DMAIC: Improve and Solution Optimisation

Introduction

The Improve phase of DMAIC converts verified causes of process problems into tested, practical and sustainable solutions. It is not simply a brainstorming exercise and it is not the stage at which a team selects the most attractive idea. Improve requires disciplined creativity, risk analysis, experimentation, economic evaluation and controlled implementation.

In the Define and Measure phases, the problem is specified and the process is described with reliable data. In Analyse, relationships between inputs and outputs are investigated and probable root causes are verified. Improve then addresses a central question: What change will produce the required performance improvement with acceptable risk, cost and operational disruption?

A technically effective solution may fail because employees cannot use it, suppliers cannot support it, information systems are incompatible, regulatory requirements are overlooked or the financial benefit is overstated. Therefore, solution optimisation considers performance, risk, cost, feasibility, customer value and sustainability together.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Generate improvement alternatives using structured creativity methods.
2. Screen and prioritise solutions using weighted decision matrices.
3. Assess solution risk using Failure Mode and Effects Analysis (FMEA).
4. Design and interpret practical experiments for process optimisation.
5. Explain mistake-proofing, simulation, piloting and implementation planning.
6. Evaluate solutions using cost-benefit and financial measures.
7. Develop an evidence-based recommendation and transition it to the Control phase.

1. The Logic of the Improve Phase

Improve follows a logical chain:

Verified cause → candidate solutions → screened alternatives → risk assessment → experiment or simulation → economic evaluation → pilot → implementation plan → Control.

The aim is not necessarily to maximise one output. Most processes have several responses. For example, a hospital discharge process may need to reduce waiting time while maintaining medication accuracy and patient safety. A manufacturing process may need to reduce defects without increasing energy consumption or operator fatigue. This is a multi-objective decision problem.

A useful solution statement identifies:

• the verified input or cause being changed;
• the expected effect on the critical-to-quality characteristic;
• the operating conditions under which the solution should work;
• potential negative effects;
• required resources and controls; and
• the evidence required before full deployment.

Improvement should also respect the distinction between common-cause variation, which is inherent in the current system, and special-cause variation, which results from an unusual event. A solution aimed at a verified common cause changes the process system. A response to a special cause may instead require containment, standardisation or removal of an abnormal condition.

2. Generating Solutions Through Structured Creativity

Creativity in Lean Six Sigma is purposeful: ideas must be connected to the verified causes and the customer or business requirement. Unstructured brainstorming can produce many ideas but may favour senior voices, familiar solutions or ideas that are easy to describe rather than effective.

2.1 Divergent and convergent thinking

Solution development has two complementary modes:

• Divergence: generate a broad range of possible solutions without premature criticism.
• Convergence: evaluate, combine and narrow alternatives using explicit criteria.

Separating these modes improves participation and reduces the risk that early criticism suppresses useful ideas.

2.2 Useful creativity methods

Brainstorming generates ideas from a defined problem statement. The facilitator should establish rules such as deferring judgement, building on others’ ideas and recording every suggestion.

Nominal group technique asks participants to generate ideas individually, share them in turn and then rank them. It reduces dominance by highly vocal participants.

SCAMPER prompts alternatives by asking whether the process can be Substituted, Combined, Adapted, Modified, Put to another use, Eliminated or Reversed.

Six Thinking Hats separates perspectives: facts, emotions, risks, benefits, creativity and process control. It is useful when a team is arguing from different assumptions.

Value-stream analysis looks for opportunities to eliminate waiting, transport, excess processing, inventory, motion, defects and unused human capability.

TRIZ-inspired contradiction analysis asks how one process characteristic can improve without worsening another. For example, a team may seek faster inspection without reducing inspection accuracy.

Analogy and benchmarking transfer ideas from another department, industry or technology. Such transfer must be tested because contextual differences can make an apparently successful practice unsuitable.

A strong facilitation practice is to convert each idea into a testable hypothesis: “If input X is changed from condition A to condition B, response Y will improve without violating requirement Z.”

3. Screening and Prioritising Alternatives

A long list of ideas requires transparent selection. Common criteria include expected impact, implementation effort, cost, time, safety, customer value, regulatory compliance and reversibility.

3.1 Weighted decision matrix

A weighted decision matrix makes preferences explicit. Let the weighted score for alternative j be:

where:

• S_j is the total score for alternative j;
• w_i is the weight assigned to criterion i;
• r_ij is the rating of alternative j against criterion i;
• m is the number of criteria; and
• the weights sum to 1 when expressed as proportions.

