How Nordic Nest built a foundation for AI, humans, and what's next

Start small and fail fast. Avoid large, long-term AI investments, since technology moves too quickly. Run small tests, build on what works and scrap what doesn't. And invest in your data foundation, because an agent is only as good as the data and instructions it works with.

The core challenge is the same: product data lives in different systems and silos, across internal teams, suppliers and manufacturers. Retailers often have it harder, because they depend on suppliers they have no mandate over. Brands can at least orchestrate processes within their own organization.

Introduction: Nordic Nest Group at a glance

Jimmy Ekbäck: Welcome. This is a conversational session about how Nordic Nest built a foundation for AI and humans, and a look at what's next in their operations. But first, let's introduce the star of this session. Jacob, could you please introduce yourself?

Jacob Rastad: Thank you for the warm welcome, Jimmy. I'm very happy to be here. I'm Jacob Rastad, CTO at Nordic Nest Group, and I work with the most fun things you can work with: tech, data, BI, analytics and warehouse operations.

Jimmy Ekbäck: Well said. Jacob and I have been sparring for a long time on the topics we'll discuss today. My name is Jimmy Ekbäck, and I'm the CTO and one of the founders of Occtoo.

Today we'll focus on the way of working with the traditional approach to PIM, Jacob's and Nordic Nest's story of transformation from that old mindset into a new one, the impact, and what's next, including Jacob's advice going forward. This is a conversational session, so please post questions and insights as we go.

Before we jump into the first topic, Jacob, can you give us some insight into Nordic Nest Group and the scale of your operation?

Jacob Rastad: Certainly. Nordic Nest Group consists of three consumer brands. Nordic Nest is the biggest and most established one, in Sweden and other markets. Then we have Svenssons, a premium furniture retailer mainly focused on Sweden and Norway. And we have KitchenTime, a kitchen specialist mainly focused on the Nordics and Germany.

We operate all three brands from Kalmar in Sweden. We have roughly 500 full-time employees, scaling up to around 700 at the end of November. We ship to roughly 70 countries each month, and we have about 90,000 SKUs in our active assortment. On top of that, there are SKUs coming next season and SKUs that have reached end of life this season. So there's a lot of data.

Jimmy Ekbäck: A lot of data, and a lot of different products in the warehouse as well. High complexity in all aspects. One interesting thing you told me: we're coming into campaign season now. How much are you actually moving through the warehouse from Black Week to the end of the holiday season?

Jacob Rastad: A massive amount. Our warehouse in Kalmar is roughly 35,000 square meters, and right now it's fully stacked. From the start of November until a week before Christmas, we're going to empty the warehouse, fill it back up, empty it again and fill it back up again. So it's going to be a busy time.

Jimmy Ekbäck: That gives you a picture of the real logistics, compared to the data logistics we'll be discussing today.

Why think bigger than traditional PIM?

Jimmy Ekbäck: First topic: why think bigger than traditional PIM? Give us your backstory. Why did you move away from what we call the traditional PIM?

Jacob Rastad: My vision for PIM started quite a long time ago. We moved into PIM about six years ago, with one of the bigger brands, and quite soon after that we started to experience a lot of performance issues. My vision was that we wanted all our product data in the PIM. It's a product information management system, right? So all data related to a product should be in the PIM.

That didn't scale. When we grew from 20,000 SKUs to 50,000 to 70,000 and so on, we got a lot of performance issues in what I now call our legacy PIM system.

In a way, traditional PIM systems are a glorified spreadsheet. You collect the data, but you don't get much help as an operator. You have a place to store all the data for a product, but very little help with how to manage, operate and orchestrate that data out into the rest of the organization.

We had these performance issues quite early, and I looked around for several years to find another solution. I haven't looked at all PIMs, but I've looked at a lot of them, and I didn't find a suitable solution for what I wanted to build. All the PIM systems I looked at required us to build a lot of infrastructure around them to implement them, which wasn't what I wanted to do.

