When a factory makes many products in small lots, changeover improvement often means shortening the time a machine is stopped. Yet a faster tool exchange does not solve a different question: which approved settings apply to this work order, on this machine, with this tool and material? This guide is for production engineering, quality, operations and IT/OT teams at factories in Thailand and ASEAN evaluating AI recommendations for changeover settings. It covers a decision-support workflow: the system proposes an approved recipe, an authorized person accepts or rejects it, a first-piece inspection gates production, and the full decision remains traceable. It does not assume that AI writes settings to a PLC without supervision.
“Settings” depend on the process. They may include approved temperature, pressure and cycle limits for molding; a tool, program and cutting parameters for machining; or fixtures, torque and test limits for assembly. Equipment capability, the quality plan and site safety procedures determine which parameters may even be proposed.
1. Separate changeover time from setting selection
SMED and time studies help separate preparation while a machine runs from work during a stop. Recipe recommendation addresses another loss: searching for an applicable approved setting when the product and operating context change. The target is fewer selection errors, missing checks and repeated first-piece adjustments, rather than a claim that AI will optimize every machine motion. Our changeover time reduction guide explains time segmentation; our injection molding parameter AI guide focuses on one process. Here the problem crosses products, machines, materials and approvals.
The same part number may run on two machines with different allowable ranges. A repaired mold may invalidate an older recipe. A new material lot may require an extra inspection. Looking up settings using only the product number conceals those differences. A useful recommendation therefore identifies an applicable candidate, its evidence and its unresolved conditions. Measure search time, version errors, first-piece results, readjustments, approval delays and recovery time separately. A shorter stop is not a success if quality risk increases.
2. Define the authoritative recipe and its applicability
Before training a model, define the record that the plant considers authoritative. At minimum, the design should distinguish recipe ID and revision; effective date; product and revision; operation; machine or machine class; mold, tool and fixture; material specification; settings and approved bounds; inspection method; approver and approval time; and withdrawal conditions. A historical setting that happened to produce an acceptable piece is not automatically an approved standard. Trials, temporary deviations and production releases need distinct states.
Do not select a recipe simply because it has the newest revision. It may belong to another machine. Preserve why an older revision was withdrawn. Store approved limits separately from values operators used in practice. Historical runs can inform ranking or anomaly detection, but cannot replace quality approval. ISA-88 is a standard series for batch process control and recipe concepts. It should not be presented as a mandatory data model for every discrete machine. Agree on the local meaning of “recipe,” “program,” “work instruction” and “inspection specification” first.
3. Join the work order to the actual shop-floor context
The next product number alone is insufficient. Link the released work-order ID, product revision, quantity, operation and assigned machine to the installed mold or fixture, tool state, material specification and lot, required tests, and maintenance or calibration status. Start with the fields that determine eligibility. Represent missing facts as unknown. Filling a missing mold ID with the most common mold can produce a dangerously confident match.
ISA-95 provides a conceptual framework for integrating enterprise planning and manufacturing operations. ISA describes Level 4 as planning and logistics, Level 3 as manufacturing operations, and Level 2 as supervisory control. This is useful for deciding which system owns an ERP order, MES execution and quality records, or equipment state; it does not prescribe an AI algorithm. For the broader ERP/MES boundary, see our production management system primer.
The OPC Foundation’s OPC UA for Machinery Job Management explains that a job order can comprise programs, recipes, metadata and parameters, in various combinations. This helps structure the work-order-to-machine interface. It does not mean every installed machine exposes that model or supports writing settings. Inspect the actual readable fields, permissions, revisions and loss-of-communications behavior. The OPC UA for Machine Tools specification may also be relevant when the equipment and supported profile match.
4. Show evidence, confidence and limits with the candidate
A screen that displays only “recommended temperature: 185°C” invites uncritical acceptance. Show recipe ID and version; match status for product, machine, tool and material; the differences from current settings; approved lower and upper bounds; the range of historical cases used; missing inputs; and the model or rule version. If confidence appears, define it precisely. Similarity to past cases is not the probability of first-piece acceptance. Do not decorate an uncalibrated score with a percentage sign.
