August 26, 2026
TL;DR: AI handles 80 percent of most industrial processes well and struggles with the long tail. Many companies obsess over that last 20 percent and won't deploy until it's solved, forgetting they already employ people who handle exactly those cases every day. The faster path: deploy at 80 percent, keep the expert in the loop, and invest in UI/UX that makes reviewing and correcting the AI's output fast. Every correction teaches the system the special processes that made 100 percent unrealistic in the first place.
Talk to industrial customers about AI for long enough and the conversation always ends up in the same place: the edge cases. Not the thousands of routine documents, orders, or quality checks the system handles well, but the strange ones. The order format one customer has used since 1997. The tolerance exception that only applies to one product family, in one plant, in winter. The conversation starts with "can AI do this?" and within ten minutes it's "but what about this one case?"
That's the 80/20 rule doing what it always does. AI handles the bulk of a process reliably, then accuracy drops off on the long tail of exceptions. Nothing surprising there. What surprises me is where the attention goes. Many industrial companies fixate almost entirely on that last 20 percent, and the fixation is understandable. It comes from a deep, earned love of process stability. These are organizations that spent decades driving error rates toward zero. A system that's right "only" 80 percent of the time sounds, to that culture, like a system that's broken.
But this framing quietly discards something these companies already have: people who are genuinely good at the hard 20 percent. The planner who knows why that one customer's orders look wrong but aren't. The engineer who recognizes the exception before the rule is even checked. This competence exists today, it's already paid for, and it's exactly what the difficult cases need. Treating the human as a stopgap until AI reaches 100 percent gets the picture backwards. The human is the solution to the 20 percent, right now.
The honest version of the goal sounds like this: yes, 100 percent correct from day one is where we want to end up. At the moment it's unrealistic, and the reason is not that the models are weak. It's that industrial companies have spent years, sometimes decades, building special processes on top of special processes. Workarounds that became standards. Customer-specific exceptions that nobody ever wrote down. An AI system has to learn all of that, and it can only learn it in production, from the people who carry that knowledge. Waiting for perfection before deploying means the learning never starts.
Meanwhile the first 80 percent sits there, and it's not small. That's the majority of the daily volume: the repetitive, well-structured work where AI already performs well and where the humans currently doing it add the least value. Leaving that on the table because the remaining fifth isn't solved yet is an expensive form of caution.
So the interesting question isn't "how do we get the model to 100 percent." It's "how do we build a system where 80 percent from the machine plus 20 percent from the human reliably equals 100." And that turns out to be less a modeling problem than a UI/UX problem.
An interface designed for this has to do two things well. First, it has to present "80 percent right" honestly. The user needs to see at a glance what the system did, which parts it's confident about, and which parts deserve a second look. A wall of output with no signal about certainty forces the user to check everything, and at that point you've automated nothing. Second, it has to make the correction to 100 percent cheap. Fixing a field should take seconds, not a detour through three screens. If reviewing and correcting the AI's work takes longer than doing the task manually, people will do it manually, and they'll be right to.
Get those two things right and something useful happens: every correction becomes training data. The special processes that made 100 percent unrealistic in the first place get captured, case by case, by the people who know them. The 20 percent shrinks. Not because someone waited for a better model, but because the system was in production, learning from the one source that actually holds the knowledge.
The companies making progress with AI in their processes aren't the ones with the most ambitious accuracy targets. They're the ones who accepted 80 percent as a starting point, kept their experts in the loop for the rest, and invested in the unglamorous work of making review and correction fast. Process stability is a fine goal. It's just not something you get by keeping the process unchanged.
