How HelloFresh Replaced 450 Spreadsheets With Real-Time Decisions

[00:00:00] Regular listeners will know I always say that I only partner with companies that align with my values and what I'm trying to build here at Tech Talks Network. And NordLayer fits that perfectly by helping businesses stay secure without adding unnecessary complexity. And every company today relies on SaaS tools, whether it's collaboration platforms, CRMs or finance systems. The common denominator here is everything now happens in the browser.

[00:00:28] But here's the problem. Most browsers were never designed with business security in mind. This is one of the many reasons why NordLayer has introduced a new business browser that puts security, visibility and control directly where work happens. And it helps protect against things like phishing, malicious sites and unauthorized access, while also giving your teams a safer and more controlled way to work online.

[00:00:55] And the best part is it doesn't change how people work. It just makes what they're already doing more secure. So if you want to learn more, please head over to nordlayer.com slash browser and see how it fits into your workflow. But now it's time to get today's guest on. Welcome back, gang. I'm recording here at Click Connect in Orlando, Florida.

[00:01:22] And it's easy to get caught up in the big conversations around AI automation and what the future could look like. But for most businesses, the real challenge isn't the future. It's the day to day reality of making better decisions faster with the data that you already have. Now, when you're running a complex operation, especially something as time sensitive as food delivery, there's very little room for error.

[00:01:52] Because you're dealing with forecasting demand, managing supply chains and coordinating logistics and making sure customers receive what they need and when they need it. So what happens when the systems behind all that start to break down? Well, to explore this today, I'm joined by Ed Dunger from HelloFresh, where he leads operational technology and analytics enablement across global teams.

[00:02:21] So today we're going to move away from theory and into the reality of what it takes to transform operations at scale. So we will talk about everything from replacing hundreds of spreadsheets with real-time data platforms to building trust in data across teams and ultimately creating systems that people actually want to use. And we'll also dig into how small incremental changes can build into something much, much bigger.

[00:02:50] And why the journey to operationalizing data and AI is often measured in years, not weeks. But enough scene setting for me. Let me introduce you to my guest right now. So a massive warm welcome to the show. Thank you for joining me here at Click Connect. For everyone listening, can you tell them a little about who you are and what you do? Hi Neil C. Hughes, thank you for having me. So my name's Ed Dunger. I currently have been working for the last four years with HelloFresh.

[00:03:19] And I've been working mainly at the market level, supporting what we would call local ops tech. So that's a range of technology from through to forecasting, through to audio creation, through to our logistics delivery network, supporting the day-to-day operations. I've just seen you on stage. You've come straight off and sat down with myself here. But for anyone that couldn't attend there, tell me about that business challenge that you first came across.

[00:03:48] So in the Click Lightning round, the business challenge we articulated there was around our supply chain network. And we've been using the UK market, Click, to help us understand the demand that we need to cover, the SKUs that we need to order, trying to predict things like potential waste so that we can mitigate that and improve our sustainability. So yeah, so Click, we went through very quickly how it's enabled us to remove loads of G-sheets out of our process, digitize it,

[00:04:18] and move to a more real-time, data-driven, decision-based ordering process. And so much of what you said there will resonate with people in similar industries. So you've got that challenge, you've got the partner, we've got technology. What is it that you did differently from before that would help you address this challenge? So I think I was using a combination of technology tools. So we've used Snowflake and Click together. And we've also used Infinity Writeback in there as well.

[00:04:44] And so we've made an interactive application that is updating in near real-time. So the buyers in our ordering team can see what is actually happening in the warehouse, what's going on with deliveries turning up. So, and they can make corrections and they can make notes and comments in it. So that fundamentally changed the way we operated away from a lot of things going on to Google Sheets, which is a valid method.

[00:05:13] But we'd rather have it that you can't accidentally overwrite something. You're not having to download and refresh the data. It's just there for that individual to then be freed up to think and make the correct business decisions. So you've gone there from having the problem, getting the right tech partners and technology and doing things differently. If you look back now, what has changed since if you have that before and after picture? So I think the biggest thing that changed for us is over this process,

[00:05:40] we've deprecated 450 Google Sheets that were supporting the team. And I think that's the biggest doing things differently is changing that mindset. It's been a two to three year journey. But if we look back at how the team operated in the past, they would all agree that they really enjoy the way that it works now and it makes their life a lot easier. So that has been the key criteria for me that our operational staff are happy. Yeah.

