The story of how Taiichi Ohno helped our Gen AI initiative
Let me take you through the unlikely journey of how an underappreciated aspect of process intelligence reshaped the trajectory of a Gen AI initiative. It is not flashy. It is not cool. But it is time tested, and it works.
Interested? Read on.
TLDR: Most Gen AI projects in operations do not fail because of weak models. They fail because the process they are meant to augment is only half understood. The real unlock comes from layering process intelligence end to end: the intended design, the execution patterns, and the human reality uncovered only through Gemba. A simple example from industry classification shows how each layer reveals hidden nuances that completely change how AI should be evaluated. If you want meaningful lifts in operational AI, start by understanding how the work is truly done.
WHY ENTERPRISE AI STALLS?
Enterprise AI adoption is nowhere near expectations.
The stock market calls it a bubble. McKinsey in its state of AI report notes that over 67% percent of projects have not taken off. Illiya Sutskever recently remarked that the ‘economic impact is dramatically behind’.
From my experience in operational AI, the missing ingredient in many cases is the lack of process intelligence.
Not model tuning.
Not larger context windows.
Not more GPUs.
Just a weak understanding of the process the model is meant to support.
MISSING LAYER – FULL STACK PROCESS INTELLIGENCE
Process intelligence is not one concept. It is a stack.
It has three layers:
* The intended design that sits in the process repository
* The execution patterns observed through process mining and task mining
* The human reality that drives how work is actually done
Most teams capture the first two. Very few understand and acknowledge the third.
Focus groups and demos cannot give you the tacit knowledge. They capture only the dominant and palatable version of the process. The exceptions, the shortcuts, the edge cases and the tribal knowledge emerge only through Gemba. That is, by observing operators in action, repeatedly and without filters.
HOW DOES THIS SHAPE GEN AI IMPLEMENTATION?
Let me illustrate this with an underwriting example.
Classifying a business into the right industry code is a foundational task. If you get this wrong, every downstream decision suffers. On the surface, the task is simple. Read the description. Compare it to a taxonomy. Pick the closest SIC or NAICS code. Prompt it. Evaluate it on a golden batch.
Done? Not really.
Here is how the understanding evolved.
STAGE 1: The Big Picture View
From the process repository and a few demos, the task looks straightforward.
You classify the business and match it to the code.
You think a simple prompt should work.
Evaluation should align with human choices. 95% of the times.
Except, it does not.
Why?
STAGE 2: The Process Mining View
Process mining shows that the average handling time is X minutes, but the distribution is bimodal. Some cases are very quick. Some take disproportionately long.
Why the split?
STAGE 3: The Task Mining View
Task mining shows that in long cases, operators do not rely only on the business website. They jump between many sources. Aggregators. Filings. News. Prior data. And sometimes pure intuition built over experience.
This is not randomness. It is accumulated judgment.
Why do they do this?
STAGE 4: Gemba Walk (or) watch
Only when you sit with operators, or better, do the task yourself, do the real insights emerge.
Questions start surfacing.
What if a business claims to do multiple things?
What if the business has almost no online presence?
When different sources list different services, which one do you trust?
What quirks in the classification system require compensating logic?
Should success really be measured by exact alignment?
**That last question flips the entire evaluation strategy.**
Many Gen AI projects fail not because the model or prompts are inadequate, but because the evaluation criteria are disconnected from operational reality.
FULL-STACK PROCESS INTELLIGENCE SHAPES AI SUCCESS
The real work is not linear. It is iterative. You slice through each layer of process intelligence, and you adjust your AI evaluation framework accordingly.
You start asking better questions. Process intelligence helps you with those questions.
This is the boring superpower. The unglamorous part that does not make headlines. Yet it determines whether AI initiative succeeds or stalls.
FINAL THOUGHTS
In Gen AI implementation, the model is rarely the problem.
The real unlock comes from the overlooked layers of process intelligence, especially the human reality you uncover only through Gemba.
Next time you want a next lift in your operational AI program? Plan a Gemba walk, or better, try becoming the operator.
PRO TIP
We were fortunate to accelerate this learning with a seasoned veteran like Steve guiding the process. Someone who has seen enough operations to know where the real constraints and real signals lie. Try finding your Steve.