blog

Transfer a Proven Improvement Between Plants: A Matched-Pilot Playbook

By: Lauren Dunford

By: Guidewheel
Updated: 
November 16, 2026
9 min read
Transfer a Proven Improvement Between Plants: A Matched-Pilot Playbook

No items found.

A cross-plant improvement transfer pilot should test one defined intervention under comparable recipient conditions. Document what changed at the source plant, choose a suitable recipient line and agree the outcome before the trial. Match the relevant jobs and shifts, record concurrent changes and compare results with a credible baseline or comparison group.

Standard definitions make the records comparable. They do not prove that the intervention will work at another site. Use the pilot to decide what transfers, what needs local adaptation and what should stop.

Pretium needed a common view of operations across a growing factory network. In Guidewheel’s case study, shared runtime, production and energy information enabled its operations team to benchmark facilities, exchange practices and prioritize improvement opportunities. That shows a practical use for cross-site learning. The matched-pilot method and numerical example below are recommendations for testing a transfer, not experiments reported by Pretium.

Pilot decision Evidence to retain
Copy the intervention Exact change and source operating conditions
Choose the recipient Comparable product, equipment, staffing and loss mechanism
Evaluate the result Fixed metric, matched periods and concurrent-change log
Expand, adapt or stop Outcome, guardrails and remaining uncertainty

Describe the intervention that produced the source result

Transfer a specific change with evidence of its effect. “Use the leading plant's process” is too broad to reproduce or test.

Write a source record naming the loss, intervention, scope, implementation date and responsible team. Identify the sequence or resource that changed. Preserve the before-and-after operating conditions, the result measure and any other action introduced at the same time.

For example, a candidate might be an approved material-staging routine that reduces waiting during changeovers. The transferable element is the staging arrangement, timing, ownership and verification. It is not the source plant's entire operating system, and it is not an instruction to alter machine safeguards or approved process limits.

Ask the source team why it thinks the change helped. Did the loss actually occur at the targeted step? Was the new method used consistently? Did inspection or scrap results change? Was improved performance sustained beyond the first supervised trial?

Use a short source card:

Field Question it answers
Loss mechanism What delay or defect was the intervention intended to remove?
Exact change What behavior, resource or sequence differed?
Prerequisites What staffing, product or equipment conditions made it possible?
Result and scope Which measure changed, where and over which runs?
Concurrent actions What else could explain the result?
Evidence owner Who can reproduce the supporting record?

A source result that lacks a clear intervention is a research lead, not a transfer-ready practice. Return to the source plant to resolve it before asking the recipient to invest time in copying an unclear method.

Choose a recipient and document the important differences

Choose a recipient that experiences the same loss mechanism and can safely test the intervention. Similar plant names or headline overall equipment effectiveness (OEE) scores are not enough.

Compare the product family and complete process route before choosing the recipient. Then check job size, equipment, tooling or recipe, staffing, shift pattern and quality requirements. Record differences that could change the intervention's effect. A staging method designed for long repeat jobs may need adaptation for short jobs with frequent material changes.

Distinguish a manageable difference from an invalid comparison. Different storage locations might require a local staging design. A different customer specification or approval process may prevent the same method from being tested at all. Let the responsible engineering and quality owners make that call.

NIST's experimental-design guidance on blocking explains controlling factors that affect the measured outcome but are not the intervention of interest. In a plant trial, those factors might include product family, crew or material lot. Keep the important conditions alike when comparing runs, and record the conditions you cannot hold constant.

Condition Source/recipient question Pilot treatment
Product, route and specification Does the work require a comparable process and accepted-output boundary? Match or restrict the scope
Job size and change type Are changeovers comparable? Compare within declared groups
Equipment, tooling or recipe Can the intervention operate within the approved conditions? Confirm compatibility
Staffing and operating window Are the required people and support available during comparable jobs? State local adaptation
Material and inspection Could quality, handling or release requirements differ? Retain as guardrail and context
Available records Can each site link its events to jobs, output and quality results? Close the data gap or narrow the comparison

Matching is not random assignment. It improves comparability but cannot eliminate every hidden difference. State that limitation when choosing the design, rather than adding a stronger causal claim after a favorable result.

Agree a bounded trial before changing the process

Fix the primary outcome, trial scope and acceptance criteria before implementation. Changing the goal after observing results makes an otherwise useful pilot hard to assess.

For the material-staging example, the primary outcome might be changeover duration from the agreed last-good to first-good boundary. Keep the same event definition in baseline and trial. Review first-pass quality, startup scrap and staffing demand as guardrails so a faster changeover does not hide another loss.

Choose comparable jobs and enough repeated observations to assess consistency. If trial order can be randomized within approved operations, use that option. If it cannot, explain the practical restriction and retain the sequence. A matched operational pilot can still inform a decision, with a narrower claim than a randomized experiment.

The multi-plant OEE implementation guide supports the shared record. Resolve standing definitions in the OEE governance definition sheet before the intervention trial; otherwise a changed denominator can masquerade as a result.