[00:06:09] So that's the key part, that they're armed with the right data. And one of HelloFresh's core DNA principles is data drivenness. So it supported that aspect of the business. And I think it's important to highlight that this wasn't an overnight success. You mentioned a two to three year journey. Everyone's happy now. They probably went through a little bit of friction along the way. What were the biggest lessons that you learned for anyone that's at the start of their journey, maybe wanting to follow in your footsteps? Any big lessons that you learned?

[00:06:38] Put things down into bite sized chunks. Don't try and write one massive specification to replace one massive set of processes. So we tackled it part by part. You've got to work alongside your operational stakeholders. So we had people from the ordering team heavily involved in helping us to not be tech, write the solution. So it was a combination effort. And then the other bit is building that trust up in the data.

[00:07:04] So that's, I think, probably the biggest part that you have to achieve, that those stakeholders know that the number being presented there, if we go and cross-reference it in any system, that they can see that granularity and know that they've got the right data to make decisions off of. And I think that was the journey of building the trust up because it's a change process. So people have to get used to switching from one way of work into another.

[00:07:32] And we took that gradually. We didn't radically force people. So we built them up. We ran things in parallel. And then once they started to see bits, I mean, one of the early wins was it predicted some waste on a particular skew line. And the team were able to then mitigate that waste.

[00:07:57] And they started to see and trust the value in, OK, yeah, that is actually right. It is actually starting to tell me stuff that I can action. It's such a crucial part that you mentioned there because I think there is a bit of friction when you introduce new ways of working, new technology. And there's almost a belief that if you just throw AI at anything now, it will make everything go away. But you need that buy-in, that adoption and those small wins to buy that and earn that trust, right? I think, as you said, it's the small wins.

[00:08:27] We definitely focused on, OK, that's made this bit of your life easier. You trust this bit that it's telling you. And that combines and it's like compound interest. Over time, it just builds up and builds up. Yeah. And it makes them feel part of the process, I would imagine, rather than being just thrown at them. And you must now do this if they feel that they're guiding it in some way. Yeah. And that guiding was critical because I've never been a buyer. I've learned a lot, but I wouldn't be able to just sit on the outside and understand the business problem.

[00:08:57] So it's a co-educational piece, I would say. Yeah. And fast forward to present day, you are responsible for everything from forecasting through to final mile delivery. So I'm curious, where are you seeing data and AI making the biggest difference in real operational performance today? Yeah. Yeah. So one of the other areas that we got the AI Innovation Award for from the data impacts by Click.

[00:09:22] We've been using ClickPredict in our logistics exception management processes. So this was actually the start of our Click journey in UK HelloFresh. We had, again, lots of G-sheets managing exception processes. So when your box has an issue within the courier network. And we, again, turned that into an interactive digital portal that our agents could work with.

[00:09:50] And then we enabled it to push data to different teams so that we could get better at communicating our customers if there was an issue. Asking production to remake orders that had failed in delivery network. Because, again, another HelloFresh DNA core principle is customer centricity. So it's not great if you don't get your food for the week. Yeah. So even if we can get it to you a day later and the product is perishable, so you can't just hold it in the courier network.

[00:10:21] So that's where we can help with the concept of recovery boxes that get it to you the next day. At least you've got your food for the rest of the week. So what we've done is we've been using ClickPredict to try and work out when items are likely to fail. So we can trigger that recovery process quicker. We can drip feed it into production. So we're not asking production to have to do a mass batch if it's there. So it's easier to drip feed it in. And ClickPredict, we built the model in a day, which was quite impressive.

[00:10:51] And it's performing really well. So it's sped up the process. It's a negative experience for the customer in the sense of their delivery is failing. But at least we're now faster at being able to take the corrective measures that we can to then ensure that they get their box. And I would imagine making those predictions really depends on how much you trust that data. And right here at ClickConnect, we're hearing a lot about the importance of trusted data.

[00:11:19] So in your world, what happens when the data is wrong or delayed? And how do you guard against that when you're making these predictions and everything kind of evolves around that data? So what we did on the prediction side is we understood business risk appetite. And I think where people go wrong with a lot of AI is they just think it's got to be 100% right. So you need to look at your current processes and think, okay, so how reliable are those?

[00:11:45] And what we found was humans were delaying, waiting for more information. So actually, the prediction was faster because it wasn't biased by human thinking X, Y and Z. I need, I want a bit more confirmation. It would say, well, actually, just trigger the recovery process now. And then what we also worked out was what's the tolerance for when the prediction gets it wrong. And we've got a reasonable margin in there.

