Michael Horn

#AIagents aren’t failing because the models are weak. They’re failing because we’re designing systems, not intelligence. Yesterday, we had a discussion with colleagues at IBM Research who worked with UC Berkeley on why enterprise #AI agents fail is worth executives’ attention—not as a technical post‑mortem, but as a leadership signal. The analysis shows that many multi‑agent systems underperform not due to LLM capability, but because of how work, roles, and accountability are designed across agents. What struck me most: the failure modes look uncomfortably familiar to anyone who has led large organizations. Misaligned objectives. Ambiguous handoffs. Too much orchestration, too little ownership. And an assumption that “more agents” equals “better outcomes.” The research makes clear that scaling agents without rethinking system design simply multiplies failure. Performance breaks down at the seams—where agents interpret instructions differently, fail to validate outcomes, or proceed without shared context. For executives, the takeaway isn’t “slow down on agents.” It’s “change how you govern them.” Some questions leaders should be asking now: 1. What decisions are we delegating to agents vs. automating tasks? Agents fail most often when they are treated as task runners rather than decision‑bearing actors. 2. Where does accountability sit when agents collaborate? Human orgs struggle with this; agent ecosystems amplify it. 3. How do we detect failure early—before output looks plausible but is wrong? The research shows that many failures pass superficial checks and only emerge downstream. 4. Are we investing as much in system design as in model selection? Most failures are architectural and interactional, not model‑driven. The uncomfortable truth: enterprise AI maturity is no longer about access to smarter models. It’s about designing systems that assume things will go wrong—and catch it when they do. The next phase of AI advantage will come from leaders who treat agent systems less like software projects and more like operating models. That’s a leadership challenge, not a technical one. 🔗 Why do enterprise agents fail? Insights from IT‑Bench using MAST: https://lnkd.in/eXaWxu8N Nicholas (Nick) Fuller Daby Sow Ayhan Sebin Yu Deng Saurabh Jha Cathy Fillare Karen Butner #IBM, #AgenticAI #executivequestions

22

OpenAI’s new State of Enterprise Report confirms what many of us are seeing every day. AI is not slowing down inside companies. It is accelerating fast. I broke down the biggest signals from the report and what they mean for operators who are shaping AI adoption in 2025. Here are the highlights: → Weekly ChatGPT messages increased 8x in the past year and the average worker is sending 30 percent more messages. → Frontier workers are pulling ahead, sending 6x more messages than the median employee while frontier firms send 2x more per seat. → 75 percent of workers say AI improved the speed or quality of their work and most saved 40 to 60 minutes a day. Heavy users saved 10 hours per week. → Structured workflows like Projects and Custom GPTs were used 19x more this year and reasoning token consumption increased 320x per organization. So what does this mean for enterprise leaders? ➡️ Treat AI usage as a capability curve. Frontier workers are not just faster, they are doing entirely new types of work. ➡️ Standardize workflows. The strongest gains come from repeatable processes, not one-off prompts. ➡️ Prioritize integration. Leading firms invest in deep system integration, data readiness, and change management because these steps determine long-term lift. The future is already here. The gap now is between enterprises that operationalize AI and those that rely on casual experimentation. If you want the deeper dive, I unpacked the full report here. 👉 Read the full analysis: https://lnkd.in/ebwf2aKA #AIAdoption #EnterpriseAI #FutureOfWork #AIProductivity #DigitalTransformation #AIOps #OpenAI

41

8 Comments

If you’ve ever tried joining Census data to H3, you know the pain. This is a clean, canonical, BigQuery-ready dataset — and there’s a free Manhattan sample to test! If you know, you know

10

1 Comment

Some exciting news today from Newton Research with the announcement of our integration with Snowflake Cortex AI. Proud to be partnering with Snowflake and our joint customer Horizon Media in driving secure, in-platform marketing analytics - Scaled through agentic AI

20