Out of that frustration, you and I, Jimmy, had a conversation one late night in Kalmar where I let my frustrations out on you. That's where we sparked the idea of how we could actually work with product data in the next era.

Jimmy Ekbäck: To be fair, you're the reason we ventured into this. We had the data flows and the data logistics, and you asked: what's the problem, why don't you solve the part where you maintain and work with the data? Which is a fair question.

PIM, MDM or something else?

Jimmy Ekbäck: You work with what we call the master product, the bigger picture, with a lot of data that content teams in a traditional PIM can't manage. Give us your take on traditional PIM versus managing the full master product. Some would call that MDM. What actually solves your problem?

Jacob Rastad: What I wanted from a PIM system goes a bit into the MDM space. But MDM is more about compliance and regulation, which we don't need. It's different if you work with pharmacy or similar, but we don't have any large regulatory requirements. So MDM isn't the right fit for us either. It's a system designed to be slow, because it's built for highly regulated environments.

What we have is a data model that's quite complex, and we wanted to work with our data in a way where it can mean different things in different contexts.

Jimmy Ekbäck: Exactly. Your model.

Jacob Rastad: That's what we tried to do in the legacy system, and it got to the point where if one employee wanted to make a change, say update 500 products, they needed to book a time slot with a specific team.

Jimmy Ekbäck: And it went down for 48 hours working through its own queues.

What changed from day one

Jimmy Ekbäck: As you went through the transformation, what were the challenges, and what did you get out of it?

Jacob Rastad: The biggest impact, from day one when we launched with Occtoo, was the complete removal of the performance bottlenecks. That was instant. We didn't need to book a time slot to work in the system at 7 p.m. From day one, we could do the work when we wanted to do the work. That's a huge win on its own, because it removes friction.

When you remove friction, the whole team starts to operate at a much higher pace, and you breed innovation in the team. So it's a really big and important impact.

The second thing, which we'll talk more about later, is how we can work with AI agents in the system. That was unexpected, because we didn't plan for it. The vision you and I set out on this journey was to have an infrastructure that enables people to operate with high efficiency. We have so many different products that we can't expect one person to be an expert on both sofas and home appliances.

Jimmy Ekbäck: And people moving in and out of roles.

Jacob Rastad: Exactly. So we wanted a platform to support that, and that's what we built together. It turns out it's an amazing fit for AI agents as well, because AI agents need that same context.

Onboarding supplier data without the spreadsheet wrangling

Jimmy Ekbäck: One of the things we moved away from with traditional PIM is that a product orchestration engine operates at an earlier stage. As a retailer, it's always a challenge to take your suppliers' and vendors' data, their models and assortments, and make it sellable in your channels. That was never handled by traditional PIMs. How important is that for your operation?

Jacob Rastad: All retailers with movement in their assortment know the pain of onboarding new products. You get some kind of export from the supplier in a format that isn't suitable for your platform at all. Then you need to wrangle the data manually in Excel or spreadsheets to make it fit your setup. None of the other PIM systems offer help, guidance or, if we're talking about AI agents, guardrails for how to do it.

That's one of the things we built together in Occtoo. When we get a data export from a supplier, we upload it to Occtoo and get AI assistance in mapping the columns and values to our format. That makes onboarding so much easier. And instead of requiring the supplier to fill out our master spreadsheet, they can just send us their data, so it's easier for the supplier too.

Again, when you remove friction from a process, you start doing more things and get more value, just by removing obstacles for the people working in it.

Jimmy Ekbäck: I felt that mindset in all our sessions and whiteboarding before we started. Your team has amazing domain knowledge. The idea was to stop trying to change the suppliers and focus on what we can do. If we want to sell those product lines, we have to do the work to make them sellable, instead of using the whip and the carrot to change their way of working. That was a great takeaway.

