DMAIC step by step: a Lean Six Sigma project from problem to control

Most improvement projects fail in one of two ways. Either the team jumps to a solution before understanding the problem, or it understands the problem beautifully and the gain quietly fades six months later. DMAIC is designed to stop both.

It's the problem-solving backbone of Six Sigma, which Bill Smith introduced at Motorola in 1986. (Motorola's original version was MAIC; the Define phase came later, once companies learned how many projects failed for lack of a clear problem.) It works just as well in an office as on a production line. This note walks through all five phases on one example, with the tools that matter in each.

The example: giving product away

A food plant fills 500 g packs of rice on a 12-head filling machine, about 2 million packs a year. The plant's rule: the average must be at least 500 g, and no pack may go below 495 g. Every gram above 500 is product given away for free. Finance suspects the giveaway is large, but nobody knows how large.

Define: agree on the problem

The first deliverable is a short project charter. The hard part is the problem statement. It should describe the gap in numbers, without blaming anyone and without hinting at a solution.

Charter element	Example
Problem statement	Average pack weight on Line 3 is estimated at 505 to 510 g against a 500 g label, costing an estimated €50k to €60k a year in product given away
Goal	Reduce average giveaway by at least 50% by the end of Q2, with no pack below 495 g
Scope	Line 3 filler, 500 g rice packs. Out of scope: other lines, packaging material
Team	Line lead, maintenance technician, quality analyst, a process engineer as project lead
Sponsor	Plant manager

Then map the process at a high level with a SIPOC, so everyone sees the same boundaries:

Suppliers	Inputs	Process	Outputs	Customers
Rice supplier, film supplier, maintenance	Rice, film, filler settings	Feed, weigh, fill, seal, check-weigh	Filled 500 g packs	Retailers, consumers, finance

Notice what the goal is not: "install a new filler". That's a solution, and it doesn't belong in Define.

Measure: get a number you can trust

First, check the measuring device. Before collecting a baseline, the team runs a measurement system analysis on the check-weigher: several people weigh the same packs several times. If the scale's own variation is a big share of the process variation, every conclusion after this is built on sand. A common rule of thumb from the automotive industry's MSA manual: under 10% of the variation from the measurement system is good, 10% to 30% may be acceptable, above 30% is not.

Then collect a baseline. Five packs every hour from each shift for two weeks gives a solid picture:

Now the project has a real number. It also has a clue: with a spread of 3.1 g, operators had nudged the target upwards to stay safely away from 495 g.

Analyze: find the causes, and prove them

Brainstorm possible causes with a fishbone (people, methods, machines, materials, measurement, environment), then test the likely ones with data. The rule in Analyze is simple: a cause isn't a root cause until the data says so.

Suspected cause	How it was tested	Result
Some filler heads fill differently	Weights split by head, compared with an ANOVA test	Confirmed. Heads 4 and 9 run lower and far more erratically
Worn valve seals on those heads	Inspection, then weights before and after swapping seals on one head	Confirmed. Spread on that head fell by more than half
Operators set a high target to compensate	Settings log compared with shift reports	Confirmed. Target raised twice after low-weight rejects
Rice from the second supplier is denser	Weights split by supplier lot	Not confirmed. No meaningful difference

This is the phase where DMAIC saves the most money, because it stops you from fixing things that aren't broken. The supplier theory was the most popular one in the room, and the data didn't support it.

Improve: fix the causes, then pilot

The fixes follow directly from the verified causes:

• Replace the worn seals, and add seal checks to the weekly preventive maintenance.
• Add a weight check per head at start-up, so a drifting head is caught in minutes, not days.
• Once the spread is down, lower the target to what the new variation safely allows.

Pilot on one shift for a week, then roll out. After the rollout:

The order matters. Lowering the target first, before reducing the spread, would have produced underweight packs. Variation came down first; the average followed.

Control: make the gain hold

This is the phase most teams skip, and the one that decides whether the project was worth doing. The deliverable is a control plan: what to watch, how, how often, and what to do when it drifts.

What	How	How often	Limit	Reaction	Owner
Average pack weight	Check-weigher, control chart	Every hour	Control limits from the new baseline	Stop, check heads, log cause	Line lead
Weight per head	Start-up check, 5 packs per head	Every start-up	Any head more than 2 g from target	Adjust or isolate the head	Operator
Valve seal condition	Visual and wear check	Weekly	Wear standard	Replace seal	Maintenance
Target setting	Settings log	Every shift	503 g, changes need approval	Reset, escalate	Shift supervisor

The control chart is the heart of it: it separates normal variation from a real change, so people react to signals and not to noise. How to read one is its own note.

Tollgate questions

At the end of each phase, the sponsor reviews the work before the team moves on. These are the questions worth asking:

Phase	Ask before moving on
Define	Is the problem stated in numbers, with no solution hidden in it? Is the scope small enough to finish in 3 to 6 months?
Measure	Did we check the measurement system? Is the baseline based on enough data, over enough time?
Analyze	Is every root cause backed by data? Did we test the popular theories too?
Improve	Does each fix address a verified cause? Did a pilot show the result? What could go wrong?
Control	Is there a control plan with owners? Is the process owner ready to take it over?

When DMAIC is the wrong tool

DMAIC is thorough, which also makes it slow. Don't use it for everything:

• The cause and fix are obvious. Just do it, and check it worked.
• The problem is small and local. A one-week kaizen event with the people who do the work is faster.
• The process doesn't exist yet. Designing something new calls for DMADV (Define, Measure, Analyze, Design, Verify) instead.
• The problem is mostly waste and flow, not variation. Lean tools like value stream mapping may get you there sooner.

A note on sigma levels

"Six Sigma" refers to a process so consistent that it produces about 3.4 defects per million opportunities. That figure assumes the conventional 1.5 sigma shift, which allows for drift over the long term. Without the shift, six sigma would mean about 0.002 defects per million, which is why some statisticians question the convention. The table uses it, as most references do:

Sigma level	Defects per million	Yield
2	308,537	69.1%
3	66,807	93.3%
4	6,210	99.38%
5	233	99.977%
6	3.4	99.99966%

Most processes never need six sigma. The point of the scale is direction: fewer defects, less variation, and the ability to prove it.

Test a fix before you touch the line ↗Vrolen models a production line, compares countermeasures before any change, and checks afterwards that the gain actually held.

Sources

• ASQ. DMAIC process: define, measure, analyze, improve, control
• ASQ. What is Six Sigma?
• Barney, M. (2002). Motorola's second generation. Six Sigma Forum Magazine, ASQ.
• U.S. Army (2011). Lean Six Sigma Deployment Guidebook, version 5.0, on tollgate reviews.
• Coşkun, A. & Ialongo, C. (2020). Six Sigma revisited. Biochemia Medica, 30(1), on the 1.5 sigma shift.
• Automotive Industry Action Group (2010). Measurement Systems Analysis (MSA) Reference Manual, 4th edition.

The plant and its numbers are illustrative.