Process Improvement with AI: A 90-Day PoC and RFP Guide for Thai Factories
For process improvement with AI in a Thai factory, the first decision is not the model. It is which loss must be reduced, whose decision must change, and by when. Comparing quality, downtime, yield, energy, and planning use cases—and defining data, acceptance criteria, and exception handling before development—turns a PoC into an investment decision rather than a technology demo. This guide covers a 90-day PoC, RFP, FAT/SAT, MLOps, Thailand PDPA, multilingual operations, and return on investment from the buyer’s perspective.
Search intent and conclusion: design decisions, not just prediction accuracy
Most readers want to know whether AI can work on their process, how long a PoC takes, what to put in an RFP, and how to prove value. Four principles answer those questions:
- Rank use cases by economic loss, ability to intervene, reproducibility, and scale—not merely by available data.
- A 90-day PoC must test operator action, integration, and fallback procedures as well as model performance.
- The RFP and acceptance plan must define baselines, evaluation periods, false alarms, missed events, leakage, drift, and retraining ownership.
- ROI should be based on avoided loss and incremental gross profit minus labor, cloud, maintenance, and change costs—not the number of “correct” predictions.
OECD’s review of AI in manufacturing discusses applications across quality, predictive maintenance, supply chains, and robotics. Its broader report on AI adoption in firms also addresses skills, data, cost, and organizational capability. The practical implication is to procure an operational change with data responsibilities, not an algorithm in isolation.
Prioritizing factory AI productivity use cases
Starting with the easiest data can produce a technically successful project with little financial value. Starting only with the largest loss can select a problem too rare or complex for a short PoC. Use one scorecard for all candidates.
| Criterion | Buyer’s question | Example recommended weight |
|---|---|---|
| Economics | Can annual loss, opportunity cost, cost of quality, or downtime cost be measured? | 30 |
| Intervention | Can people change settings, inspect, maintain, or reschedule after the signal? | 20 |
| Data readiness | Can time, product, machine, and outcome records be joined? | 20 |
| Testability | Can a control, historical baseline, or phased rollout support comparison? | 15 |
| Scalability | Can the method be reused on another line, product, or site? | 10 |
| Risk | Can safety, customer, legal, and personal-data risks be controlled? | 5 |
These weights are recommended design examples, not universal standards.
Quality: target cause isolation, not only defect prediction
Quality AI can link material lots, settings, environment, tools, operations, and inspections to surface conditions associated with higher risk. Existing customer specifications and approved inspection rules should remain authoritative; AI can begin as an additional warning and cause-assistance layer.
Do not accept overall accuracy alone. Measure misses, false alarms, performance by product and defect, and operator review time. With rare defects, a model that always predicts “normal” can look accurate. See our related guide to AI quality management for Thai manufacturing.
Downtime: connect warning to maintenance planning
Anomaly detection on vibration or temperature has little value unless the workflow states who checks it, within what time, how parts are reserved, and when a planned stop replaces emergency repair. If failure history is sparse, standardizing condition monitoring and inspection logs may be more defensible than rushing into supervised learning.
Track unplanned downtime, mean time to repair, warning-to-intervention time, emergency-maintenance share, unnecessary inspections, and ignored alerts alongside model metrics.
Yield: embed engineering constraints in recommendations
Yield analysis connects material, recipe, equipment settings, and quality outcomes. Correlation is not causation. Lock equipment capability, safety limits, and customer-approved ranges as constraints, and require authorized review before trials. Moving from advice to automatic control requires a separate safety and change assessment.
Energy: normalize consumption by production conditions
Total electricity cannot distinguish higher output from lower efficiency. Define meaningful intensity measures by product, weight, runtime, weather, startup, idle state, and utilities. AI can forecast demand, identify abnormal consumption, or propose operating sequences, but should be evaluated against quality and delivery objectives as well as energy.
Planning: preserve real shop-floor constraints
Planning models must account for due dates, changeovers, capacity, people, materials, tooling, and maintenance. Do not dismiss spreadsheet adjustments as “non-standard”; identify the constraint behind each override. Logging the proposed plan, executed plan, override reason, shortage, expedite, and changeover creates the feedback needed for improvement.
Preparing manufacturing data for AI analysis
Success depends more on semantic consistency than on column count. PLC data at one-second intervals, MES production records, ERP orders and items, quality databases, and maintenance logs cannot describe one event if their clocks and identifiers do not align.