The rating scale must be defined, for example 1 = very poor and 5 = excellent. All criteria must be oriented consistently: a high rating should always mean a desirable result. If “implementation cost” is used, a low cost should receive a high rating, or the criterion should be renamed “cost attractiveness.”

For Alternative A, the score is 0.35×4 + 0.20×5 + 0.20×4 + 0.15×5 + 0.10×3 = 4.25. Although Alternative B has a stronger expected defect reduction, Alternative A is preferred under these stated weights because it is easier to implement and has greater customer benefit.

The matrix does not prove that A is objectively best. It provides an auditable decision structure. Sensitivity analysis should be used: if a small change in weights reverses the ranking, the team should recognise that the decision is fragile and gather more evidence.

3.2 Screening gates

Some criteria are not trade-offs. Legal compliance, patient safety, information security and essential technical specifications may be must-pass gates. An alternative failing such a gate should not be selected merely because it scores highly on cost or speed.

4. Failure Mode and Effects Analysis

FMEA is a preventive method for identifying how a proposed or existing process could fail, why it might fail, what effect failure would have and how the risk can be reduced. It is especially valuable before implementation because a new solution can introduce new failure modes.

Typical FMEA fields include process step, failure mode, effect, cause, current prevention controls, current detection controls, severity, occurrence, detection and recommended action.

Traditional FMEA uses a Risk Priority Number:

where S, O and D are ordinal ratings, commonly from 1 to 10, for severity, occurrence and detection difficulty. A high detection rating means poor detectability; it does not mean that detection is good. RPN is a prioritisation aid, not a probability of failure and not a precise measure of risk.

For a failure mode with S = 8, O = 4 and D = 6, RPN = 8×4×6 = 192. The team should not automatically treat 192 as more urgent than every lower RPN. A severity-10 failure may require action even when occurrence is low, and modern FMEA practice often gives priority to severity and action thresholds rather than relying on RPN alone.

FMEA actions should preferably reduce:

• severity, by designing out the harmful condition;
• occurrence, by preventing the cause; or
• detection difficulty, by improving detection before the output reaches the customer.

Prevention is normally stronger than detection. An alarm that identifies a wrong component is useful, but a keyed fixture that makes the wrong component impossible is generally more robust.

5. Designing Experiments for Optimisation

When the team is uncertain which process settings produce the best result, designed experiments (DOE) provide a structured way to learn. DOE changes selected factors deliberately and observes one or more responses. It is more informative than changing one factor at a time when factors may interact.

5.1 Factors, levels and responses

A factor is an input that can be controlled or deliberately varied, such as temperature, speed, staffing level or software setting. A level is a selected value of a factor. A response is the measured output, such as defect rate, cycle time or strength.

A two-level full factorial design with k factors has:

basic treatment combinations, before adding replicates or centre points. Thus, three two-level factors require 2³ = 8 combinations. Replication estimates experimental error; randomisation reduces bias from time-related or nuisance effects; blocking controls known nuisance variation such as batch or shift.

5.2 Main effects and interactions

A main effect describes the average change in a response when one factor changes level. An interaction exists when the effect of one factor depends on the level of another. For two factors A and B, a simple model may be written:

where y is the response, β₀ is the intercept, βA and βB are main-effect coefficients, βAB is the interaction coefficient, xA and xB are coded factor values, and ε represents unexplained error. The model is appropriate only when its assumptions and adequacy have been examined. In practice, residual analysis, replicated observations and subject-matter knowledge are required.

5.3 Practical DOE discipline

Before running an experiment, define the response operationally, specify measurement methods, identify safety constraints, determine factor ranges and confirm that the process can be returned to a safe state. Do not deliberately test unsafe conditions merely to obtain statistical information.

If the response is a defect proportion, normal-theory methods may be inappropriate, especially with small samples or proportions near 0 or 1. Generalised linear models or appropriate attribute-data methods may be required. Statistical significance should also be distinguished from practical significance: a small improvement can be statistically detectable but operationally irrelevant.

6. Mistake-Proofing and Robust Process Design

Mistake-proofing, or poka-yoke, designs the process so that an error is prevented or immediately made obvious. It should address the likely human interaction with the system rather than assume perfect attention.

Examples include:

• connectors that fit only the correct port;
• barcode verification before order dispatch;
• software fields that reject impossible dates;
• fixtures that allow loading in only one orientation;
• checklists integrated at the point of work; and
• automatic interlocks that stop a machine under unsafe conditions.

A warning light is weaker than a prevention mechanism because it depends on noticing and responding. However, prevention may not be feasible in every context. In such cases, detection should occur as close as possible to the point of error, with clear escalation and recovery instructions.