[00:12:12] And that was agreed with the business stakeholders that actually, even if it gets it wrong this amount of times, it's still more profitable to trigger the process and have a few false positives. And we got the business to define where that false positive rate tolerance should be acceptable. So we'd then be able to put in checks to say, okay, is it still performing within that tolerance? If it isn't, that's when you need to interdict.

[00:12:41] Otherwise, just let the process run. And you get the power of that speed of AI processing within the boundaries that you've given it. Love that. And another one, I've also read that you're working across highly automated distributed centers in everywhere from the UK and Germany too. So how do you connect data analytics and physical operations in a way that actually improves outcomes on the ground, that evolves around, again, that customer centricity that you mentioned?

[00:13:10] So I suppose, Noah, it's getting people used to knowing where there's a good, reliable, single source of truth. So we've invested in that process that where I think the friction comes is when someone's got their spreadsheet. Someone else has got their spreadsheet. Someone else has got their numbers. And you don't know the, and they've got different data currencies. And they may have done the calculations slightly different. I think that's the biggest driver for operational friction.

[00:13:36] And it was one of the key areas that we addressed in both the United Kingdom and in our German slash DAC market. But again, that takes a lot of effort. It's not something that you easily do overnight. You've got to convince those individuals that, yep, this is a good digital source and it's easy to use. I think that's the other thing. You don't want someone to have to be a highly experienced data engineer to get hold of the data.

[00:14:03] You want nice, easy, queryable tables that they can go in and pull their ad hoc needs out of. And then agreeing the cadence, I think, has also been very helpful that we all understand the currency of the data there and what that means when you're doing a business process. I would say that that's probably been the biggest part that we've had as we've got two new automated distribution centers, one in Barleben in Germany, one in Derby in the UK.

[00:14:28] And as we've scaled those operations up, that project of supporting it with good, reliable data has really helped those. As issues arise, as you scale a new concept and a new building, it's helped the operational teams really be able to get hold of quickly the data they need to then understand what do I need to address and what do I need to sort out. I'm from Derby originally. Whereabouts is that? It's on Green. I think it's Green Park, it's called.

[00:14:55] It's a new eco park in Derby. There's Greg's next to us as well. And it's expanding rapidly, that business park, actually. It's changed since the one HelloFresh building. There's now multiple units springing up. So good for Derby. Brilliant. Love it. And obviously, you're also dealing with complex logistic challenges from ensuring deliveries arrive on time to also maintaining product quality. Where has AI or analytics had maybe the most tangible impact on customer experience?

[00:15:24] And the reason I ask that is there's so much hype around AI. But hearing these very real stories of, hey, it can make a measurable difference, a measurable impact. So I think it's back to this exception management piece for us. It's trying to interdict faster. So for our product, it has an ultimate shelf life that it can sit in the current network. It's not like an item of clothing that can get you three days, four days late. You might not be happy, but your jumper hasn't degraded.

[00:15:51] So for us, the AI piece is, as I mentioned, about making that process faster, removing the bias of a human, necessarily wanting that confidence of the data says X, Y, and Z, so that you can then trigger your processes and not rely on a human there waiting to press send or something like that.

[00:16:13] So where we've had great success with the click part of our logistics application is that it's pushing instantaneously, asking production to remake the box, getting the customer that communication to say, it's not great. There's a problem, but we're doing something to solve it. And I feel if you can communicate faster with your customers, again, what people hate is not knowing. So where's my box at 6 p.m.?

[00:16:43] So it's like, I need to cook dinner tonight. Whereas if we've known since 1 p.m. that that's not going to happen, at least we can tell you that in a really fast manner so that you can do something about what you're going to do for your evening food. And also be aware that we're triggering X, Y, and Z process to hopefully correct your experience.

[00:17:07] And as demand grows and systems will inevitably become more complex, what have been the biggest bottlenecks in scaling your data and scaling your AI capabilities? Is that something that you're still working towards now? Yeah, so HelloFresh rapidly scaled during COVID. So there's a lot of mixed systems.

[00:17:30] And I think that's probably been the biggest challenge is getting consistent, reliable data that you can join together from different platforms that have been developed by different people quite often who have left the company. So I think that's the biggest challenge for AI, that in my view, it needs good, well-structured rules and data. And also just saying to AI, oh, just go and answer this big question.

[00:17:55] I think actually breaking it down into more little microservices of AI that then feed into each other really helps to construct it to then get you a reliable answer that it can come out with. I think too many people just go, oh, I should just be able to ask it this big, wide-ranging question.

