Why Most AI Pilots Stall — and What Actually Makes Them Pay Off

Jan de Vries · 2026-11-05

Most companies have tried AI by now. Far fewer have managed to make it stick. That gap between the pilot and the production floor is where a lot of money quietly disappears, and it's exactly the problem the folks at AppliedAI set out to solve when they started building Opus back in 2021. Opus isn't another chatbot wrapper. It's an AI transformation platform that rewires how enterprises actually run — and ties every single change back to the bottom line. ### The real reason pilots fall apart Here's the thing. AI models aren't the bottleneck anymore. When a pilot fails, the reasons are almost always operational. Old systems are painful to connect. Nobody measures the results. And governance shows up as an afterthought, if it shows up at all. Xenia Schmahl, who leads Opus' push into the DACH region, puts it bluntly: three things kill pilots. - They're built on assumptions instead of the real process, so the first exception breaks them. - Integration and governance come too late. - Nobody owns a clear target, and there's often no way to measure success at all. Sound familiar? It should. This is the pattern across most of Europe right now. ### Redesign first, automate second > "AI pays off when we redesign the process, not just automate it." That line is the whole philosophy in a nutshell. Opus starts by mapping how a process really runs today — the messy version, with all its workarounds — and measures a baseline: time, cost, quality, error rates. Then it redesigns the workflow for AI and calculates the expected return before anything gets built. Once the new workflow goes live, every run and every decision — whether made by an AI agent or a person — gets recorded. Over time, that builds a digital twin of the organization's processes. Companies can show what happened, why, when, and based on which data. In regulated industries, that audit trail isn't a nice-to-have. It's the price of admission. ### Why DACH, and why now The DACH market is demanding. Regulation is strict, data has to stay in the region, and buyers expect governance from day one. But that's precisely what makes it promising. Banks, insurers, and healthcare providers there run huge, document-heavy back offices — exactly where reliable, well-governed AI creates the most value. Opus draws on its Work Knowledge Graph, a proprietary model of over a million enterprise processes organized by industry and jurisdiction. That's the shortcut that lets teams prove a single process on real data within weeks instead of quarters. ### How to measure success beyond cost savings Cost savings are easy to report, but they're rarely the biggest win. Schmahl recommends five measures: - **Capacity:** Can the team handle more work without adding headcount? - **Quality:** How accurate is each field, for each case type? Measure on test data that was never used for training — a single "97% accurate" number hides too much. - **Exceptions:** A high automation rate means little if wrong results slip through. - **Ownership:** Someone has to own the outcome, not just the tool. - **Return:** Value should be clear before the project starts and provable after go-live. ### Start small, but start for real The advice here is refreshingly practical. Pick one process. Prove it on real data within weeks. Expand from there. And know where AI doesn't belong — if an existing system already handles a process end to end, leave it alone. Not every process that can use AI should use AI. That's a lesson a lot of enterprises in the US are learning the hard way right now, too. The ones that get it right aren't the ones with the flashiest models. They're the ones that redesigned the work before they touched the tech.