|
Welcome back! Talking to business leaders about AI these days can often feel paradoxical. On one hand, AI is becoming a steady part of company software budgets, especially productivity tools like chatbots and AI coding software that individual employees use. On the other hand, there’s a persistent frustration that AI isn’t living up to the hype when it comes to meaningfully automating work. That contradiction was evident when we spoke to a range of customers for our story on Monday about AI agents, including how customers increasingly need hand-holding from Amazon, Anthropic, OpenAI and other providers to make the agents work reliably. Another theme that emerged: AI can be useful for a wide range of work but quickly loses its lustre when companies try to automate tasks that need a high level of accuracy. For instance, one frustrated cybersecurity executive told us that using Microsoft’s Security Copilot, an AI agent meant to automate security analysts’ work, was essentially like lighting money on fire because the data it provided was often inaccurate. Separately, a customer support executive at Bosch Power Tools said the company has held off from releasing a chatbot that explains how tools work because it frequently hallucinated incorrect information that could potentially get customers hurt. Still, those companies say they’re getting value out of other AI tools that aren’t quite as ambitious. For instance, Bosch is using an AI tool from SAP that reads customer support queries and automatically routes them to the right human employee—that process used to require humans to route support tickets, but now the SAP tool accurately routes 95% of the over 1 million annual tickets automatically, the company said. But the broader struggles with AI could explain why so many AI firms are offering to get hands-on with customers to work out the kinks in their AI products. Still, not every customer is enough of a heavyweight to get the white-glove treatment from AI labs. In the short term, that appears to be setting up a system of AI haves and have-nots. Snowflake’s Agent Gets General The agent race is getting crowded. Database firm Snowflake on Tuesday said it made its AI agent for answering business questions generally available to customers following months of testing. It said more than 1,000 customers had already used the product, Snowflake Intelligence, to launch 12,000 agents that can search for and get data on their sales or other corporate metrics, tapping information that resides in Snowflake databases or in apps such as Microsoft Teams and in Salesforce apps. Snowflake also said it had reached a deal with SAP so that joint customers can make data from SAP’s enterprise resource planning apps available to Snowflake tools. Snowflake CEO Sridhar Ramaswamy said at a press event that the Intelligence product is akin to ChatGPT’s deep research tool, which spends extra processing power to search through a variety of data source, “but powered by your own important enterprise data, both structured as well as on structured.” For instance, Snowflake itself developed a “go-to-market” agent so if an employee can type that they’re going to meet with a particular client and the agent generates a report that describes what Snowflake features the client is using and what kinds of support tickets they’ve filed. Numerous enterprise and AI developers have launched such agents, in part as a reaction to early traction from AI search startup Glean. But Snowflake and some of its customers, including ServiceNow, said Snowflake’s product is especially accurate, relative to other agents. Snowflake manager Jeff Hollan said at the press event that the company’s agent relies on the most advanced models from firms such as OpenAI and Anthropic but that Snowflake had rigorously tweaked them to be accurate. TS Imagine, which provides software to financial firms to evaluate risks, uses the Snowflake agent to quickly generate dashboards that contain numerous charts, said Thomas Bodenski, chief operating officer. “Rome wasn’t built in a day, but for sure dashboards are,” he said. We’re curious to see how much or how little hand-holding Snowflake customers will require to use the new product, given Ramaswamy’s prior comments about utilizing more forward deployed engineers to assist clients with AI. We’re also interested to know how much customers are willing to pay for the service. Ramaswamy said his main goal is to get customers using the product before he worries about making money from it.
Thank you for reading the Applied AI newsletter! I’d love your feedback, ideas and tips: aaron@theinformation.com. If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.
|