Retailer vs brand: the same data logistics challenge

Jimmy Ekbäck: There are a lot of brands here today, and that was a question from the chat. You're a retailer with a fast-moving assortment. But if you're a brand with your own designs and a PLM inside your organization, you have the same data logistics problem. You lose context when you move between systems. What would your take be from a brand's perspective?

Jacob Rastad: The core challenge is the same. You have data about products living in different systems and different silos, with different vendors and suppliers. Say you're a 100% D2C brand. You have a manufacturer somewhere that supports you with data about the products you design. Some of that data you create yourself, in a PLM or something else, then send to the manufacturer and get something back.

That data logistics challenge is the same. It's just other actors involved. I would venture to guess that in the system landscape of the future, we won't have a separate PLM and a separate PIM system, because it's just friction along the way.

Jimmy Ekbäck: I'd also say retailers have it harder, because there's someone outside the organization you have no mandate over. Inside your own organization, you at least have the mandate to orchestrate the processes and choose your system landscape. But the general process is very similar.

The unexpected revelation: guardrails for humans work for AI agents

Jimmy Ekbäck: What have been the biggest revelations in terms of impact so far?

Jacob Rastad: As I said, the biggest impact was the speed at which we can work with our product data, from day one.

The really big revelation was a bit unexpected. You could call it luck, but luck is an outcome of how prepared you are. It was how much help we get from having built our product model with all the context around it, and building guardrails that were meant for humans from the start. Having those guardrails available to an AI agent that can help with both onboarding new products and working with products is mind-blowing.

When you connect your own harness, whether you work with Claude, ChatGPT or a custom-built one, to your different data sources, the impact on how much you can produce is mind-blowing. It goes from "I get some nice assistance in a chat and increase my productivity by 10 or 15%" to an impact of orders of magnitude, 2x, 3x, 5x, 10x, when you connect multiple systems together.

The best example: I type into Claude, which is our primary harness at Nordic Nest, "Look at the sales data from the last month and identify any potential cross-sell we haven't mapped." Claude looks at the sales data, which we have in BigQuery, does a quick analysis and identifies products that seem to correlate but don't have any relation set up in Occtoo.

Then I say, "Please add a cross-sell, upsell or spare part relation to these products in Occtoo." It uses the Occtoo MCP server, adds all of that and says, "OK, I'm done." And we have all these new cross-sell and upsell opportunities out on the website within a few minutes. It's such an enabler.

Jimmy Ekbäck: We knew early on, with "bring your own agent", that innovation would happen there. But it's interesting that the things we built to make humans more efficient were the same things agents want: to know what good and great look like. That's something we found out on this journey together, and it seems so easy when you say it.

Jacob Rastad: Definitely. There's a great saying that great design should be obvious. Great designs aren't obvious to produce, but they are obvious once you've produced them.

Jimmy Ekbäck: In hindsight, everything looks easy.

Jacob's take on AI: start small, fail fast

Jimmy Ekbäck: Let's go for the biggest topic. Nordic Nest is very impressive in the way you foster innovation and try things out. It looks easy from the outside, but you actually do a lot, and that's the difference from many who just talk. Give us your findings and your stance on the two letters: AI.

Jacob Rastad: If everyone listening takes only one thing with them today, it's this: try things out and test them at a small scale to see if you get any traction.

There are so many companies and consultants saying you need a big AI transformation project and need to spend millions on it. I'm very keen on the philosophy of failing fast. Try something small and see if it works. If it works, build on it. If it doesn't, scrap it. It won't cost you much, because you only tried a small thing.

That's the most important thing we do as an organization to achieve good outcomes. When I say we have a lot of really nice ROIs on AI projects, big and small, it's because we've also done two or three times as many small things that didn't work out.

Jimmy Ekbäck: And it's not fun to talk about those.

Jacob Rastad: It's like the Edison saying. He tried a thousand different ways to build a light bulb before he built one that worked. It's the same here. Try things, see what sticks, and build on it.