Robust design seeks acceptable performance despite normal variation in noise factors such as temperature, supplier differences, operator experience or demand. It should not be used to conceal a fundamentally incapable process or to avoid understanding root causes.

7. Simulation as a Decision Tool

Simulation represents the behaviour of a process so that alternatives can be evaluated without exposing the live operation to unnecessary risk. Discrete-event simulation is suitable for queues, service systems, production lines and logistics; system dynamics is useful for feedback, accumulations and long-term policy effects; agent-based simulation can represent interacting individuals or entities.

A credible simulation requires:

1. a clearly stated purpose and boundary;
2. verified logic, meaning the model has been built correctly;
3. validated behaviour, meaning the model represents the real system adequately for its intended use;
4. realistic input distributions or scenarios;
5. consideration of warm-up, run length and random variation where applicable; and
6. sensitivity analysis showing how conclusions change under assumptions.

Simulation outputs are estimates, not facts. A model predicting a 20% reduction in waiting time does not establish that the reduction will occur unless the assumptions, data and implementation conditions are credible. Stakeholders should be able to inspect the model’s logic and limitations.

8. Economic Evaluation of Solutions

An improvement must create value or satisfy a justified strategic, safety or regulatory requirement. Cost-benefit analysis compares relevant incremental benefits and costs.

A simple net benefit is:

For a recurring annual benefit and cost, a simple benefit-cost ratio is:

A BCR greater than 1 indicates that estimated benefits exceed estimated costs under the stated assumptions. Present value is used when cash flows occur at different times. For a cash flow CFt received at time t, discounted at rate r:

Here r is the discount rate per period and t is the number of periods. The formula assumes discrete compounding and cash flows measured consistently in the same currency and period.

Payback period is a simple measure of the time required for cumulative cash benefits to recover the initial investment. If annual net cash benefit is constant:

This is an approximation and ignores discounting, benefits after payback and uncertainty. Costs should include training, data migration, validation, downtime, maintenance, quality assurance, change management and unintended consequences—not only purchase price.

Benefits may be hard savings, cost avoidance, capacity release, revenue protection, risk reduction, customer retention or regulatory protection. Teams should distinguish measured benefits from assumptions and assign ownership for benefit realisation.

9. Pilot Testing and Implementation Planning

A pilot is a controlled, limited implementation used to test technical performance, human usability, operational effects and control requirements. It is not a demonstration arranged only to confirm the preferred solution.

A pilot protocol should specify:

• scope, location, duration and participants;
• baseline and target measures;
• inclusion and exclusion conditions;
• training and support arrangements;
• data collection methods;
• safety, privacy and escalation provisions;
• success and stop criteria; and
• decision rules for adoption, redesign or abandonment.

Before-and-after comparisons can be misleading when demand, staffing, seasonality or case mix changes. Where feasible, use a comparison group, staggered rollout or repeated observations. The evaluation should consider both average performance and variation. A lower mean cycle time accompanied by unstable performance may not be a genuine improvement.

An implementation plan typically includes a work breakdown structure, responsibilities, resources, milestones, dependencies, communication plan, training plan, risk register and transition criteria. A RACI matrix can clarify who is Responsible, Accountable, Consulted and Informed.

The Improve phase ends successfully when the selected solution is demonstrated under representative conditions and the Control phase has the information needed to sustain it. Control plans should identify key process measures, owners, reaction rules, audit frequency and procedures for responding to drift.

10. Integrated Illustrative Case: Reducing Laboratory Sample Rejection

This is a fictional illustrative case designed for learning.

Background and problem

A hospital laboratory reports that 7.2% of emergency-department blood samples are rejected because of labelling errors, insufficient volume or haemolysis. Rejected samples delay diagnosis and may require repeat venepuncture. The project team has already confirmed that rejection is not evenly distributed: most events occur during high-demand periods and involve handover between collection and transport.

Analysis and solution generation

The team maps the process and verifies three important causes: labels are printed away from the patient, collection tubes are sometimes selected before patient identification is completed, and pneumatic-tube transport is used for samples vulnerable to haemolysis. Ideas include bedside label printing, a barcode scan before collection, a revised tube-selection guide, transport restrictions and a short competency intervention.

Prioritisation

The team applies a weighted matrix using defect-reduction potential, patient safety, feasibility, cost and implementation speed. Bedside label printing and barcode verification score highest, but the team identifies that the two changes should be evaluated together because their effects may interact.

FMEA

A proposed failure mode is “barcode scanner unavailable during peak demand.” Severity is rated 8, occurrence 5 and detection 7, giving RPN = 280. Rather than merely instructing staff to check the scanner, the team proposes a daily readiness check, spare charged scanners and a defined downtime procedure. A second failure mode is “correct patient selected but label printed for another encounter.” The solution includes a two-identifier confirmation and a system lock requiring confirmation before printing.