Use a trial charter with:

  1. The exact intervention and approved local adaptation.
  2. Recipient scope, baseline and comparison method.
  3. Primary outcome, unit and event boundary.
  4. Required repetitions and the planned analysis.
  5. Quality, safety and delivery guardrails.
  6. Concurrent changes to hold constant or record.
  7. Stop conditions, owner and decision date.

Capture the fields that could change the conclusion. Confirm that both sites can supply them for the agreed operating window. If one site records only machine activity while the other records accepted output, reconcile that boundary before comparing results. Keep run-level results so reviewers can distinguish a consistent effect from one unusual job.

A pilot should stop or pause if the intervention breaches an approved safety or quality condition, undermines required delivery, or cannot be implemented as specified. Document the reason. A trial that identifies an unsuitable transfer can still prevent an expensive network-wide mistake.

Read a matched pilot without claiming more than it proves

Compare the recipient's change with what happened under comparable conditions without the intervention. Treat the result as bounded evidence, not automatic proof of cause.

The following trial is illustrative, with invented averages. A recipient line and a comparison line each contribute 10 baseline changeovers and 10 trial-period changeovers. The agreed measure is mean minutes per changeover. The intended product and changeover mix is matched across the periods.

Line Baseline mean Trial-period mean Reduction
Recipient using the new method 50 minutes 35 minutes 15 minutes
Comparison retaining its method 48 minutes 44 minutes 4 minutes

The recipient improved by 15 minutes per changeover. The comparison improved by four minutes. The difference in those changes is 15 − 4 = 11 minutes per changeover. Across 10 trial changeovers, that corresponds to 110 minutes of recovery beyond the observed comparison trend.

This arithmetic does not make the 11-minute difference a proven causal effect. The illustration has no randomized treatment assignment or run-level variation estimate. A material change, supervisor intervention or unequal job mix could still explain part of it. Review those possibilities before deciding to expand.

In a real trial, inspect the individual changeovers. Compare the agreed statistic within the matching groups and review variation, excluded runs and guardrails. If one difficult job disappeared from the trial, the favorable mean may reflect scheduling rather than the intervention. Do not change from mean to median simply because the alternative gives a better result.

Retain the comparison-line record too. An improvement there may reflect a plant-wide change that affected both lines. Record it rather than ignoring it. If no suitable comparison exists, use a clearly labeled before-and-after design and narrow the conclusion accordingly.

The pilot's decision is practical: whether the evidence justifies another bounded application under stated conditions. Statistical significance, persistence and network-wide financial value require additional evidence appropriate to those claims.

Capture the transferable method and the local adaptations

Write down what the recipient actually used and what the result supports. Keep local adaptations visible so the next plant can assess fit.

Disposition What justifies it Next step
Adopt locally Outcome improves with guardrails intact and credible implementation Confirm persistence in normal operations
Adapt and retest A documented local difference limits the result Revise one bounded element and repeat
Hold Evidence is incomplete or concurrent changes obscure the answer Resolve the specific gap
Stop transfer Method fails feasibility, guardrails or outcome Record why and return to alternatives

Separate the intervention's required elements from local choices. A required material-ready check may transfer, while the staging location and responsible role differ. Preserve the rationale so local flexibility does not remove the mechanism that produced the source result.

Guidewheel's software, manufacturing experts and machine information can support cross-plant operating reviews. Optional integrations may add the context an agreed implementation needs. The intervention's approval, comparison design and acceptance decision remain responsibilities for your team.

Discuss a specific cross-plant improvement trial with Guidewheel.

Frequently asked questions

Does the best-performing plant supply the best pilot?

It supplies a candidate only when a specific intervention explains a relevant result. A high score may reflect product mix, staffing or measurement differences. Transfer the documented mechanism, then test it where comparable conditions exist.

Is matching the same as randomization?

No. Matching aligns observed operating conditions; randomization determines treatment or run allocation by chance. Use the feasible approved design and describe its limits. Matching cannot remove every unobserved difference.

What if the comparison line also improves?

Retain that result and investigate shared changes. Compare the recipient's improvement with the comparison trend rather than attributing its whole before-and-after change to the intervention. The remaining difference still needs scrutiny.

When should a transfer trial stop?

Pause or stop when a pre-agreed safety, quality or delivery condition fails, or when the intervention cannot be implemented as specified. Record the evidence and owner. Do not trade an essential guardrail for a favorable headline measure.

Can a successful pilot justify network-wide rollout?

It can justify the next decision within its demonstrated scope. Confirm persistence and assess each new plant's relevant conditions before expansion. A successful recipient trial does not establish universal fit across the network.

About the Author

Lauren Dunford is the CEO and Co-Founder of Guidewheel, a FactoryOps platform that empowers factories to reach a sustainable peak of performance. A graduate of Stanford, she is a JOURNEY Fellow and World Economic Forum Tech Pioneer. Watch her TED Talk: the future isn’t just coded, it’s built.

GradientGradient