Edited from a conversation with James Gurd on the Inside Commerce podcast, episode 363.

The world is obsessed with AI and what it's doing to shopper behaviour. Martin, you've said the interesting part is somewhere else.

Shopper behaviour is evolving, and we see that across our merchants — more careful spending, more consideration before buying, more time spent shopping with chatbots before deciding. A continued shift from quantity to quality. And that consumer behaviour drives efficiency pressure internally at brands. But it's evolution, not revolution.

Where we actually see the revolution is not on the outside. It's on the inside — what's happening inside the brand. Six months ago, agents were like interns on a triple espresso: incredibly fast, but not really useful for doing real work. Today, a fairly large share of daily e-commerce operations work is agent-doable. That's where the revolution is happening.

What counts as real work?

Merchandising: writing product text, translation, tagging products correctly, quality-controlling product entries, managing campaigns. Order management: someone wants to edit an order, cancellations, refunds, an order is stuck and nobody knows why. Customer operations: the most common question our merchants get is where is my order, and the second is how is my return going. And B2B — retail orders, restocks, invoice questions.

The agents can do that work today. What has been lacking is an agentic system for them to work in. An agent doesn't sit and click around in an admin UI. An agent needs tools built for agents. So we've released around 300 tools that agents can use, live in production, so they can go in and do real operational work.

For people who are new to the terminology — what actually is an MCP server?

MCP is an open standard for how you expose tools to an AI agent.

As humans, we interact with software by clicking around a user interface. Right now we're recording a podcast, and there's an interface with buttons: pause my audio, pause my video, leave the recording. An MCP is that set of tools, for an agent. It's not a user interface — it's language-based, where the agent reads and interacts with the software through text. An MCP version of this recording software might have a tool for the agent to leave the podcast. But there's no door icon. There's a text description saying: here's a tool you can use if you want to leave the recording.

That's where the world is heading. Offering good tools that agents can understand, adapted to how an agent thinks, is really the key to making agents productive and safe to use.

How big a task can you actually hand to an agent?

We're at a point where you can give a larger task to an agent. By larger I don't mean translate this product description from English to Japanese. I mean handle the localisation of the entire site for the Japanese market. The agent picks up that complex task, breaks it into sub-tasks, and spins up sub-agents that carry out the work. Technology-wise, we're there now.

The open question is the human one. As an e-com operator, how do I hand over managing a long-tail market to an AI colleague? Where do I actually sit — a desktop app, a specialised tool? How do I handle follow-up? Do I invite the agent to my board meeting? Probably not. But how does this work when agents are joining the workforce?

You've started talking internally about agent experience. What does that mean?

It's a big shift we've made. Before, it was about customer experience, user experience, developer experience. Now we really need to talk about the agent experience.

It's a cultural shift as much as a technical one: we don't get stuck on how the admin panel looks, because the agents are not using the admin panel. It's completely fine to ship a feature where the look and feel isn't the most important thing — maybe it's more important that the agent understands how to use it.

An agent thinks very differently from a human. We can't take the tools we give humans, hand them to agents, and call it an agent version. As humans we recall things from long ago. Agents don't: either something is in the prompt, or it isn't, and if it isn't, the agent has absolutely no idea. As humans we can scroll quickly through large amounts of information and find what we need. If you dump a huge document into an agent, you get context rot and the agent becomes completely dumb. So the tools have to be built for how an agent actually works.

And the second half of that is the playbooks.

Think of the agent as, in some regards, a brilliant colleague — and in many regards really, really stupid. We want deterministic output, and that's hard. The way you get it is strict playbooks. If you're processing a refund request, this is how you do it: step by step, double-check this, double-check that. Imagine managing a junior colleague in an operations team. You don't let them think too much. You give them a playbook. Then you aggregate those playbooks into handling larger and larger tasks.

We don't believe the agents should live only in Centra — we think that's wrong. The agents work for the brand, and like any other employee they can work across multiple tools: figure something out in one, take action in another. There's a big and important job in building playbooks and agents that work across the whole stack, and a real opportunity for agencies to lead there.

How do you stop an agent going rogue?

The framework is still evolving and nobody knows exactly what it will look like. But there are three types of guardrails.

First, permissions. What the agents should and shouldn't have access to — a hard stop. Exactly what they need and nothing else. There's a second reason to limit access, too: an agent's context gets eaten up by things it doesn't need to know, and it gets less smart. So there are performance reasons as well as safety ones.

Second, hard controls. Programmatic limits that stop the agent if it starts doing something stupid — a lot of edits, or reaching for bulk tools it shouldn't use.

Third, and this is the one that's become most popular: you run another model to check the main agent's work. Before each tool use, a second agent looks at what the first is trying to do. Does this seem correct, given what the user asked? Should it be allowed? If it's a no, it gets a no. If it's a maybe, it escalates to the human to decide.

What does this mean for the operations teams doing that work today?

It's about trust and permissions. You wouldn't hire an intern and give them access to everything without double-checking anything. It's the same with agents, and how you bring them on board.

But what it means is that operations teams doing fairly repeatable, scriptable work — those jobs will be automated by agents. We believe the volumes are going to move there, and it's moving now.

We're experimenting with a few handfuls of clients right now. No one knows exactly how this is going to work. The technology is there, the agents are smart enough, we've built the tools. The question is what it looks like in the workplace — what it means when agents are working side by side with humans to operate a brand. It changes everything.

Martin Jensen is co-founder and Chief Innovation Officer at Centra. Listen to the full episode of Inside Commerce with James Gurd.

Ready to put AI agents to work in your eCommerce team? Download our guide Agentic Operations for Fashion Commerce and explore 80+ playbooks. want to go further? Learn more about Centra MCP.