Pilot and experiment

A four-week pilot is conducted on one emergency-department shift. The team compares three conditions: current process, bedside printing alone and bedside printing plus barcode verification. Shift and day are recorded because demand may affect performance. The primary response is rejected samples per 100 samples; secondary responses are collection time, patient complaints and staff-reported usability.

The pilot shows that the combined intervention reduces rejection substantially, while bedside printing alone produces inconsistent results. However, collection time initially increases because staff are unfamiliar with the scanner. Additional practice and interface simplification reduce the time penalty. The team does not claim that the pilot proves long-term success; it identifies the combined solution as suitable for a broader, monitored rollout.

Economic and operational evaluation

Costs include printers, scanners, software configuration, training and maintenance. Benefits include reduced repeat collections, reduced laboratory rework and lower delay-related clinical risk. The financial case is accepted only after separating hard savings from risk-reduction benefits. The implementation plan includes phased deployment, weekly review during the first month, equipment readiness checks and a control chart for rejection rate.

Outcome and learning points

The case illustrates that solution optimisation is iterative. The best solution was not the cheapest isolated change; it was a system intervention whose components supported one another. FMEA revealed operational weaknesses, the pilot exposed a usability problem, and economic evaluation prevented an exaggerated claim. The Control phase must ensure that the improved process remains reliable during peak demand, staff turnover and technology downtime.

11. Governance, Ethics and Contemporary Technology

Improvement teams have ethical obligations. Patient, employee and customer safety must not be traded casually for efficiency. Data used in pilots and AI-enabled optimisation must be lawful, relevant, secure and appropriately anonymised. People should understand when automated recommendations affect them, and a responsible human should be able to review consequential decisions.

AI can support idea generation, process mining, anomaly detection, simulation input preparation and prediction of failure risk. However, AI-generated recommendations may reproduce historical bias, misinterpret context or create false confidence. Models require validation, version control, access controls, monitoring and clear accountability. An AI tool should not replace process observation, expert judgement or the obligation to test a change safely.

Contemporary improvement practice increasingly combines Lean visual management, digital process mining, sensor data, remote monitoring and advanced analytics. These technologies are valuable when they improve decision quality and process capability—not when they add data collection without a clear operational purpose.

12. Common Failure Modes in the Improve Phase

Several errors recur:

• selecting an idea before verifying the cause;
• confusing enthusiasm with evidence;
• using an unweighted matrix whose criteria duplicate one another;
• treating RPN as an exact probability;
• changing many factors without a learning plan;
• running a pilot too small or too short to represent normal variation;
• ignoring adverse effects on safety, workload or equity;
• counting theoretical savings as realised benefits; and
• handing over a solution without ownership, reaction rules or training.

A mature team documents assumptions, uncertainty and rejected alternatives. This creates organisational learning and makes later review possible.

Chapter Summary

The Improve phase of DMAIC transforms verified causes into tested and economically justified solutions. Effective teams begin with structured creativity, then screen alternatives using explicit criteria and non-negotiable compliance gates. Weighted decision matrices improve transparency, while FMEA identifies new and residual risks. Designed experiments reveal factor effects and interactions under controlled conditions; simulation supports safe evaluation of complex systems. Mistake-proofing prevents or detects errors at the point of work. Cost-benefit analysis clarifies financial and strategic value, but must include implementation and lifecycle costs. Pilots test technical performance, usability and operational consequences before full deployment. Finally, implementation planning connects the improved process to Control through ownership, monitoring and reaction rules. The central principle is evidence-based optimisation: a solution is selected not because it is appealing, but because it performs, is safe, is feasible and can be sustained.

Key Terms

DMAIC: Define, Measure, Analyse, Improve and Control, the structured Six Sigma improvement cycle.

Improve phase: The DMAIC stage in which solutions are generated, tested, selected and prepared for implementation.

Divergent thinking: Expanding the range of possible ideas without premature evaluation.

Convergent thinking: Narrowing alternatives using evidence and decision criteria.

Weighted decision matrix: A tool that combines criterion weights and alternative ratings into comparative scores.

Screening gate: A mandatory requirement that an alternative must satisfy before further comparison.

FMEA: Failure Mode and Effects Analysis, a preventive method for assessing process failure risks.

Risk Priority Number: The product of severity, occurrence and detection ratings in traditional FMEA.

Mistake-proofing: Designing a process to prevent errors or make them immediately evident.

Designed experiment: A planned test in which factors are deliberately varied to study their effects on responses.

Factor: An input variable deliberately changed or controlled in an experiment.