The first stage can simply search and rank approved recipes. If interpolation is eventually allowed, contain it within bounds approved by engineering and quality, and require a separate process before a new value becomes the production standard. An unapproved mold, unknown material, overdue calibration or out-of-bounds candidate should yield “no recommendation” and escalation. Evaluate false eligibility, abstention reasons, human corrections and first-piece outcomes as well as top-choice accuracy. A shadow run, where operators cannot see the model’s proposal, can reveal disagreement without altering the current operation.
5. Fix the handoff between AI, people and the machine
Document the sequence: read the released order and actual context; filter eligible approved recipes; show a candidate and caveats; have an authorized person accept, edit or reject it with a reason; set the machine under the existing work instruction; inspect the first piece; then authorize production. In this guide, acceptance on the screen does not itself change the machine. Any future write interface is a separate project requiring vendor specifications, interlocks, safety review and change control.
Manufacturing checks order and installed items. Process engineering checks setting differences and machine fit. Quality owns the inspection and deviation route. Define the night-shift delegate, loss-of-network procedure and urgent-order exception. Record identity, time, displayed candidate, edit and rejection reason so the decision can later be reconstructed.
NIST’s AI Risk Management Framework uses Govern, Map, Measure and Manage functions. Its Core includes defining, assessing and documenting human-oversight processes. Use that as a governance reference, not as a certification of a plant installation. Put ownership, model-change approval, operator objections and incident reporting into an explicit responsibility matrix.
6. Treat first-piece inspection as a quality gate and a data source
Passing a settings screen is not proof that the machine, material and tool produce an acceptable part. Define product- and operation-specific checks such as dimensions, appearance, weight, electrical response or torque. Link measured values, gauge and calibration status, test method, inspector, time, disposition and approval to the work order. One accepted first piece does not guarantee all later units; in-process controls still apply.
If the first piece fails, avoid saving only the last value from a sequence of trial adjustments. Keep each attempt, its recipe revision, reason and measurement. Require a deviation approval before crossing authorized bounds. Distinguish a failed test from an omitted test, a changed test method or a pending retest in training labels. Otherwise the model can learn from a missing inspection as if it were success. For a pilot, compare like product, machine and material contexts, and examine whether the order mix changed. Do not manufacture a universal improvement percentage or payback period.
When feeding inspection results back into the model, preserve units, gauge resolution and the revision of the acceptance specification. A part accepted against an old drawing must not silently become a pass against a new tolerance. Distinguish a piece that passed only after rework from one that passed on its first attempt. If material, mold maintenance and inspection methods all change together, do not assign the outcome to the AI setting alone. Version the dataset, and retain the original record, correction reason and approver when labels are corrected.
7. Make rollback more than restoring the last number
After a failed first piece or a process alarm, recovery may involve the last valid approved recipe and revision, installed tool, material lot, program, inspection plan and machine state. The selected recipe in software may not match the machine if someone manually edited a value. Verify physical state before restart.
Specify who stops production, how work in process is held, how a prior valid condition is restored, how a test run and new first-piece inspection are conducted, and who authorizes restart. Test those steps in FAT/SAT. Preserve start and end times of setting application and links to order, lot or serial, and inspection. The previous recipe may no longer apply to the current material or tool; in that case stopping and escalating is the correct behavior.
An audit trail should connect input, model version, proposal, human decision, actual setting and first-piece outcome. Specify IDs, synchronized time, before/after values, corrections, retention and access. The practical test is whether a quality complaint can be traced from one work order through that complete chain.