Our customer service agent, which we started building three years ago, has been rebuilt about three times since then, with improvements along the way. We started small, got a few things to work, moved on to the next type of question, realized it needed something else, rebuilt it, and it stacked on top of that. It's like compounding interest. You add a few bits along the way, and then the ROI really starts to snowball.

Connect your agent to everything

Jimmy Ekbäck: You learn as you go. AI is technology too. You've talked about guardrails, life cycles and your data approach, and you work a lot with MCP and Claude. How will this impact users, workers and organizations?

Jacob Rastad: The impact you get when you connect your harness, your agent, to all of your infrastructure is mind-blowing. It was for me the first time I connected it to my email. Then you connect it to other systems, like your ERP, your OMS or your data analytics.

For every system you onboard to your agent, the output potential increases exponentially. Going from one to two increases it a lot. Going from two to three increases it even more, because the agent can move between your systems and take actions in them. The big output doesn't come when you connect one or two. It comes when you add three, four, five.

There are still a lot of vendors with their own AI sticker, their own chatbot in their system, which isn't connected to anything else. It only works with that system's data.

Jimmy Ekbäck: Elaborate on that. There will be agents inside services, helping with things like onboarding, and we have a mutual project on that. But where should innovation take place: inside the services, or outside them, given how fast AI moves?

Jacob Rastad: There's extremely big value in having one agent with access to all your different systems. It enables big automation projects that reduce personnel or other costs. But it also enables all your employees to make better decisions, take more action and do more, day to day, which increases your output as a whole. In hindsight, that's where I would start, especially since the models have gotten so much better compared to just a year ago. The threshold is much lower now than it was even six months ago.

Once you've connected all your systems and data sources to one agent, you'll find processes that are a bit too difficult, or that need extra governance or guardrails. That's when you go from your general harness, like Claude, to building a more specialized agent for that specific process.

So there's room for both, because they solve different things. One doesn't replace the other. They're different tools for different needs.

Jimmy Ekbäck: That's so important, because at some point tokens and cost will be a factor. If you have life cycles, controls and guardrails, there are boundaries for the agents to operate within, and they can be more efficient because you give them more direction and more data context. General agents and your AI harness versus specialized agents and agent workflows. That's architecture as well. Old truths will remain true. It will just be faster: faster right, or faster wrong.

What's next: Jacob's advice

Jimmy Ekbäck: We've touched on some of the questions from the chat. As time is scarce, let's summarize. What's your last piece of advice?

Jacob Rastad: The best advice I can give is about the speed at which things are moving and the capabilities emerging from that. We've only scratched the surface of what's going to be possible.

Don't make a huge investment in a big AI project. I know someone at an insurance company that made a huge investment in building their own agent two years ago. Today it's completely worthless, because all the other agents are much better.

So do small things, try them out, see if you get value from them and build on that. Don't sign a three-year deal with an AI solutions company. Sign a three-month deal, build something and see if it works. If it doesn't, scrap it. Things are moving too fast to fall into the sunk cost fallacy.

Jimmy Ekbäck: And how important is your data in all of this?

Jacob Rastad: It's the foundation we build everything else on. If your input data is poor, whether it's product data, process descriptions or your internal wiki and how-tos, your agent will be poor as well. The agent does exactly what it's told. It won't cover for any discrepancies or misinterpretations in how you instruct it.

The good news is that it's much easier now than half a year ago to update your internal wiki and how-tos, because the harnesses are so good at it. You can just prompt them or talk to them. So really take advantage of that.

Jimmy Ekbäck: Classic small steps and fail fast, which I live by too. The only way to make something work is to fail a lot.

Thank you so much, Jacob, for making time for this. It's always a pleasure. We could have gone on for twice as long, and I don't think this is the last time we'll do this. For those who want a deeper dive into the foundation we've discussed, please register your interest. Thank you to everyone who joined, and thank you, Jacob.

Jacob Rastad: Thank you for having me. My pleasure.