Minimum data dictionary
| Item | Define explicitly | Common problem |
|---|---|---|
| Event time | Time zone, clock source, start/end semantics | PLC, server, and manual clocks differ |
| Production unit | Relationship among lot, batch, serial, and order | Traceability breaks at split or merge |
| Asset ID | Line, machine, unit, and sensor hierarchy | Old tags survive an equipment change |
| Quality outcome | Value, limit, retest, scrap, and rework | Only the final result is retained |
| Missingness | Not measured, network loss, stopped, out of range | All cases become zero or null |
| Version | Recipe, parameter table, model, and program version | Before/after comparison is impossible |
Create a data contract covering names, types, units, update rate, latency, missingness, retention, and owner. Calibration and equipment changes must trigger change notifications because they can shift data distributions.
Audit label quality
If defect, failure, and stop-reason labels vary by person, the model learns ambiguity. Review a representative period and define inclusion, exclusion, decision owner, and decision deadline. Preserve later corrections.
Split evaluation data by time
Manufacturing records are autocorrelated. Random splitting can put near-identical lots or operating periods in both training and test data. To simulate deployment, train on earlier periods and test on later ones, including product changes and post-maintenance conditions.
A 90-day factory process-improvement AI PoC
Ninety days is a recommended design example for one process and one operational decision, not a guaranteed duration. Add a readiness phase when extraction, contracting, or site access needs more time.
| Period | Main work | Exit decision |
|---|---|---|
| Days 1–10 | Confirm process, loss, decision, data, and risks | Approve problem statement and baseline |
| Days 11–25 | Extract, join, audit, and freeze evaluation set | Approve dictionary and quality report |
| Days 26–45 | Build simple baseline, model options, and explanation view | Establish value beyond simple rules |
| Days 46–65 | Back-test, analyze errors, set thresholds and workflow | Agree provisional acceptance criteria |
| Days 66–80 | Run shadow mode, user test, and integration test | Confirm people can act on the output |
| Days 81–90 | Estimate value and risk; decide scale or stop | Record Go, Conditional Go, or No-Go |
Write a one-sentence problem statement on Day 1
“Improve quality with AI” is too broad. A usable statement identifies object, timing, user, intervention, and outcome—for example: “Before inspection results are available, show the molding supervisor a risk alert and prompt checks within approved parameter ranges to reduce the selected defect loss.”
Compare against a simple baseline
Compare advanced models with current rules, moving averages, control charts, last value, or product-family averages. If complexity creates only a small incremental gain, the simple method may win after explanation, maintenance, and computing costs.
Observe decisions in shadow mode
Show recommendations without directly controlling equipment. Record who saw each alert, whether it was accepted, why it was rejected, and the actual outcome. This tests timing, explanation, authorization, and standard work—not only the model.
RFP requirements for process improvement with AI
An RFP should make proposals comparable and reduce contract ambiguity. “AI optimization” and “high accuracy” are not acceptance criteria unless the evaluation method is defined.
| RFP area | Buyer specifies | Evidence requested |
|---|---|---|
| Business scope | Process, user, intervention, exclusions | To-be workflow and RACI |
| Data | Period, grain, missingness, interfaces, export limits | Data needs, quality diagnosis, mapping |
| Performance | KPI, frozen set, thresholds, segmented results | Reproducible results and error list |
| Non-functional | Response, availability, monitoring, backup | Architecture, tests, recovery procedure |
| AI governance | Explanation, approval, version, drift, retraining | Model Card, change log, monitoring design |
| Security | Access, encryption, logs, vulnerability, transfer | Data flow, contract controls, evidence |
| Exit | Return of data, settings, models, documents | Export format and deletion procedure |
| Cost | Initial, usage, support, change, rollout | Assumption-based quote and three-year TCO |
The RFP should also state buyer inputs: process experts, ground-truth owners, interface contacts, test equipment, and approval deadlines. A vendor cannot resolve operational semantics alone.
Proving value with KPI, baseline, and counterfactuals
A before/after difference is not automatically an AI effect. Demand, product mix, material, weather, maintenance, staffing, and parallel kaizen may all change. Design the best feasible counterfactual: what would have happened without the AI intervention?
Four KPI layers
| Layer | Example KPI | What it tests |
|---|---|---|
| Model | Precision, recall, MAE, false alarm, miss | Prediction quality |
| Workflow | Response rate, intervention time, adoption, review time | Whether people used it |
| Process | Defect rate, downtime, yield, energy intensity | Whether the process changed |
| Finance | Avoided loss, incremental margin, operating cost, payback | Whether investment is justified |
Define numerator, denominator, exclusions, aggregation, source, cutoff time, and owner. Where possible, use a comparable line and difference-in-differences. Otherwise consider phased rollout, alternating operation, or matching similar product-shift combinations. Never compromise safety or customer requirements for experimentation.
Choose the comparison method
Set the smallest economically meaningful effect before the PoC. Rare failures may not provide enough events in 90 days; state that limitation in the RFP and combine proxy measures with longer follow-up.