8. Write acceptance scenarios into the RFP
An RFP saying only “optimize changeovers with AI” invites proposals for unrelated scheduling, vision and autonomous control products. Name the operation and product family, machines, recipe repository, source of orders, tool and material identification, inspection system and approval roles. Then describe what a user sees, when the system must abstain, and what releases production.
| Area | RFP question | Acceptance evidence |
|---|---|---|
| Master and versions | Can product revision, machine, tool and material constrain a recipe? | Logs excluding obsolete or wrong-machine recipes |
| Recommendation bounds | What happens with unknown material, expired calibration or an out-of-range value? | “No recommendation” and escalation |
| Human approval | Can users delegate, edit or reject with reasons? | Role-based operation and audit export |
| First-piece gate | Can production release be blocked on failure or missing test? | Tested site scenario |
| Recovery | What if the previous recipe is no longer applicable? | Hold, restore, retest and restart records |
| Connection | How are readable fields, retries, offline work and clock alignment handled? | Interface contract and fault tests |
| Model operation | Can model version, training data and reevaluation be traced? | Change log and evaluation report |
Compare quotations on a shared scope: machine connectivity assessment, master cleanup, recipe approval redesign, historical data preparation, roles and screens, FAT/SAT, training, translation and support. At a Thai site, check that Thai operator labels and Japanese or English management views resolve to the same IDs. There is no responsible single “market price” without knowing interface count and data condition.
Set the data contract and owners before acceptance testing
As an RFP appendix, list each exchanged field with its name, meaning, type, unit, required condition, authoritative system, editor, update time and withdrawal rule. For example, ERP may own the work-order ID; a quality-controlled recipe register may own the recipe version and approval state; the shop-floor confirmation may own the actually installed mold ID; and the inspection record may own first-piece disposition. Each plant must agree on its own allocation. A field shown on the AI screen does not mean that screen may edit the master. Corrections belong in the authoritative source, which then publishes the corrected version and effective time.
The ID crosswalk must account for product suffixes, mold repair numbers, machine asset IDs, and material-lot leading zeroes. For CSV exchange, specify encoding, delimiter, units, decimal separator and time zone. Ambiguous UTC versus ICT timestamps can reverse the apparent order of approvals and lot changes around a shift boundary. A “take the newest arrival” rule can let delayed old data overwrite the actual state. Give messages unique IDs and versions; make retries idempotent; hold automatic acceptance when events arrive out of order.
Manage concrete test records in one table shared by the plant and vendor. The normal case has an approved product, machine, mold and material match. Boundary cases should include: an old product revision despite a matching product number; the same machine model but an excluded asset ID; a repaired mold read under its old ID; a known material lot with unknown specification revision; calibration expiry at the date boundary; a recipe withdrawn that day; an approver whose role has just changed; a duplicate work-order message; stale installation data arriving after connection recovery; and a first-piece retest that might overwrite the earlier failure. For each case, specify not only “show candidate,” “abstain,” “hold approval” or “block production,” but also whom to alert and what log must remain.
Rollback acceptance must prove more than a button. Confirm that the prior recipe version applies to the current material; identify affected work in process; verify actual settings through machine readback or physical check; record a new passing first piece and the restart approval. If no valid previous recipe exists, the system must keep production stopped and notify process engineering and quality rather than invent a replacement value. Requiring the complete evidence chain to be retrievable from the work-order ID lets buyers compare operating capability instead of demo-screen appearance.
9. Use FAT and SAT to test failures, not just a happy path
FAT can use a vendor test environment and simulated data; SAT needs the real machine, network, users and inspection route. Include same product on another machine, newly revised mold, new material lot, unapproved recipe, obsolete version, connection loss, duplicate message, clock drift, failed first piece and missing night-shift approver. Agree on expected results and evidence before the test.
Test whether the model abstains when it should. After model changes, replay representative cases for increased false eligibility. Even without AI-to-PLC writing, compare the screen’s selected condition with the condition read from the equipment or verified at the machine. Define SAT closure in terms of unresolved defects, roles, auditability, first-piece gate, recovery and trained operators. There is no universal industry pass percentage for “AI accuracy”; set criteria for the site’s quality risk and test sample.
