Bardeen

95% of AI pilots are failing! 📉 🤖 That’s the headline finding from MIT’s new GenAI Divide report, based on 300+ enterprise AI deployments. Despite $30–40B in investment, only 5% of pilots create measurable ROI. ...Why? 🔹 Most tools don’t learn—they forget context, can’t adapt, and break in real workflows. 🔹 Big companies run the most pilots but scale the least. 🔹 70% of AI budgets go to sales & marketing, while the highest ROI lives in back-office automation. 🔹 Internal builds fail twice as often as external partnerships. But there’s hope. The report shows how the best builders and buyers are crossing the “GenAI Divide”: ✅ Focus on learning-capable systems with memory. ✅ Start narrow, integrate deeply, then expand. ✅ Partner instead of building everything in-house. ✅ Treat AI like a business service, not just software. And a fascinating insight: the real AI adoption is already happening in the “shadow AI economy”—employees using personal ChatGPT accounts daily while corporate pilots stall. The next wave? Agentic AI—systems that remember, adapt, and coordinate across the web. The window to lock in these capabilities is closing FAST. 👉 If you’re investing in AI today, ask not “Which model?” but -> "Am I automating the right thing?", and -> “Does it learn and integrate with our workflows?” At Bardeen, we built our Work Intelligence Platform to solve that exact adoption problem. Hit me up if you want to get real ROI from your AI pilots! https://lnkd.in/djEWXUCY (link to full study in first comment)

Do not invest in an AI tool to do something for the first time. Invest in an AI tool to do something you already do much faster.

Seems like we’ve entered the „optimization“ area now of LLMs. We’ve been there with Computer Vision models a few years ago, the big breakthroughs, weather CNNs then or LLMs now are „miked“ with the obvious things: all available training data on as much compute as you can get your hands on. That scales amazingly in the early phases of the logistic growth curve but it does flatten out at some point (now). The key now is the known machine learning optimization techniques to squeeze more performance out of the models and, more importantly, make more specialized models for each use case. GPT-5 is that for tool calling (Agentic workflows) and coding tasks (where the money is!). Much much cheaper and more optimized for generic tasks, and the „routing models“ are going to become important parts, to figure out which model to use for which exact use-case and situation. Exciting times ahead and I think this marks the first inning of real value creating with LLMs in the application of the models to real world and business problems. Researchers in the best labs all over the world will surely come up with the next stepping stone to propel model performance, but for now it’s game-time for optimization and application 🚀🚀🚀

We are seeing something VERY similar. In the world of GTM automation, there are SO many tools, SO many symptoms, and the approach of the last few years has been to attach a single tool (or tools) to a single symptom/problem. Example: "my team doesn't have a good way to reach out to leads, so we buy ZoomInfo, Apollo or Clay to 'get leads' faster". This is a symptom of a larger issue: our team doesn't have operational bandwidth to reach out to the quantity of prospects we need to drive our TOFU. GTM leaders are under-diagnosing the larger disease/issue. Like, big time. All the time. Bardeen's Work Intelligence tool goes under the hood and watches your team work, producing actionable, specific data to drive operational decision making. Not gut instincts or obvious symptoms. We surface the REAL underlying issues before they spread into resentment and likely tools being underused.