Interaction: A condition in which the effect of one factor depends on another factor.

Replication: Repeating experimental conditions to estimate variation and experimental error.

Randomisation: Assigning experimental runs in an order that reduces systematic bias.

Simulation: A model-based representation used to study process behaviour under alternative conditions.

Pilot: A limited implementation used to evaluate a solution before wider deployment.

Net benefit: Total estimated benefit minus total estimated cost.

Benefit-cost ratio: Present value of benefits divided by present value of costs.

Control plan: A documented arrangement of measures, owners and reaction rules for sustaining performance.

Residual risk: Risk remaining after improvement actions have been implemented.

Review Questions

Short Answer Questions

1. What is the purpose of the Improve phase in DMAIC?
2. Why should divergent and convergent thinking be separated?
3. What are the three traditional dimensions used in FMEA ratings?
4. Distinguish a factor from a response in a designed experiment.
5. Why is a pilot preferable to immediate full-scale implementation?

Long Answer Questions

1. Explain how a weighted decision matrix can support, but not replace, managerial judgement.
2. Describe how FMEA should be used to improve a proposed solution rather than merely rank risks.
3. Explain the role of interactions in designed experiments and why one-factor-at-a-time testing may miss them.
4. Discuss the financial, operational and ethical considerations in evaluating an improvement proposal.
5. Develop an implementation plan for a process change, including training, ownership, risk management and transition to Control.

Analytical / Application Questions

1. A team rates three solutions against impact, feasibility, safety and cost. Explain how you would test whether the preferred solution remains preferred when the criterion weights change.
2. A proposed automation project reduces average cycle time but increases variation and creates a new safety failure mode. Analyse whether the project should proceed and identify additional evidence required.
3. Design a pilot for reducing customer-contact waiting time. Specify the response measures, comparison approach, potential confounding variables, success criteria and Control-phase measures.

Multiple Choice Questions with Answers

Question 1

Which sequence best represents disciplined solution optimisation in DMAIC?

A. Implement, brainstorm, measure, control

B. Generate, screen, assess risk, test, pilot and implement

C. Control, analyse, purchase, train

D. Measure, implement, estimate benefits, define

Correct Answer: B. Generate, screen, assess risk, test, pilot and implement

Question 2

In a weighted decision matrix, what does a criterion weight represent?

A. The number of employees involved

B. The financial cost of the alternative

C. The relative importance assigned to the criterion

D. The probability that the solution will fail

Correct Answer: C. The relative importance assigned to the criterion

Question 3

If a failure mode has severity 8, occurrence 4 and detection 6, what is its traditional RPN?

A. 18

B. 48

C. 192

D. 384

Correct Answer: C. 192

Question 4

Which action is generally the strongest form of mistake-proofing?

A. Placing a warning poster near the workstation

B. Asking an operator to remember an additional inspection

C. Designing a fixture that physically prevents incorrect assembly

D. Reviewing the error at the end of the month

Correct Answer: C. Designing a fixture that physically prevents incorrect assembly

Question 5

How many basic treatment combinations are in a two-level full factorial experiment with four factors?

A. 4

B. 8

C. 12

D. 16

Correct Answer: D. 16

Question 6

What does an interaction mean in a designed experiment?

A. The response has no measurement error

B. The effect of one factor depends on the level of another factor

C. All factors have identical effects

D. The experiment has no control variables

Correct Answer: B. The effect of one factor depends on the level of another factor

Question 7

A benefit-cost ratio greater than 1 generally indicates that:

A. Estimated costs exceed estimated benefits

B. Present value of benefits exceeds present value of costs

C. The project has no implementation risk

D. The payback period is exactly one year

Correct Answer: B. Present value of benefits exceeds present value of costs

Question 8

What is a principal purpose of a pilot?

A. To guarantee that no further monitoring is needed

B. To test a change under limited and controlled conditions

C. To replace all process analysis

D. To avoid collecting baseline data

Correct Answer: B. To test a change under limited and controlled conditions

Question 9

Which statement about simulation is most accurate?

A. Simulation results are facts independent of assumptions

B. Simulation eliminates the need for process validation

C. Simulation can compare alternatives, but its conclusions depend on model validity and inputs

D. Simulation is appropriate only for manufacturing processes

Correct Answer: C. Simulation can compare alternatives, but its conclusions depend on model validity and inputs

Question 10

Which item most directly connects an implemented solution to the Control phase?

A. A list of untested ideas

B. A control plan with owners, measures and reaction rules

C. A brainstorming attendance sheet

D. A supplier brochure

Correct Answer: B. A control plan with owners, measures and reaction rules