Amazon has spent the past year building tools that let AI agents work directly inside its advertising systems. This week at unBoxed, that turned into something concrete: a new way for AI tools to plug into Amazon’s advertising data, and a specific new capability – already live in Pacvue Agent – that changes how fast a brand or agency can get an answer out of Amazon Marketing Cloud.
That work was the subject of a session at unBoxed featuring Amazon Ads’ Liz Joyce and Pacvue Chief Product Officer Sunava Dutta. Here’s what was covered in that session, what Amazon announced, what it actually means for your team, and how Pacvue is putting it to work for you today.
What Amazon Announced
For years, getting the most value out of Amazon’s advertising data meant relying on a specialist who could translate a business question into the right technical format and hand back a report. Amazon’s Ads MCP Server, which opened for beta in February, is built to remove that translation step. It lets AI tools connect directly to Amazon’s advertising systems and act on your behalf, instead of every company having to build that connection from scratch one piece at a time.
What Amazon announced at unBoxed was the first of what it calls “Expert Tools” available through the MCP Server: the AMC SQL Generator. In simple terms, the AMC SQL Generator takes a plain-language business question and turns it into a ready-to-run query.
What does that mean for your team directly? In short, it lets you ask a business question the way you’d ask a colleague (i.e., “how are new-to-brand customers responding to our streaming ads this quarter?”) and get back a validated answer without anyone needing to know how to write a query.
How Pacvue is Putting Amazon’s MCP Server To Work
This unlock is huge for enterprise brands and agencies, because AMC has always held rich answers to full-funnel questions. But reaching those answers has usually meant waiting on a data team, adapting an existing template, or bringing in outside help. Amazon lowering that barrier means more of the people who actually own the business question can go get the answer themselves.
Pacvue joined the beta for this capability early, and it’s already live inside Pacvue Agent. Pacvue Agent acts as the connecting layer, taking your plain-language question, routing it to Amazon’s Expert Tools, and returning the insight and ready-to-run query. Together they turn “I have a question” into “here’s your answer” in one step.
The results of this pairing speak for themselves: answers to AMC questions that used to take hours or days now come back up to 80 times faster, and the time spent gathering the underlying query has dropped by roughly 90%. In practice, that’s the difference between waiting on a report next week and getting an answer during the meeting where the question first came up.
It’s also worth noting what this doesn’t change, because speed should not come at the cost of oversight. A person is still in control of anything that touches your live campaigns. The AMC SQL Generator and Pacvue Agent work synergistically to surface an insight, write the underlying query, and even build an audience. But before that audience gets added to an active campaign, someone on your team has to confirm it and decide the timing.
On Stage at unBoxed
At this year’s unBoxed, Pacvue Chief Product Officer Sunava Dutta joined Amazon Ads’ Liz Joyce on stage to talk through what it actually takes to put an AI agent to work on a brand’s advertising data. Their conversation highlighted some of the biggest considerations – and potential mistakes – brands should keep in mind when evaluating and working with AI agents:
- Most AI agents fail because of access, not intelligence. Liz opened by naming a pattern showing up across the industry. Most companies that experiment with AI agents never get them to deliver lasting value, and the reason is almost always that the agent can’t reach the right data and context, not that the underlying AI isn’t smart enough. The same AI, pointed at your actual account data versus not, produces two very different experiences. That’s a useful filter the next time you’re evaluating a new AI ad tool. The question worth asking isn’t “how advanced is the model?”; it’s “what can it actually see?”
- Getting an answer out of AMC no longer requires a fully formed question. With a combo like Amazon’s AMC SQL Generator and Pacvue, what used to follow one of two paths (clicking through a rigid set of filters or writing SQL) can now happen as a conversation, where teams work alongside an AI agent that Sunava described as acting as a sort of “sparring partner.” Instead of arriving with a fully formed question, you can go back and forth with the agent to sharpen what you’re actually trying to figure out, then explore from there.
- The most common mistake teams make is vague requests. Asked what he sees most often, Sunava pointed to too-simple prompts, like “help me increase my ROAS for this campaign.” That leaves out the details that really matter (what counts as a win, what’s off-limits, what assumptions to make) so the agent has to guess. Sometimes that produces an answer that’s obviously off. But other times, it can produce an answer that looks right but isn’t, and you may not see the difference until later. Sunava’s advice is simple: be specific about what you’re asking and keep a saved set of instructions and assumptions you’ve already tested, so you’re not starting from scratch with every request.
- Reviewing an agent’s output isn’t just one person’s job. Sunava described it as a split according to whoever is closest to each kind of decision. The tech provider checks that that agent itself is working correctly, whoever holds budget authority owns anything touching spend, and whoever’s running the campaign day-to-day owns performance calls.
- Teams trust AI to think, not to spend. As Sunava put it, “we’ve become comfortable letting agents think for us, but we’re not quite there letting agents spend.” He went on to add that this is the right instinct for where the industry actually is right now. Liz offered an analogy that captures it perfectly: an agent is a lot like a new team member. You wouldn’t hand someone full access to your dashboards and reporting on day one and tell them to go make a million-dollar decision, and the same logic applies to an AI agent handling your ad spend.
Liz closed the session with four pieces of advice for building agentic workflows. Translated into what they mean for brands adopting those workflows, they’re worth keeping on hand as a quick checklist before rolling anything out to your team:
- Get your own data in order first. Any agent, no matter how capable, is only as useful as what it can actually see. Make sure your account structure and campaign history are clean before you expect a sharp answer.
- Decide what “better” looks like before you start. Know whether you’re trying to save analyst hours, catch things faster, or surface insights you’re currently missing entirely.
- Set your own rules before you need them. Decide as a team what you want an agent to do on its own versus what should always need a person’s sign-off, before you’re deciding it under pressure in the middle of a campaign.
- Don’t accept a black box. Any tool you adopt, Pacvue’s or otherwise, should be able to show you why it did what it did, not just hand you the output.
What This Means for Your Team
If you’ve tried plugging an AI tool into your ad data and gotten back something generic or watched an early pilot fizzle out after the initial demo, you’re not alone. That’s the norm right now, not the exception, and it usually comes down to the agent not having access to the right data and context about your specific business. By combining forces, Amazon’s AMC SQL Generator and Pacvue Agent close that gap. One brings deep AMC expertise, and the other brings the account context to apply it to, so what comes back is a real answer about what’s really happening in your business.
To see this in action inside Pacvue Agent, book a demo.
To learn more about AMC SQL Generator and all of the announcements Amazon made at unBoxed this year, join us for unBoxed unPacked: Keynote Breakdown & Live AMA on October 7. Register here.
