Open-source Ecommerce

June 26, 2026

How Ops, Marketing, and Customer Care Do More With AI Agents

AI agentseCommerce operationsHeadless eCommerceMCPopen source eCommerce

The fastest-growing group of AI agent users is not engineers. It is the non-technical staff who never had a safe way to act on the store themselves. Spree Commerce tooling lets those same people harness AI agents to clear the busywork and do more.

Key Takeaways

Last verified: June 2026

Why it matters: Non-technical staff get store work done through Claude Cowork or any AI agent.

What it saves: No need to ask developers or pay in an app store when the admin panel falls short.

What you get: Ask in plain language for anything from a one-off fix to a scheduled job or an integration.

What’s different with Spree Commerce: Because the platform is open source, the agent works from the core team’s documented know-how.

The bottleneck was never the work. It was who was allowed to do it.

Think about what slows a store down on an ordinary Tuesday. A flash sale needs to go live before lunch. A few thousand prices need a one-time correction. Customer service needs to reach everyone caught in a shipping delay. None of it is hard. Most of it still waits.

Some of it the admin handles fine. You can filter orders and export a CSV without anyone’s help. The work that piles up is the bulk changes and the oddly specific pulls the admin makes slow or has no way to express:

• A catalog-wide flash sale. Marking four thousand products down by twenty percent, then back up on Monday, is an afternoon of clicking. Described in a sentence, it takes seconds.
• An audience the built-in filters do not cover. Everyone who bought a hat but not a scarf in the last ninety days, in Germany, for one campaign. That is a precise question, not a saved filter, and today it waits on someone who can write the query.
• The exact list of people hit by a problem. Every customer with an undelivered order that includes the delayed item, so support can email them before they complain. The data is all there. Pulling that one list on demand is the part that stalls.

None of this is hard work. It is work sitting behind a question most of the team has no way to put to the store directly. So it waits.

That gap has nothing to do with talent and everything to do with access. The person who needs the work done is rarely the person allowed to do it. So the store collects small jobs that only engineering can touch, and the people closest to the customer wait.

What changes that is not another dashboard button. It is giving the people who run the business a way to act directly, and giving them an assistant that knows how.

What does “the psychological cost of action” actually mean?

It means the quiet tax you pay before you even start something unfamiliar. The hesitation, the worry about breaking something, the sense that this is really someone else’s job. When that tax is high, people defer. When it drops, they try.

Workplace culture expert Jessica Kriegel named this directly in a June 2026 piece in Axios on the spread of AI agents at work. “Agents are reducing what I’d call the psychological cost of action,” she said. They “make unfamiliar work feel more approachable, which means I start sooner, experiment more, and spend less energy worrying about what I don’t know.”

That same report, built on data from OpenAI and researchers at Columbia, Duke, and the University of Pennsylvania, found that non-developers are the fastest-growing group of agent users. In a sample of people using agents, 80.6% delegated at least one task an experienced person would have spent more than 30 minutes on. A quarter handed off work that would have taken a full day.

Read that as an operations lead, not an engineer. The point is not that agents write code. The point is that the marketer, the merchandiser, and the support manager now attempt work they used to route to someone else. The cost of trying went down, so more got tried.

This is not a fringe habit. In the Stack Overflow 2025 Developer Survey, 84% of people said they use or plan to use AI tools, up from 76% a year before. The behavior is mainstream. The question for a store is whether the platform is built for it.

You talk to the agent. The agent does the work.

Here is what makes this work for someone who has never written a line of code. You do not learn anything new. You say what you want in plain language, the way you would brief a capable new hire, and the assistant does it on your store.

For that to work, the assistant needs two things. It has to understand your store, and it has to be allowed to act on it.

It understands your store by reading guides written by the people who build the platform. These are called agent skills, and they are plain instructions that tell the assistant how your store actually works, so it follows the right steps instead of guessing. The assistant reads the live documentation the same way whenever it needs to check something.

It is allowed to act because you give it a key. You create that key in your store’s settings and decide exactly what it may see and change. The assistant uses the key to do real work, like fixing a price or pulling a report, and it can do nothing the key does not permit.

That is the whole arrangement. You ask in plain language. The assistant, which now understands your store and holds a key you control, gets it done.

What does it take to set this up?