10. Scope a pilot and its KPIs
Choose one operation and machine group with frequent changes, some recipe discipline and traceable first-piece results. Capture the current approval process. Run recommendations in shadow mode. Fix false eligibility before exposing suggestions to approvers. Retain the human decision and first-piece gate throughout.
Measure candidate coverage, false eligible suggestions, recipe search time, first-piece acceptance and readjustment, scrap, approval wait, recovery time and record completeness. Note changes in product mix and operating conditions. A pilot is evidence for an investment decision, not a commitment to production deployment. End it with lists of conditions suitable for recommendation, conditions reserved for human judgment, conditions requiring abstention, and missing data to collect.
11. Common implementation mistakes
Training on every historical setting mixes trials, temporary deviations, mistakes and pre-maintenance conditions. Label approval state before training. Selecting by product number alone misses machine, mold, material and revision. Showing only a confidence number conceals missing data. Claiming success without first-piece outcomes ignores quality. Skipping recovery tests leaves the most important action unproven. Assuming an interface from a brochure overlooks differences between old machines and firmware versions. Data ownership, approval, testing and recovery must precede a model-performance claim.
12. Establish operation at a Thai plant
A headquarters-approved revision may coexist with an old paper instruction, a PC copy and a program stored in the machine. Identify the authoritative copy and inventory differences for the pilot area before attempting full migration. If the machine cannot expose settings automatically, begin with documented manual verification while planning which fields to structure later. A photograph alone does not support reliable numeric comparison.
Translation is also data governance. A Japanese mold term, Thai shop-floor name and English ERP field must refer to the same object. Give products, machines, tools, materials, recipes and test characteristics stable IDs, and display local names beside them. Use an approved glossary for warnings and instructions. If headquarters maintains product revisions while the site maintains installed-tool status, updates may arrive at different times. Hold recommendations when the combined state is inconsistent; retain both event time and effective time.
Train staff to reject or escalate a candidate, record reasons, hold material after a failed first piece, and switch to the approved offline procedure. Train managers on model updates and reversal. Collect early reports of wrong labels and ambiguous terms, and distinguish a quality defect from a usability improvement. This order allows value from authoritative recipes, order matching and audit history even before a predictive model has adequate data.
13. FAQ
Can we start when recipes exist only in spreadsheets?
Possibly. First add revision, applicable machine, tool and material, approval and withdrawal state to each record and resolve contradictions. Even a non-AI search with approval history is useful.
Should AI write settings straight to the PLC?
The workflow here does not. It recommends; a person approves; settings follow the existing work instruction. A write interface requires separate design, safety review, change control and acceptance tests.
Does this apply beyond injection molding?
It can apply to machining, assembly, testing and packaging where approved tools, programs and conditions change by order. Parameters and quality gates differ by operation; one model may not transfer across them.
What if historical data are sparse?
Begin with rule-based eligibility and evidence. Abstain for underrepresented conditions and reevaluate when data accumulate.
Who should own the system?
One possible split is process engineering for applicability, quality for inspection and deviations, operations for the physical process, and IT/OT for data and interfaces. Document final authority and exceptions for the actual organization.
What does it cost and when does it pay back?
There is no credible universal figure. Machine interfaces, master quality, approval complexity and inspection integration drive costs. Compare proposals against the same RFP and calculate payback from measured site data.
Conclusion
AI changeover recipe recommendations are a controlled decision process, not a number generator. Establish approved recipes and revisions, match order to physical context, keep an authorized human decision, gate production on first-piece inspection and rehearse rollback. In the RFP, FAT and SAT, demand evidence for wrong versions, missing facts, failed inspections, connectivity loss and recovery as well as normal cases.
TOMAS TECH can help Thai factories map recipes, work orders, first-piece checks and approvals across production management, MES and machine data. Share the target operation, equipment, changeover frequency and current recipe repository through our contact form to identify a practical pilot scope and the data that should be fixed first.
