Six Sigma had a discipline that modern analytics projects quietly dropped: you earned the right to build a solution. I earned my Black Belt at Motorola, the company that created the method in the 1980s, and the part that has aged best is not the statistics. It is the order of work.
The solution came fourth
The method runs in five steps: define, measure, analyze, improve, control. The fix belongs to the fourth. Before a team could propose one, it had to size the problems in front of it, pick the one worth solving and run small studies to find which factors drove it. Only then did the investment follow.
Measurement was expensive back then, so the method forced economy. Nobody could afford to measure everything, so the first argument was always about what deserved measuring. That argument was the discipline.
Cheap building changed the order
Today building is cheap and fast, so teams start there. A model takes weeks, a pilot takes a quarter, and a demo can be ready before anyone has written down what it is for. What got skipped is the part that made the old projects pay: proving which problem is worth the investment before making it.
The evidence points the same way. In 2024, RAND researchers interviewed 65 experienced data scientists and engineers about why AI projects fail. The cause they heard most often was not the model or the data. It was a misunderstanding of what problem the project was meant to solve. Their advice to leaders reads like a Six Sigma charter: focus on the problem, not the technology.
Example
A voice bot and 45 roles that were still needed
In July 2025, Commonwealth Bank of Australia, the country’s largest bank, announced that 45 roles in its customer service business were no longer needed. The cuts were linked to a new AI voice bot designed to reduce the calls reaching its staff.
A month later the bank reversed the decision. The Finance Sector Union said call volumes had in fact been rising, and that the bank was offering overtime and putting team leaders on the phones to cope. The bank apologized to the employees and said its initial assessment “did not adequately consider all relevant business considerations,” which meant the roles were not redundant.
The tool was built and running. What had not been done was the first step of any improvement project: measuring the problem before acting on the solution. The decision rested on a result the bank’s own call data could have tested first.
The cost that enforced the discipline is gone
Cheap solutions made expensive problem definition feel optional. It isn’t. It just stopped being enforced by cost.
When a solution cost millions to build, someone had to prove the problem first, because the budget would not clear without it. At the price of a pilot, nothing forces the question, so the proof gets skipped.
Earn the right to build
Bringing the discipline back does not require the old bureaucracy. It requires four answers in writing before the build starts.
- Size the problems. Put a number on each candidate: money, hours or service lost per year.
- Pick one. Choose the problem whose size justifies the investment, and say which ones are waiting.
- Find the drivers. Run a small study on existing data to see which few factors move the result.
- Define done. State the measure that will show the fix worked, and who will check it.
This is why we start with a diagnosis to bring clarity to the problem before anything is built. A team that cannot give the four answers has not earned the build yet.
Sources: RAND, “The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed,” 2024; Finextra, “CommBank reverses plan to replace call centre staff with AI,” 2025; HRD Australia, “CBA reverses AI-triggered employee layoffs,” 2025.