My latest in today's print edition of The Wall Street Journal: AI work assistants are off to a slow start. AI work assistants were designed to provide businesses a relatively easy avenue into the cutting edge technology. It isn’t quite turning out that way, with CIOs saying it requires a heavy internal lift to get full value from the pricey tools. “It has been more work than anticipated,” said Sharon Mandell, CIO of Juniper Networks, who is testing tools from several vendors but doesn’t feel ready to put any into production. Tools like Copilot for Microsoft 365 or Gemini for Google Workspace aim to access large bodies of enterprise data—including emails, documents and spreadsheets— deliver reliable answers to questions such as “what are our latest sales figures?” But that isn’t always the case—in part because the enterprise data they are accessing isn’t always up-to-date or accurate and in part because the tools themselves are still maturing. Mandell said if she asks a question related to 2024 data, the AI tool might deliver an answer based on 2023 data. At Cargill, an AI tool failed to correctly answer a straightforward question about who is on the company’s executive team, the agricultural giant said. At Eli Lilly and Company, a tool gave incorrect answers to questions about expense policies, said Diogo Rau, the pharmaceutical firm’s CIDO. Vendors have noted the issue. “As companies started using Copilot, people started finding data that companies didn’t know they had access to, or that they realized wasn’t as fresh or as valuable as it could be. And then they realized, ‘Oh, we’ve got to do more,’” said Jared Spataro, corporate vice president of AI at Work at Microsoft. Read the full, unlocked story here for how CIOs and vendors are aiming to solve the issue: https://lnkd.in/eKKPcDjt #tech #cio #ai #artificialintelligence
It was good to see the focus on the need to prioritize data management. LLMs don't provide the level of precision data management required in the enterprise. Enterprise data management can certainly help, but it doesn't eliminate some of the problems with LLMs. Manual data management is very expensive, making it much more difficult to achieve an ROI. It's primarily a system design challenge. Precision data management has been part of the core of the KOS invention for more than 20 years. Happy to discuss if and when the WSJ is ready to move on from covering the same few products that have inherent design flaws.
Organizations are turning on builtin copilots and turning them off for a few reasons. (1) copilots are exposing weakness in security and governance controls in the underlying apps and data systems. Huge risks of sensitive info being exposed broadly internally or even externally (2) underlying data needs curation for different use cases and it needs to be automated. (3) quality of answers suffers with it also Teams are back to drawing boards to first automate the underlying data controls and then turn it on. Copilots are bound to happen, just needs more prep and automation for it. Well placed article Isabelle Bousquette Sharon Mandell
This is very well written and your observations are accurate. Kudos!
This well articulates that LLMs are not wholly suitable for the Enterprise, specifically for automating manual human tasks that require specific data and accurate decisions / responses. One interesting by-product of our Domain Specific Model (DSM) approach, helps us to actually identify underlying data issues, allowing us to make the data more reliable, cleaner, and ultimately a source of truth that automates (97%+) of that specific workflow.
Thanks for including me in the conversation.
Excellent analysis, Isabelle Bousquette and The Wall Street Journal In the enterprise context, relying solely on copilots and AI assistants is not sufficient. Businesses require robust platforms that can prepare and organize unstructured data for effective utilization by Generative AI tools. At Altilia, we are developing a platform designed to enable companies to create multi-structured knowledge bases. Our #AI models are adept at organizing unstructured data scattered across various organizational silos. They classify content and accurately extract data and metadata, regardless of the format or type of unstructured data source available within the organization. This content and data are then indexed and stored in multi-structured knowledge bases, which take the form of knowledge graphs enhanced with semantic indexes. This innovative approach allows AI assistants and AI robots we allow to create by the platform to harness the extensive unstructured data within enterprises. These digital workers or colleagues can then collaborate closely with enterprise users, answering crucial business questions to empower decision-making processes or automate tasks and operational processes.
Isabelle Bousquette Generally true but you need to look deeper like with Orby AI building AI agents for the enterprise https://www.linkedin.com/posts/sardire_orby-is-building-ai-agents-for-the-enterprise-activity-7212086915301355520-IoZs?utm_source=share&utm_medium=member_desktop “Orby’s platform observes how workers do their work in order to automatically create automations for complex tasks that require some level of reasoning and understanding,” Liu, Orby’s CEO, explained. “An AI agent installed on a worker’s computer effectively watches, learns and generates automations, adapting the model as it learns more.” Orby’s purported secret sauce is a cloud-based generative AI model that’s fine-tuned to complete customer tasks, such as validating expense reports. The model relies partly on symbolic AI, a form of AI that leverages rules, such as mathematical theorems, to infer solutions to problems. “Humans are kept completely in the feedback loop,” she added. Cheers..Steve AI startup advisor 'force multiplier' https://www.forcemultipliersteveardire.com
Great input from Sharon Mandell! Always wonderful insights! Isabelle Bousquette, your article on the challenges of adopting AI work assistants in enterprises is well-written. Your point about data is spot on and was the focus of a recent SIM San Francisco Bay Area panel on the critical importance of a robust data foundation for AI strategy. Making the case to boards and CXOs is crucial, as poor data quality renders even the most advanced AI solutions ineffective—truly a case of "garbage in, garbage out." Understanding the technology architecture is vital for strategic and investment planning. If it helps, I have shared my thoughts on GenAI for enterprises, including a segment on reality checks which supports your observations: https://www.linkedin.com/posts/benjamindai_activity-7211432463833653248-nzWL?utm_source=share&utm_medium=member_desktop #disruptiveinnovation #genai #datastrategy #technologyrift #architectureiskingandqueen #architecture
Co-Founder and CEO at SimplyPut
1mo“It isn’t quite turning out that way, with CIOs saying it requires a heavy internal lift to get full value from the pricey tools.” This is a complaint we hear a lot! It’s why SimplyPut has spent so much time to focus on taking care of the heavy lifting of understanding company data, with mechanisms for data experts to verify accuracy. This ensures everyone always get *trusted* answers to their data questions. Our core value is enabling anyone in a company to obtain reliable data answers - accurate enough for a board deck.