|
Welcome back! Businesses looking to expand their internal use of artificial intelligence from a select few testers to a large number of workers are hypersensitive to increased costs. Over the last couple of years, that has meant many of their AI projects have never made it out of the pilot phase. Veolia, a 220,000-person provider of water treatment, waste management and energy services, seems to have found a middle ground. While the company, a spinoff of Vivendi based near Paris, has rolled out an internal chatbot to all of its employees, it has chosen to limit the number of staff that can access certain features, like image-generating tools. As it stands, only staff working on Veolia’s marketing, sales and product teams are able to generate images using the company’s chatbot, said Fouad Maach, group head of architecture and engineering, who has overseen the technology’s rollout. Other staff can use the chatbot to help answer questions about information technology or human resources policies, or to help them generate emails and documents. More than 60,000 people now regularly use the chatbot, Maach said, up from 48,000 late last year and 2,000 in 2023. Staff get to choose which model they want to use for each conversation, including those from Cohere, Mistral AI, Meta Platforms and Anthropic. Anthropic’s Claude is the most popular, Maach told me, because its staff has found it excels at a wide variety of tasks. The company buys the models over Amazon Web Services’ Bedrock, which hosts AI models from many vendors. A smaller number of workers also get access to image generation tools. They use it to create visuals for purposes like sales presentations or customer support materials, Maach said. The image tool uses OpenAI’s image-generating software, GPT-4o, as well as models from Stability AI, which helped commercialize one of the earliest widely used image models. Most images are created with Stability’s models, Maach said. Maach wouldn't specify how much Veolia spends on running its AI chatbot, but he said the amount was similar to the cost of purchasing a subscription from a vendor like OpenAI or Anthropic. OpenAI’s publicly listed price for its Team plan, geared toward businesses, is roughly $25 to $30 per user each month. That means Veolia would pay more than $15 million per year on AI subscriptions if it were using OpenAI. In 2024, Veolia generated 44.7 billion euros ($51 billion) in revenue. “We reach a very low cost compared to this kind of subscription model,” Maach said. In recent years, AI developers have been laser-focused on making their models as powerful as possible. Matt Fitzpatrick, CEO of Invisible Technologies, which provides data developers use for AI model training, thinks businesses might benefit just as much from small models that excel at very specific tasks, like extracting data from PDFs. Such narrow uses for AI models will become increasingly common in enterprises over the next couple of years, as business leaders start to realize the difficulty of using systems set up to handle a variety of tasks, Fitzpatrick said at The Information’s Financing the AI Revolution event in New York on Monday. Another challenge for businesses has been making AI software work alongside older technology they have already introduced. That’s one reason startups and other smaller companies have so far been able to benefit more from AI than larger businesses have, Fitzpatrick said. Large companies will have no choice but to rebuild much of their internal software and processes to get the benefits of AI, he added. It’s easier to get results from AI “if you’re a digitally native company starting from scratch,” Fitzpatrick said. “But if you have to reconfigure the way your business currently works, that’s harder.”
| 
