's API catalogue evolves. If you've ever maintained a deep API integration, you know that the initial build is the fun part — it's the ongoing maintenance that slowly erodes your soul. That burden effectively disappears.

The Bigger Picture (Because There's Always a Bigger Picture)

Amazon isn't doing this out of the goodness of their corporate heart. This fits neatly into a much larger strategic play. Andy Jassy said last year that AI agents could meaningfully expand how consumers shop online. Amazon's advertising business — which grew 24% year-over-year to $17.7 billion — sits right at the intersection of AI-mediated discovery and commerce infrastructure. Building an MCP server that lets AI agents manage ad campaigns with minimal human input isn't just a nice developer tool. It's laying the railway tracks for a future where the boundaries between advertising, fulfilment, and AI-powered shopping blur into something we don't quite have a name for yet. ("Adshopalgorithmapocalypse" is available if anyone wants it.)

There's also a standards war brewing. A competing protocol called AdCP was announced last year, and we're entering what could become a protocol standardisation period — think of it like the early days of programmatic display when everyone was fighting over OpenRTB. Whoever controls the key integration points controls how AI-mediated advertising evolves. And right now, Amazon is making a pretty aggressive play for that position.

The Bottom Line for Sellers

Here's the uncomfortable truth that I keep coming back to: this isn't really about whether AI agents will become standard in Amazon advertising. That's happening. The question is whether you adopt them before or after your competitors do, and whether "after" still leaves you with enough market share to care about.

The MCP server meaningfully lowers the barrier to entry for agentic campaign management. You don't need a dedicated engineering team anymore. You don't need to maintain custom API integrations. You need an AI agent, a connection to the MCP server, and the strategic judgment to point it in the right direction.

That last part — the strategic judgment — is the bit that's still firmly in human territory. AI can execute campaign builds at machine speed, but it still can't look at your product catalogue and intuit that your bestselling yoga mat appeals to a completely different customer than your competitor's nearly identical yoga mat. That nuance, that understanding of your customer beyond what the data says? That's yours to keep.

At least until the next IAB meeting. At which point, honestly, all bets are off.

P.S. If you're still manually duplicating campaign structures across markets in 2026, I'm not judging you, but I am looking at you with the same expression my partner gives me when I say "I'll just check one more thing" at 11pm and emerge at 2am having reorganised my entire campaign taxonomy. We both know better. Act accordingly.

• Apple could quietly win AI by default as native apps are easier to agent-build, its ecosystem feeds premium data like health and messages, and Mac minis become the go-to hardware for personal agents.

• Claude can now build full slide decks by itself using Gamma and Google Drive, pulling your data, visuals, and brand style automatically, a glimpse of how AI-native workflows should feel in 2026.

• Newsletters are surging as a core owned channel, with open rates climbing past 41%, weekly sends winning attention, and clarity, consistency, and respect beating urgency tricks as social platforms grow unstable.

• Gemini is testing a ChatGPT chat import feature, letting users carry over long-running context, but the convenience comes with data tradeoffs since imported chats feed directly into Google’s model training.

• A major study shows AI brand rankings are wildly inconsistent run to run, making “AI rank tracking” misleading, but repeated prompts can still reveal meaningful visibility patterns across large samples.

• Reddit says AI-powered search could become a major growth and revenue driver, with AI Answers users exploding from 1M to 15M in a year as it blends traditional search with generative results.

• OpenClaw shows the future of AI agents is powerful but dangerous, automating real-world tasks like car buying while exposing massive security risks, signaling big productivity gains are coming before safety is ready.

📈 Zupyak: Do SEO research and create high-ranking content with AI, without using expensive agencies and contractors. Used by 400,000+ professionals.

🗒️ Intent by Upflowy - Turn your leads' behaviour into AI Summaries.

🗣️ izTalk - Overcome language barriers with instant AI translation

🌐 Permar - Generate a conversion-rate-optimized landing page with a single prompt.

🪄 Snack Prompt - Magic: An optimization tool that upgrades your AI workflow with integrations with Bard, Claude, Midjourney, & more.

🖊️ Anyword AI: Instantly analyze every piece of content you've ever published, so you know which messaging works best on your website, socials, and email channels.

😍It was great, all I needed

🙂Good, not great

🙄It sucked