Not much, and you never have to write code. For someone non-technical, the assistant is Claude Cowork, which you use as a chat app. Developers use the same skills inside coding tools like Claude Code, Cursor, and Codex, so the whole team works from one playbook.

It comes down to two things: a key, and an assistant that knows your store.

You create the key in your store, under Settings and then API Keys, and tick exactly what it may do, whether that is reading reports only or also making changes. The assistant can do only what the key allows.

Then you connect Claude Cowork to the live documentation and add the agent skills so it follows the store’s conventions. The agentic setup guide has the documentation address and the exact steps.

Give Cowork the key, and from then on you just ask in plain language. The work gets done.

What can a non-developer get done on the store now?

Almost anything the admin can do, plus the bulk and scheduled work the admin was never good at. With an agent that knows the platform and a permissioned key, the daily grind opens up to the people doing it.

Daily operations, in bulk or on a schedule. Mark four thousand products down by twenty percent before a sale, in one pass instead of an afternoon of clicking. Pull yesterday’s orders for finance every morning without asking anyone. Sweep the catalog overnight and set sold-out variants out of stock before the store opens.

Optimization at scale. Spin up a customer segment, attach a contract or volume price with Price Lists, and launch a segment-specific offer through the promotions engine across thousands of products at once. The merchandiser describes the goal. The agent does the repetitive part.

Reporting and reconciliation. Ask an agent to compile a month of marketplace orders with each vendor’s commission, tax, and shipping broken out, then hand finance a reconciled export. The data already lives in the store. The agent assembles it on request instead of someone rebuilding the spreadsheet by hand.

Each of these used to be either a manual slog or a ticket in the dev queue. Now it is a request, made in plain language, run within limits you set.

Build beyond the core without a custom-development project

The next step up is changing what the store can do, and here too the assistant does the building. Because the platform is open source and every part of it is reachable, you are not boxed in by the features that came in the box, and you are not waiting for an app store to sell you an add-on.

Want a report the admin does not include, a rule for how orders get routed, or a quick tie-in to a tool your team already uses? You describe it, and the assistant builds it against your store, using the same know-how the skills gave it. Small, useful changes that used to mean a development project become a conversation.

It is also the quickest way to test an idea before you commit to it. An assistant can stand up a rough connection to a marketing or accounting tool, on the real store, so you can see whether it earns its place.

If it does, that is the point where a developer turns the prototype into a permanent connection through the Admin API. Working through your assistant is for getting things done and trying things out. A built integration is for the connections you decide to keep.

That is the difference between renting a fixed set of features and owning a system you can shape. The store does what your business needs, not only what a vendor’s roadmap allows.

How do you automate the repeats, and keep it safe?

You hand the recurring work to an agent and scope exactly what it can touch. Ad hoc requests are the start. The real payoff is automating the jobs that come back every day or every week.

A weekly price refresh, a nightly stock sweep, a morning sales report for the team: each becomes a standing instruction your assistant runs on a schedule, with no person typing it each time. The dull, repeatable work moves off people’s plates without anyone giving up control.

Control is the reason this is safe to do. Every access key is limited to exactly the resources you choose and to reading only or reading and writing. A key for pulling reports cannot change a price. In development, access starts read-only by default, so exploring breaks nothing. A repeated command will not double-apply, so a retry will not refund twice. For coding agents specifically, a Claude Code plugin adds safety hooks that block destructive database commands before they run.

So the boundary is yours to draw. More people and more automation can act on the store, while you decide precisely who can do what.

Lower the barrier, and teams try more

This is where it adds up to something bigger than saved time. When acting on the store is cheap and safe, a team stops rationing its ideas. The promotion that was not worth a developer’s week gets tested this afternoon. The segment nobody had time to build gets built and measured.

That is the experimentation dividend, and it compounds. The businesses that pull ahead will not be the ones with the most engineers. They will be the ones where the people closest to the customer can move on an idea the moment they have it.

The pairing is what makes it real: an AI agent that knows the platform, speaking directly to the store on your team’s behalf, inside limits you set. The know-how, the skills, and the interface in one place. That is what drops the cost of action from a project to a sentence.

Lower that cost, and a business stops waiting to find out what works. It just tries.

Want to see what your team could run with an agent? Get started with Spree Commerce, read how agent skills make an assistant a platform expert, or see what an open source Admin API makes possible.

Frequently Asked Questions