FAT and SAT: accept the operating system, not only the model
Factory Acceptance Testing validates the integrated solution in the supplier or staging environment. Site Acceptance Testing validates it in the actual plant. Include missing data, network loss, sensor replacement, unknown products, and permission failure.
FAT test examples
- Reproduce results on a frozen test set.
- Report performance by product, asset, and shift.
- Detect out-of-range, missing, duplicate, and time-shifted inputs.
- Trace model version, features, settings, and approver.
- Control API retries, duplicate processing, and timeout.
- Verify role-based access, audit logs, and restore.
SAT test examples
- Receive PLC/MES/ERP data at the required rate.
- Validate Thai, English, and Japanese user interfaces.
- Process alerts according to standard work and RACI.
- Test breaks, shift changes, night work, and planned stops.
- Return safely to the existing method when AI is unavailable.
- Measure support escalation and recovery time.
Record open issues, conditional acceptance, owners, and deadlines. A successful PoC and production acceptance are separate gates.
MLOps and drift in production
Materials, products, machines, tooling, recipes, seasons, cameras, and inspection standards change. Monitor data drift, concept drift, and process drift.
Minimum operating register
| Controlled object | Minimum record |
|---|---|
| Data | Source, period, quality, transformations, permission |
| Model | Purpose, algorithm, features, evaluation, limitations |
| Deployment | Version, date, approver, assets, rollback target |
| Monitoring | Input/output distribution, outcome, incidents, usage |
| Change | Reason, impact, test, training, approval, notice |
Do not automatically retrain at every threshold breach. First distinguish a data fault, equipment change, and specification change. Protect retraining data from wrong labels and recent-product overfitting.
The NIST AI Risk Management Framework organizes work around Govern, Map, Measure, and Manage. ISO/IEC 42001 is an AI management-system standard; it is not certification of an individual model’s accuracy. An RFP should ask which controls and evidence apply to this project, not merely cite a standard.
Thailand governance and PDPA
Determine whether personal data is necessary. Even process records may relate to individuals when combined with operator IDs, images, voice, location, performance, or attendance. Review purpose, lawful basis, notice, access, retention, processors, cross-border transfer, deletion, and incident response with legal counsel and the DPO. This article is not legal advice.
ETDA’s AI governance information, its Generative AI Governance Guideline, and the Expanded ASEAN Guide on AI Governance and Ethics are useful references for organizational responsibility, risk, transparency, and human oversight. Confirm applicable law and contractual interpretation for each project.
Define accountability at the Thailand site
When headquarters owns the model and the Thai plant owns data and operations, define data owner, process owner, model owner, IT operations, security, DPO, and approver in a RACI. For 24-hour production, include night-shift response and escalation.
If generative AI explains causes or work instructions, separately evaluate hallucination, confidential prompts, prompt attacks, source provenance, and repeatability. Generated text alone should not change machine settings.
Multilingual shop-floor operation
Japanese management, regional English, and shop-floor Thai may coexist. Translation is insufficient if the meaning of an alert and required action diverge.
| Operational item | Localization requirement |
|---|---|
| Alert | Show cause candidate, urgency, deadline, and owner briefly |
| Standard work | Use the same version and approval date in every language |
| Training | Rehearse normal, false alarm, miss, and outage scenarios |
| Records | Use reason codes plus free text for analysis |
| Meetings | Freeze KPI definitions and share one dashboard |
Maintain a glossary for asset and defect names and preserve original free text alongside machine translation. Training must explain when to use the AI, when to doubt it, who may stop, and where to escalate. Our article on factory daily-report AI automation provides a complementary starting point for multilingual records.
ROI calculation
Avoid double counting. If defect reduction and yield improvement refer to the same scrap, do not add both. Use incremental gross profit—not revenue—only when demand and capacity support additional sales.
Annual net benefit = avoided quality loss + avoided downtime loss + energy savings + incremental gross profit + labor benefit − annual operating cost − annual change cost
Payback (months) = initial investment ÷ (annual net benefit ÷ 12)
Illustrative scenario—not an actual result or guarantee
Assume annual loss of THB 12,000,000, an AI-addressable share of 50%, a 10% improvement within that scope, and 60% attribution to the intervention:
Avoided quality loss = 12,000,000 × 50% × 10% × 60% = THB 360,000/year
Add only non-overlapping benefits and subtract cloud, monitoring, maintenance, user time, and retraining. If initial investment is assumed at THB 1,200,000 and annual net benefit at THB 600,000, simple payback is 24 months. These numbers are calculation examples, not market prices, customer results, or promised outcomes. Test pessimistic, base, and optimistic cases.
