How Unilever created an AI chatbot to mine its sustainability data

The global sustainability team at Unilever gets a lot of data requests. Customers need emissions numbers for Scope 3 reporting, for instance. And employees on other teams want to include sustainability metrics in bids for new business. 

Alongside these discrete asks, a genuinely substantial request lands around once a week. “Not just a one-off data point,” said team member Francesca Kennedy Wallbank. “This is a full questionnaire with multiple tabs and over 100 questions.”

Earlier this year, Wallbank wondered whether AI could help with this important but decidedly unexciting work. With the help of Unilever’s in-house AI experts, she uploaded a “knowledge bank” — a curated set of sustainability reports and other documents that contain answers to common questions — to Microsoft Copilot Studio, a tool used to build and manage AI chatbots. 

Then she began training her fledgling creation by firing questions at it. Wallbank and colleagues assessed each answer and passed the feedback to the AI team, which used the information to create successively more useful iterations of the chatbot. 

Gaps in the chatbot’s knowledge appeared as it fielded test questions. Some customer-facing employees, for example, wanted to know which Unilever products were covered by the European Union’s forthcoming Regulation on deforestation-free products (EUDR). “So we went to the regulatory affairs team to understand how Unilever is dealing with EUDR and then fed that back into the knowledge bank so it had the right information,” said Wallbank.

Going live

Around a month ago, the chatbot was ready to be shared with the rest of the company. “Rather than trying to find the right team and the right Excel document to review and understand the regulation, they now can be told the information instantly and then what their next step should be,” Wallbank explained.

As well as saving time, the chatbot also gives teams access to more data than would have been practical previously, which in turn allows them to strengthen bids in which sustainability is important. “By providing this more detailed data, we’re able to win more business,” Wallbank added.

Unlike publicly available AI systems, such as Claude and ChatGPT, the chatbot doesn’t look beyond its knowledge bank, which currently contains around 25 documents. Wallbank started by uploading company-level documents and is now working on adding more granular information, such as data the company collects on plastics, nature and specific products.

Despite the focus on carefully controlled information — the chatbot even accesses Unilever’s intranet — every answer comes with a link to its source, and users are advised to click through to verify the data before sharing it. The “heavy lift,” as Wallbank sees it, is finding the right data in the right documents. 

Support, not replace

There’s tentative evidence that AI is already causing some firms to scale back hiring in roles that AI has proven adept at, such as entry-level coding. But Wallbank pushed back on the suggestion that her chatbot might replace Unilever sustainability employees. “The idea is to support employees and to save time for important, high-quality tasks they weren’t able to do previously,” she argued. 

In the long term, she added, she imagines the AI itself being replaced: One of the next things she wants to work on is a centralized repository that customers can access without needing to ask Unilever.

“Where I would love to go is an industry coalition where we agree on the data that customers need and then we can automatically share this data with them,” Wallbank said. “So you don’t need this back and forth anymore. That is the future.”