What Two Years of Building AI into a Warehouse Has Taught Us

Building AI into warehouse and fulfilment operations wan’t the primary goal of Helm according to Jade Mills, it was solving real world problems and it’s revolutionising the lives of warehouse managers, from pre-built warehouse performance dashboards to on the fly reports and app building.

Helm are approaching new warehouse tools one problem at a time and with the power of AI it’s transforming operations in ways that simply wouldn’t have been possible just a couple of short years ago: A warehouse manager always knows exactly what they need and exactly how long it used to take to get it.

Two years ago, “AI in the warehouse” meant demand forecasting and the occasional pilot with a mobile robot. Today, a warehouse manager can describe a report they need in a single sentence and watch it get built live, against real stock data, in about the time it takes to make a coffee.

That pace of change has surprised almost everyone in this industry, us included, and it’s worth sharing what we’ve learned from building into it steadily rather than chasing a single big launch. We’ve been shipping AI into warehouse and fulfilment operations for two years now, and the thing that’s changed fastest isn’t any single feature.

It’s what a good operator can reasonably get done in an afternoon. We think that pace is one of the most exciting things happening in eCommerce right now, and we’ve tried to build our way into it deliberately, one real problem at a time. We Didn’t Start With AI. We Started With the Problem.

Every AI feature we’ve shipped has come from watching an operation get stuck on the same problem repeatedly, not from a decision to “use more AI”. That distinction has shaped everything we’ve built.

We started with the boring end of automation, because that’s where the pain actually was: a rule engine to take repetitive decisions off someone’s plate, automatically selecting shipping services, flagging orders that needed a human look, catching and fixing address errors before they became failed deliveries.

It didn’t get much of a launch moment at the time. It was us quietly fixing the thing that was costing our customers money, which is a less exciting story than a product announcement, but a genuinely useful one if you’re the person running the floor.

The AI assistant came next, built for a specific reason: frontline teams were spending hours a day answering the same handful of questions. “Where’s my order?” Returns policy. Tracking updates. We didn’t set out to build something impressive. We set out to give people their afternoons back, and that’s what it’s done.

The step we’re proudest of is the one that changes the category rather than just improving it. Earlier this year we launched AI App Building, powered by Claude through the Helm AI MCP Server, our Model Context Protocol connection to live warehouse data.

We built it because we kept watching the same scene play out: a warehouse manager knows exactly what report they need, stock movement by SKU and shift, put-away times by zone, an exceptions screen showing which orders need a decision before they can move.

A shift handover report, the kind that tells the next team exactly where the warehouse stands, used to mean joining a development queue and waiting for something that’s out of date by the time it lands. Now they describe it in plain language, Claude builds it against live Helm WMS data, and it’s tested in a sandbox before it goes anywhere near production.

It’s live in minutes rather than sprints. It’s the clearest example we have of AI genuinely growing what a small operations team can do without growing the team. Where We Go From Here The next step for us is doing more of exactly this.

Pre-built warehouse performance dashboards are already in development: the kind of view that today only exists if someone has had the time to build it by hand. It’s the same gap we closed for one-off reporting, but closed this time for every operation from day one, not just the ones who happened to ask.

Every one of those gaps is an opportunity for the AI’s role to grow a little further, and for the people running the operation to get a little more of their time back.

We’re genuinely encouraged by how quickly the wider industry is moving with us on this, even though we know we don’t show up often enough yet when buyers ask AI tools which warehouse platforms to consider. Closing that gap matters as much as closing the ones inside the warehouse, and it’s a good sign for everyone building in this space, not just for us.

The businesses getting the most out of it right now tend to be the ones building steadily. One real problem solved properly, then the next. And we’d rather be one of those than chase the biggest possible headline in a single quarter. That’s the approach we’re sticking with: build for the next real problem, prove it works, then build for the one after that.

Slower than a single big launch, but it’s how you end up with AI that an operation actually trusts enough to keep