A quality team approves an AI vision model for a glass-container line. The model performs well for weeks, then false rejects begin to rise. The forming process is stable. The inspection equipment is functional. The cause is less obvious: a lighting adjustment, new mold set, and model-threshold update changed the inspection system in combination, but each change was reviewed by a different function.
Traditional continuous improvement governance is built to control changes to the process. AI adds another object that can change: the intelligence system interpreting the process. Data definitions, model versions, thresholds, prompts, features, interfaces, and confidence rules may shift even when equipment and standard work remain unchanged.
That is how AI changes continuous improvement governance. Manufacturers must govern two connected baselines, then make five recurring decisions about how an AI-supported improvement is admitted, released, operated, changed through learning, and retired.
Continuous Improvement Now Has Two Controlled Baselines
A conventional improvement changes some combination of material, method, machine, measurement, environment, or work practice. The organization evaluates the change, updates controlled information, trains affected people, verifies results, and sustains the new condition.
An AI-supported improvement contains a second baseline that can influence the same result.
| Controlled baseline | Typical contents | What can change |
|---|---|---|
| Operating-process baseline | Equipment configuration, material requirements, process parameters, control plan, inspection method, maintenance condition, standard work, qualification, and response plan | Tooling, settings, suppliers, methods, limits, staffing, product mix, inspection, or work sequence |
| Intelligence baseline | Intended use, data sources, context variables, model or rule version, thresholds, output meaning, uncertainty treatment, interface, permitted action, monitoring, and fallback | Training data, features, model version, prompt, rule logic, threshold, data pipeline, user interface, confidence requirement, or automation boundary |
The baselines are not independent. A new resin, mold, camera, route, product family, work instruction, or inspection standard can invalidate assumptions in the intelligence baseline. A new model version or threshold can change how the operation responds even though the process documentation is untouched.
Dual Baseline Governance for AI Enabled Continuous ImprovementDual-Baseline Governance requires every significant change to one baseline to trigger an impact check against the other. The purpose is not to send every adjustment through a large committee. It is to prevent a change from escaping control merely because it occurred in software, data, or model configuration rather than on the factory floor.
Why Existing Improvement Governance Is Not Enough
AI introduces four governance complications that conventional improvement systems often handle poorly.
Evidence Can Change Without the Process Changing
A sensor mapping, defect label, data-cleaning rule, image crop, or context join can change what the model sees. The operation may appear unchanged while the evidence entering the decision has moved.
Performance Can Drift Without a Formal Revision
Product mix, supplier behavior, equipment wear, environmental conditions, and user behavior can move away from the conditions represented in development data. No one intentionally revised the model, yet its fitness for use has changed.
Learning Creates Pressure to Update Faster
AI teams can retrain models, modify prompts, or tune thresholds quickly. Continuous-improvement teams may welcome the responsiveness, but speed can bypass the process knowledge, validation, training, and controlled-documentation disciplines that protect manufacturing performance.
Responsibility Is Distributed
The process owner controls operating results. The quality function controls product and system requirements. The AI team controls model development. Technology teams control data and deployment. Frontline users encounter actual conditions. A decision that sits between those responsibilities can remain unowned even when every function completes its assigned work.
The answer is not a separate bureaucracy for every model. It is a governance design that makes joint responsibility visible at the moments when authority matters.
The Five Governance Decisions
Dual-Baseline Governance organizes oversight around five recurring decisions. These are not project phases or another implementation framework. They are management decisions that recur across the life of an AI-supported improvement.
1. Admit: Is AI an Appropriate Part of This Improvement?
The Admit decision determines whether the proposed AI role is legitimate for the performance problem and process context. It should confirm that the problem is operationally meaningful, that established improvement methods remain intact, and that the proposed use of AI is proportionate to the risk.
Required evidence includes:
- A defined performance gap and process owner
- The specific AI task and intended user
- Known alternatives, including non-AI countermeasures
- Potential effects on safety, quality, delivery, cost, people, customers, and compliance
- A preliminary decision boundary and fallback
The decision does not select projects across the portfolio; that belongs to use-case prioritization. It decides whether AI is an appropriate mechanism inside this particular improvement.
2. Release: May the Capability Enter Operating Use?
The Release decision determines whether the process baseline, intelligence baseline, workflow, users, and evidence are ready for a defined operating scope. Model accuracy alone cannot answer that question.