[00:18:14] So an example was last night at the awards dinner, I was talking to a human rights lawyer, an actress, and she was saying about how a young refuge could be called a young refuge. It could be a refuge for young people.

[00:18:30] And I think if you ask AI to just try and work that out from scratch, whereas if you've asked a part of AI to prep that data for you beforehand and do that task of, okay, align all of these similar terminologies, then that can then feed into the next stage of the piece of analysis that you want it to do. Looking at everything that you've built so far and the improvements you've put in place,

[00:18:57] is there one lesson you wish you'd known earlier when it comes to operationalizing data and AI in a supply chain environment? Anything stand out there? So operationalizing data, I think the one bit that I would have known earlier would have been knowing your structure that you want to feed into AI

[00:19:19] and having known more about how AI operates, I think understanding the foundations you need to build are that if you don't build a good foundations, like for a house, it's going to crumble down and you might get subsidence and you will start, for example, AI hallucinations. So I think the bit which I would have liked to have known is how best to have structured that data for the AI tools that are available now.

[00:19:48] Now, unfortunately, they weren't available back then. So it's a bit of a catch-22 problem. But I think as we all explore now and we know that AI is going to be that consistent norm within expectations of business operations, I think really getting to grips with what's the structure of the data setups you need to feed into those tools and how do you need to break it down really will help make them faster and easier to adopt.

[00:20:16] And for people listening, business leaders and people from organizations that are trying to move from pilot phase, from experimentation to real world impact, any advice that you'd give them based on your experience on your two to three year journey? Yeah, so my advice would be don't worry about failure. I do quite strongly believe in fail fast, fail often. And I think people try to approach these AI projects with, oh, it's going to be this massive business saving.

[00:20:44] It's going to be this massive outcome and effect. And by that massive waterfall nature of tackling the problem, it quite often fails. And the business pace, it doesn't deliver and then meet expectations. So I think tackling little small chunks is quite good. You can start to show the value of the return that you'll get in. And before you know it, if you've combined four or five small bits together, you've got one large piece that is having quite a significant effect.

[00:21:13] And that may take you six to 12 months. But when you reflect back, you've actually moved the goalpost quite a distance over that period of time. Fantastic advice. And before I let you go, before you fly home, one of the great things about any event, I think, is you get to spend time with people you wouldn't normally and bounce ideas around. You said that a few days ago you were having an award ceremony with a human rights lawyer. Those little moments of serendipity.

[00:21:43] When you take all the conversations you've had, what are you going to be thinking about and reflecting on that long flight home? For me, I think it's the MCP piece that was announced yesterday. And I quite liked the ClickFound. What Mike Capone said was it's that freedom piece. And I think being able to use a range of tools to tackle a problem is very handy within that software architecture.

[00:22:09] And especially with the pace of how things change, that you might have something that's great now, but then something better is coming along in six months' time that you might want to switch out with. And I think having that, I think the biggest thing I've picked up is having that flexibility is quite a key concept. Awesome. Well, I will add a link to HelloFresh website and indeed your LinkedIn for anybody that would like to carry the conversation on. But you've literally just come off stage. It's lunchtime. I'll let you go off and get something to eat. But thank you for stopping by today.

[00:22:39] Brilliant. Thank you, Neil. One of the things I loved about this conversation is there was no overnight success story here. Yep, it was a two to three year journey of gradually changing how people work, building trust in the data and proving value step by step. And I think this is something that often gets lost in the wider conversations around AI.

[00:23:03] Yeah, we often talk about what technology can do, but far less about what it takes to get people to trust it, to adopt it and rely on it in operational environments. And I think there was a very important point here around starting small, rather than just trying to solve everything at once. Breaking problems down into smaller pieces, delivering quick wins and building momentum over time.

[00:23:29] That idea of compound progress, I think, is something that really stood out. And then, of course, there's the human side of it all. Making sure the people using these systems feel part of the process, not replaced by it, by giving them the tools to make better decisions rather than forcing change upon them. All critical points. So as you think about your own journey with data and AI, the question isn't just what technology you adopt.

[00:23:58] It's how you bring people with you, how you build trust and how you turn those small improvements into meaningful, lasting change. Love to hear your thoughts on this one. Are you seeing transformation happen in small, incremental steps? Or are you still being asked to deliver those big, shiny results overnight? Please share with me your insights, your experiences, your stories at techtalksnetwork.com. But I'm afraid we're out of time once again.

[00:24:28] I'll be back again bright and early tomorrow with another guest. But thank you for listening today. And I'll catch up with you all again tomorrow. Bye for now.