The World Economic Forum’s 2025 Global Lighthouse Network announcement presents advanced-site transformation examples. Any performance figures reported there represent selected Lighthouse leaders’ aggregated or individual cases; they do not guarantee results for an average factory or this project. Use your own baseline for ROI.
Vendor evaluation scorecard
| Area | Example recommended weight | Focus |
|---|---|---|
| Business understanding and value | 20 | Loss, decisions, counterfactuals |
| Data and model | 20 | Diagnosis, reproducibility, errors, limits |
| OT/IT integration | 15 | PLC/MES/ERP, failure modes, cybersecurity |
| Operations and MLOps | 15 | Monitoring, change, retraining, rollback |
| Security and governance | 15 | PDPA, access, logs, transfer, processors |
| Team and local support | 10 | Thai support, response, training |
| TCO and exit | 5 | Assumptions, data and configuration return |
Weights are design examples. Make safety and change control pass/fail gates where relevant. Give every bidder the same data, deadline, and KPI. Evaluate how they explain errors and scope changes, not only their feature list and polished demo.
Common failure patterns
Calling a technical demo a PoC
A good chart on historical data does not test action timing, ownership, integration, or fallback. Require shadow mode and user acceptance.
Accepting accuracy alone
It hides class imbalance, costly misses, and product differences. Review confusion matrices, segmented results, cost weighting, and false-alarm workload.
Outsourcing data meaning to the vendor
Only the plant can define operational semantics. Assign process and IT data owners to approve the dictionary and quality issues.
Connecting AI directly to automatic control
Stage adoption from advice, to approved action, to limited automation. Do not replace safety PLCs or existing interlocks without a dedicated safety process.
Funding the PoC but not operations
Include monitoring, retraining, data incident response, and training in three-year TCO and staffing.
Translating a headquarters template without localization
Thai equipment, shifts, approvals, connectivity, PDPA, and support differ. Separate global standards from site deviations and validate them in SAT.
FAQ
What is process improvement with AI?
It uses manufacturing data to support decisions through prediction, anomaly detection, cause suggestions, parameter recommendations, and planning. It includes collection, interfaces, workflows, monitoring, and governance—not only a model.
Which factory AI productivity use case should come first?
Prioritize a measurable loss, an actionable signal, joinable data, and a comparison feasible within the PoC. A very large but extremely rare or non-actionable loss may not suit the first project.
How many years of manufacturing data does AI analysis need?
There is no universal number. Representativeness across seasons, products, failures, and process changes matters more. Confirm enough target events, joinable time and IDs, and consistent labels.
Will a factory process-improvement AI PoC always finish in 90 days?
No. Ninety days is a recommended example for a narrow scope. Sensors, reviews, data-export approval, and rare-event observation can extend it. Separate readiness from model validation.
What acceptance accuracy should the PoC use?
There is no universal percentage. Derive thresholds from the cost of misses and false alarms, safety, and customer impact. Combine segmented model metrics with workflow and process KPIs.
How should the RFP address model ownership?
Separate rights to raw and processed data, features, code, configuration, weights, results, and logs. Define export, deletion, reuse, and third-party model restrictions at termination.
Does Thailand PDPA apply to process data?
Pure equipment values may not be personal data, but linkage to operator IDs, images, voice, location, performance, or attendance may change that. Consult the DPO and legal counsel.
Can generative AI perform root-cause analysis?
It can retrieve logs and manuals and draft hypotheses, but sources, hallucination, and confidential input must be controlled. Require evidence and human approval; never operate equipment from generated text alone.
Can a small factory adopt process-improvement AI?
Yes, but a custom model is not always the best first step. Compare visualization, control charts, rules, and standardized daily reporting, then invest only where AI adds measurable value.
Summary
Effective process improvement with AI connects economic loss to an operational decision. Select an actionable use case, establish a data contract, and use a focused PoC to compare simple baselines, shadow operation, counterfactuals, and FAT/SAT. Put MLOps, Thailand PDPA, multilingual work, and exit provisions in the RFP, and base ROI on your own measured baseline.
If you are assessing use cases, planning a PoC, drafting an RFP, or connecting PLC/MES/ERP systems in Thailand, you can contact TOMAS TECH at the concept stage. We can help structure the test scope, including non-AI alternatives.
References
- OECD: AI in manufacturing
- OECD: The adoption of artificial intelligence in firms
- World Economic Forum: Global Lighthouse Network 2025
- NIST: AI Risk Management Framework
- ISO: ISO/IEC 42001 AI management systems
- ETDA: Future AI Governance
- ETDA: Generative AI Governance Guideline
- ASEAN: Expanded ASEAN Guide on AI Governance and Ethics