A release record should identify the approved product, process, site, user, model version, data configuration, threshold, action boundary, monitoring method, fallback, and expiration or review condition. Any untested condition should be named rather than hidden inside a general approval.
The AI-Enabled Operational Excellence Framework designs the full capability. Governance uses that design to make an explicit release decision and preserve the approved baseline.
3. Operate: How Are Exceptions, Overrides, and Emerging Risks Handled?
Operating governance begins after approval. It covers low-confidence output, missing data, unfamiliar conditions, conflicting evidence, user overrides, unexpected recommendations, process excursions, and signs that the model no longer represents the work.
The operating rule should make four things plain:
- When the AI output may be used
- When a person must confirm, reject, or escalate it
- When the process must revert to a known fallback
- Which events require investigation or suspension
This is narrower than redesigning the enterprise operating model. It is the control needed to keep one approved capability within its intended conditions.
4. Learn: Which Evidence May Change the Process or Intelligence Baseline?
AI can reveal new patterns and collect rich operating feedback. Governance must determine which observations are signals, which are hypotheses, and which justify a controlled change.
A recommendation should not retrain itself into authority. Before a model, threshold, prompt, response plan, or process standard changes, the organization should know:
- What evidence triggered the proposed change
- Whether the cause lies in the process, intelligence system, or both
- What conditions and risks are affected
- What validation is required
- Who approves the revised baselines
- How users and controlled information will be updated
This is where the discipline behind bounded experimentation and verified learning becomes essential. More frequent evidence should improve the quality of change, not weaken the standard for accepting it.
5. Retire: When Must the Capability Be Suspended, Rolled Back, or Removed?
Continuous-improvement governance often focuses on adoption and sustainment, while retirement remains informal. AI makes that omission risky. A model can remain visible and influential after its data, process assumptions, support resources, or intended use have expired.
The Retire decision should define triggers such as persistent performance loss, unacceptable false actions, unavailable support, major process redesign, obsolete data sources, unresolved security or compliance concerns, or a superior non-AI method. Retirement also requires a known replacement or fallback and retention of the evidence needed for traceability.
Who Must Participate in AI Improvement Governance?
Governance should assign accountability by decision, not create a ceremonial list of stakeholders. A practical responsibility model includes five perspectives.
| Perspective | Core accountability | Evidence contributed |
|---|---|---|
| Process owner | Operating performance and process consequences | Target condition, process state, constraints, and outcome evidence |
| Improvement or method owner | Integrity of the problem-solving and improvement method | Problem definition, causal analysis, countermeasure logic, and sustainment method |
| Intelligence owner | Fitness, limits, monitoring, and configuration of the AI contribution | Model evidence, data lineage, uncertainty, version, and technical monitoring |
| Quality, safety, or risk authority | Independent challenge and approval where consequence requires it | Requirements, risk evaluation, validation criteria, audit evidence, and escalation |
| Frontline user | Correct use and direct evidence from actual work | Usability, abnormal conditions, overrides, local context, and observed effects |
The same person may hold more than one role in a small organization. What matters is that the evidence and authority are explicit. The broader Industrial Intelligence management system owns enterprise roles and routines; Dual-Baseline Governance applies them to the specific decisions created by AI-supported improvement.
Case One: Glass-Container Inspection Drift
A glass-container plant uses AI vision to identify finish defects, checks, blisters, and dimensional anomalies. The model was approved for a defined family of containers, camera configuration, lighting setup, mold set, line speed, and defect taxonomy.
Over several weeks, false rejects rose. Maintenance replaced lighting components. Production introduced a new mold set. Quality refined one defect category. The AI team adjusted a confidence threshold to restore throughput. Each action was reasonable in isolation; together, they changed both baselines.
Under Dual-Baseline Governance:
- The lighting and mold changes trigger an intelligence-impact review.
- The defect-taxonomy revision triggers a review of labels, training evidence, and performance reporting.
- The threshold change is treated as a controlled intelligence-baseline revision, not a harmless software setting.
- Release evidence is rechecked for the affected container family.
- False rejects, escapes, throughput, and user overrides are reviewed together.
The governance mechanism does not prevent improvement. It preserves the relationship between the inspection decision and the evidence supporting it.
Case Two: Metal-Stamping Die Protection
A metal-stamping operation uses tonnage signatures and acoustic signals to warn of die wear, feed problems, and emerging damage. The system recommends inspection or controlled stoppage before a severe event. After die maintenance, the signal pattern changes. Operators override repeated warnings because the die has just been serviced. The AI team proposes recalibrating the threshold from new production data.
