By Adonis Partners
Ninety-five percent of AI initiatives fail to deliver a return. That number gets repeated often enough that it has started to sound like a technology problem. It isn’t. In most operations Adonis Partners has worked inside, the AI wasn’t the failure point. The process underneath it was.
AI does not fix a broken process. It automates it faster, which means it also fails faster and at greater scale.
Two things show up over and over in operations that struggle to get value from AI.
The first is automating a process nobody has actually mapped. Leadership approves an AI initiative because a vendor demo looked compelling or a competitor announced something similar, without first walking the floor to confirm what the current process actually does, where it breaks, and why. When the AI layer goes in, it inherits every workaround, every manual patch, and every undocumented exception the team has been quietly managing for years. The output looks fast. It is fast and wrong.
The second is treating AI as a headcount reduction tool instead of a throughput tool. Executives approve budget to eliminate roles, not to improve the operation those roles support. That framing puts the initiative on the wrong side of the P&L from day one. Cost cutting without a corresponding process redesign just moves the bottleneck to wherever the remaining people are stretched thinnest.
Both patterns share a root cause: leaders trying to automate a system they do not fully understand.
This is not a call for more documentation. Adonis Partners’ consultants see plenty of operations with process maps sitting in a shared drive that nobody follows. Understanding a process means knowing where variation actually comes from, which steps are load-bearing and which are legacy habit, and what the real cycle time looks like when nobody is watching the clock.
That is root cause work. It is the same discipline behind Lean Six Sigma and DMAIC, applied before automation instead of after a failed pilot. Define the problem in operational terms. Measure the current state honestly, including the parts that are embarrassing. Analyze where the actual constraint sits. Only then does it make sense to ask what should be automated and what should be redesigned first.
Skipping straight to automation because the technology is available is the same mistake operations leaders have made with every prior wave of tooling, ERP included. AI just runs the mistake at a faster clock speed.
A few signals tend to predict which AI initiatives are heading for the 95 percent:
The business case for the initiative describes a technology capability, not an operational outcome. “We’re implementing AI-driven scheduling” is a technology statement. “We’re cutting changeover-driven downtime by 20 percent” is an outcome, and it forces the team to define the current process before touching a tool.
Nobody on the project can explain the current process without opening a slide deck. If the people closest to the work can’t describe it from memory, the AI layer is about to encode someone’s best guess as permanent logic.
The project timeline has no measurement phase. Teams that jump from “select vendor” to “deploy” skip the step where they would have discovered the process wasn’t ready.
Operations that get a return from AI tend to do the unglamorous work first. They document the current state with real cycle time and defect data, not estimates. They fix the process issues that don’t need a single line of code, standard work, clear handoffs, defined escalation paths, before adding automation on top. Then they automate the parts of the process that are stable enough to trust running without a human catching the exceptions.
That sequencing is slower at the start and faster everywhere after. It’s also the difference between an AI initiative that shows up as a case study and one that shows up as a write-off eighteen months later.
AI is not the risk. An unexamined process is. Leaders who understand their operation in detail, who can name the actual constraint and the actual variation sources, are the ones positioned to get real value from automation. Leaders who skip that step are buying speed for a system that was never going to hold up at scale.
If the process hasn’t been root-caused, that’s the place to start. Not the AI vendor list.
Adonis Partners helps operations leaders find the real constraint before they automate around it. Talk to a process improvement consultant about your roadmap.
