At 112 years young, Holcim is proving that age and innovation can go hand in hand. With headquarters in Switzerland, a presence in 70 countries worldwide, and $27 billion in net sales, the construction building materials company is using cutting-edge technology to help the customer ordering process run more smoothly. To transform what was once a manual system into something speedier and easier to use, Holcim recently began experimenting with a generative-AI (gen-AI)-enabled mobile solution that allows customers to place cement orders quickly through a messaging app. This frees up Holcim’s sales staff to focus on higher-touch, more complex orders. As part of the C-Suite Growth Talks series from McKinsey’s Growth, Marketing & Sales Practice, Juan recently sat down with McKinsey’s José Carluccio to talk about the great promise this pilot is showing. The following is an edited version of their conversation.
José Carluccio: What does the journey look like for a customer when they want to order cement?
Juan Beltrán: Traditionally, it’s been very straightforward. If someone wants cement, they call a sales rep and ask them, “Hey, can you please send me another truck of cement?” And that’s all. The sales rep looks at the contract, the materials available, and the address where the company needs to deliver this order. They have a conversation with the customer. Then we ship it. It’s very simple.
José Carluccio: How do you make a simple process even simpler with gen AI?
Juan Beltrán: We want to allow customers to order by text. We’re doing that by going to a well-known, well-adopted solution for many people. That’s WhatsApp. And we’re using gen AI to try to understand the customer’s natural language when they speak, because when they order, they may not use SKU numbers or product names.
José Carluccio: What has your experiment looked like?
Juan Beltrán: We conducted a pilot last year in Spain to test an AI-enabled copilot customer ordering solution. Instead of picking up the phone to make a call, a customer can open up WhatsApp and request a cement truck. Our tool recognizes the customer, brings up their order history, and makes proposals for the next order. The AI model uses natural language when responding to the customer and makes suggestions for products to order. The customer can then make modifications to the order, including the number of trucks, the delivery time, the type of material, and where we should ship the cement. Then they can confirm the order. Then we ship it.
We wanted to see if this was something our customers would like. We got amazing feedback. They said it’s been very convenient for them. It’s made it a lot easier to place their orders anytime, from anywhere.
José Carluccio: What have the results been?
Juan Beltrán: With this AI-enabled pilot, around 66 percent of our first-order proposals were accepted because they’re accurate and reflect what customers want. There’s also been increased adoption with this tool compared to other customer portals; we managed to move from a 25 percent adoption rate to a 93 percent adoption rate.
José Carluccio: We see a lot of hype around gen AI. Getting value out of it is not easy. What have been some of the challenges on your journey so far?
Juan Beltrán: The use case is straightforward, but generative AI is quite new and there’s not a ton of expertise around.
During testing, sometimes our gen-AI-enabled tool could be too chatty, or overexplain things, or even recommend products that weren’t part of our product portfolio. These are the main challenges. We need to handle them by restricting possible answers to queries.
Latency is also very important. The customer expects a human-like interaction. They expect to receive a message like someone is writing it. They don’t expect the message or product proposal to appear too fast. They expect that it takes a little bit of a time lag. But the response from us also needs to be quick enough. If it takes five or ten minutes, the customer is going to abandon the process. These are some of the tricky things we encountered during our testing.
José Carluccio: Can you give an example of something the gen AI tool got wrong during testing?
Juan Beltrán: During testing, one customer tried to see if he could order a famous Spanish cold tomato soup called gazpacho through this copilot tool. The tool identified gazpacho as a new SKU and proposed sending a truck full of gazpacho soup. It was funny. But no, we didn’t deliver a truck full of soup.
José Carluccio: What’s your secret sauce to get such high adoption rates?
Juan Beltrán: It’s critical to meet customers where they are and offer a solution with tools they’re already using. In this case, it’s WhatsApp because we know our customers are already using it. They don’t need to learn how to use something new. They don’t need to select products from a catalog. They don’t need to overthink. They just need to use what they use already. It’s also important to let them use natural language.
José Carluccio: How do you inject your AI models with a human touch?
Juan Beltrán: We try to make things as human as possible. Humans decide the tone of voice to be used for the interactions and the kinds of messages we send. Humans define the whole process. There are some cases where customers only need to interact with the copilot tool to get their needs met. But there can also be cases where they need more help. The LLM (large language model) may not handle certain queries effectively. That’s why we always offer customers the option to interact with a human agent.
José Carluccio: For this project, there are a lot of moving parts. There’s a messaging platform, a CRM (customer relationship management) system, a cloud provider, and more. How do you manage the complexity of a project like this when there are so many different systems?
Juan Beltrán: It’s not always smooth. Sometimes we work in parallel with two different models that are providing intelligence to the tool. But we make sure everyone knows what their roles and responsibilities are in the project. We try to provide clear accountability and transparency. And someone from our Holcim team took the lead in the project, jointly with McKinsey, making sure we were moving in the right direction and aiming toward a clear goal, which was to provide a fantastic customer experience. Everyone was very committed. Excitement makes collaboration easier.
José Carluccio: How long did this pilot project take? What was it like to prepare, set up teams, and stand up the proof of concept?
Juan Beltrán: The most consuming part of the whole project was the prework: looking at the requirements, finding the right partner, and understanding where the technology stands. It took some time to determine if the right technology existed to provide the solution we needed. This took around seven months. The actual development of the solution was quite fast. It took two to three months in total. The project took more time to find the right technology than to develop the solution itself.
José Carluccio: You now have proof of concept in Spain. How are you envisioning the next wave of AI-enabled ordering?
Juan Beltrán: We’re rolling out this solution to all regions in Spain. We want to make sure it works. We want to capture customer feedback, fine-tune the solution, and make sure it’s valuable for customers. Then in 2025, we’ll expand to other countries. We’re starting in Europe, then we’ll go global. This is an easy business line compared to others. We have more complex ones where there are many more products, many more SKUs, and many more ship-to addresses. Finally, we want to expand to other channels. WhatsApp is, of course, one of the most used channels, but phone calls are still popular with our customers. It’s a very industrial business, and when customers need cement, they don’t wait for someone to ask them. They pick up the phone, they call straight away, and say, “Hey, I need cement for today or tomorrow.” We need to find a way to add this technology behind phone calls and emails as well. LLMs will need to handle much more complexity. That’s next for our road map, and it’s very exciting.