The governance question is not simply whether the new threshold reduces alarms. The team must determine whether maintenance restored the intended process, whether the post-maintenance signal is normal, whether a sensor or installation condition changed, and whether the proposed threshold reduces protection against an actual failure mode.
The process owner, maintenance owner, intelligence owner, and safety or quality authority review the two baselines together. The change can then be validated for the serviced die condition, released within a defined scope, and monitored for missed events as well as nuisance alarms.
The Minimum Governance Evidence Pack
Governance becomes burdensome when every review starts from scattered documents and verbal history. A concise evidence pack should make the current state inspectable.
- Purpose and scope record: performance aim, process, users, intended AI role, and exclusions
- Dual-baseline record: approved process and intelligence configurations with their relationship
- Decision and fallback record: permitted actions, confirmation, escalation, and reversion method
- Validation record: technical performance, operating trial evidence, boundary conditions, and known limitations
- Operating evidence: outcomes, countermeasures, overrides, low-confidence events, incidents, and drift indicators
- Change history: what changed, why, supporting evidence, approvals, affected conditions, and effective date
- Retirement readiness: suspension triggers, fallback owner, retained records, and replacement plan
This evidence supports internal assurance without turning the governance review into a generic audit tutorial. Manufacturers can connect it to existing quality-system practices, including the discipline described in ISO 9001 internal auditing, while applying AI-specific criteria to intended use, data, model behavior, and impact.
How GAO AI Accountability Framework Supports Governance Direction
The U.S. Government Accountability Office developed an AI Accountability Framework organized around four complementary principles: governance, data, performance, and monitoring. Its practices are intended to help managers, auditors, and assessors consider responsible AI across selection, implementation, and continuous monitoring.
The GAO framework is broader than manufacturing continuous improvement. This article applies its accountability direction to the operating relationship between a physical process and the intelligence influencing decisions about that process. Dual-Baseline Governance does not reproduce the GAO framework; it gives operational-excellence teams a specific control mechanism for connecting governance, data integrity, operating performance, and ongoing monitoring.
Governance Should Accelerate Responsible Improvement
Poor governance creates two opposite failures. Too little control allows model, data, and process changes to drift apart. Too much undifferentiated review sends low-risk adjustments and high-consequence changes through the same slow path.
Proportionate governance distinguishes scope, consequence, reversibility, novelty, uncertainty, and automation authority. A low-consequence advisory tool may use delegated review and frequent monitoring. An AI function influencing safety, product release, validated processes, or customer commitments needs stronger independence, evidence, and change control.
The objective is not to slow AI improvement. It is to make change faster where evidence and authority support it, and deliberately slower where consequences require deeper validation.
Control the Process and the Intelligence Together
AI does not remove continuous-improvement governance. It expands what governance must see. Manufacturers now have to control the process that creates the product and the intelligence system that influences how people understand and improve that process.
Dual-Baseline Governance makes that relationship explicit. The five governance decisions keep authority focused where it matters: admitting an appropriate use, releasing a complete capability, managing real operating conditions, learning through controlled change, and retiring systems that are no longer fit.
When both baselines remain aligned, AI can accelerate improvement without quietly weakening the discipline that makes improvement sustainable.
Frequently Asked Questions
What is AI continuous improvement governance?
AI continuous improvement governance is the set of decisions, responsibilities, controls, and evidence used to ensure that AI-supported improvement remains fit for its intended process, users, risks, and operating outcomes.
What is Dual-Baseline Governance?
Dual-Baseline Governance jointly controls the operating-process baseline and the intelligence baseline. A significant change to equipment, materials, methods, data, models, thresholds, interfaces, or action boundaries triggers an impact review across both.
Does every model change require executive approval?
No. Approval should be proportionate to consequence, scope, reversibility, uncertainty, and authority. The governance design should delegate low-risk changes while preserving independent review for high-consequence applications.
How is AI governance different from model monitoring?
Model monitoring measures technical or operating behavior. Governance determines who interprets that evidence, what conditions require action, who may approve changes, when use must be restricted, and how decisions remain traceable.
How does governance relate to the AI-Enabled Operational Excellence Framework?
The six-block framework designs the operating capability. Governance controls the recurring decisions that admit, release, operate, change, and retire that capability throughout its useful life.
