If you are considering the “introduction of AI risk assessment for non-routine work” in a Thai factory, the first thing to decide is not how to make AI score hazards. It is who plans work such as equipment shutdowns, jam clearing, die changes, temporary repairs, and trial runs; what information they use; what must be checked on site; and who authorizes the work. AI can draft JSA (Job Safety Analysis) and permit-to-work documents from past procedures and equipment information, and it can point out inconsistencies or omissions in what is written. However, hazard identification, verification of isolation, matching against actual site conditions, and approval to start or restart remain the responsibility of competent, authorized people. This article provides an end-to-end framework for personnel in manufacturing, maintenance, safety, IT, and procurement at Thai sites, covering implementation scope, operations, RFPs, and FAT/SAT.
The Scope of AI Support for Risk Assessment of Non-Routine Work
Non-routine work falls outside daily standard operations, and the equipment, location, personnel, and time of day may vary from case to case. For example, when a workpiece gets jammed in a conveyor on a production line, simply knowing that it was stopped from the normal operator screen is not enough to fully assess pneumatic energy, parts that may fall under gravity, transfer from adjacent equipment, residual charge, and other hazards. Even if the equipment model is the same, conditions differ after a jig change, during removal of protective covers, or when night-shift support personnel are involved. OSHA guidance on hazard identification also indicates that low-frequency tasks such as maintenance and startup/shutdown have their own specific hazards. However, this is U.S. guidance and does not directly demonstrate compliance with Thai law.
What AI support should do is “fill in the blank space before people start thinking.” By cross-checking work requests, asset registers, manufacturer procedures, past JSAs, and change histories, it can propose candidate work steps, energy sources to verify, related departments, interference with simultaneous work, and required evidence. Each candidate should include its source, applicable conditions, revision, and points where confidence is low. Workers and supervisors then review the actual equipment and revise the draft, while safety personnel and permit authorizers decide whether the controls are actually workable. Do not establish an operating rule that says, “AI judged it safe, so approve it.”
In this article, JSA means a method of breaking one job into steps and examining the hazards and controls for each step. A permit-to-work is a control procedure that specifies the scope, period, equipment condition, isolation, related parties, start conditions, and completion/handover. HSE explicitly states that simply issuing a permit does not make the work safe. Even if digitized, a permit screen is meaningless unless it matches the physical conditions in the field. Introducing AI for non-routine work should begin not with document automation, but with a design that makes this alignment easier to verify.

Narrow the Target Work: Where to Start With Maintenance, Changeovers, or Trial Runs
For an initial rollout, it is practical to narrow the scope to one type of work, one equipment group, and one permit workflow. Candidates include clearing conveyor jams, changing dies on packaging machines, and replacing pump seals. The selection criteria are not just “high frequency.” Confirm that standard procedures exist, that differences in job conditions arise in the field, that past permit records are available for reference, and that there are site leaders who will participate in the assessment. If you immediately expand to high-hazard confined space work or complex simultaneous jobs, it becomes difficult to separate assessment failures from workflow design failures.
Once the work type is selected, observe the full process from initiation through equipment handover. Even if the requester writes “jam clearing,” the job on site may actually break into multiple stages such as removing covers, stopping upstream feed, isolating the air supply, releasing stored pressure, entering the work area, and conducting a trial run after restoration. Record who performs each stage and where the necessary information is scattered across paper, Excel, maintenance systems, and verbal communication from the team leader. If you define AI input fields first, you risk ending up with fields the site cannot actually fill in, creating a danger that blank items will be filled with plausible-sounding guesses.
The PoC scope sheet should state the target equipment, work categories, responsible shifts, whether contractors are involved, input documents, output items, approvers, excluded work, and stop conditions. For example, you may include “die changes during planned stoppages on Packaging Line A,” while explicitly excluding emergency restoration immediately after breakdowns, live electrical work inside control panels, and piping line opening. If excluded work is detected, the AI must not reuse a similar permit; instead, it should display “out of scope – follow current procedure.” Testing safe-side behavior for out-of-scope work from the beginning is what builds trust on the shop floor.
From Work Request to Permit: Seven Gates Where Humans Retain Responsibility
The first gate is the work request. Confirm the asset ID, location, reason for the request, intended work, and requested time, and do not allow vague entries such as “general repair” to pass through unchanged. The second gate is document revision control. Confirm whether the equipment manufacturer’s procedure, electrical and pneumatic diagrams, post-modification configuration, and isolation point list match the actual equipment. If only outdated drawings are available, the AI should not fill in isolation points by guesswork; it should wait for on-site confirmation.
The third gate is the JSA draft. Break the work into steps and line up the hazards, affected persons, existing controls, additional controls, verification methods, and responsible persons. The ILO workplace risk assessment guide presents the sequence of identifying hazards and persons at risk, defining controls, recording responsibilities and deadlines, and reviewing them. AI can support this sequence, but it cannot guarantee discovery of new hazards that are not visible on site.
The fourth gate is on-site verification. The responsible person and the workers confirm the equipment number, access route, adjacent work, materials, lighting, scaffolding, and energy sources, and they record any differences. The fifth gate is implementation and verification of controls. Isolation of electrical power, pneumatic pressure, hydraulic pressure, gravity, and other energy, along with release of residual energy and methods to prevent startup or restart, must be carried out according to the procedure defined for the target equipment. OSHA lockout/tagout materials explain hazardous energy control during maintenance and servicing, but those U.S. regulations must not be reinterpreted as legal obligations for Thai factories. Here, they are used as practical design references.
The sixth gate is the permit. The permit authorizer reviews the JSA and site conditions and decides whether the necessary controls and evidence are in place. AI may warn about unfilled fields, mismatched equipment IDs, expired procedures, and overlapping simultaneous work, but it must not finalize electronic signatures or permit status on its own. The seventh gate is completion and handover. After the job, record retrieval of tools and parts, restoration of covers, withdrawal of related personnel, trial-run conditions, handover of the equipment, and outstanding items for the next shift. HSE also emphasizes coordination for work spanning multiple permits or shifts. An approval gate is needed not only at the start, but also at completion.
| Gate | Assistance delegated to AI | What humans confirm and decide | Evidence to retain |
|---|---|---|---|
| Request | Indicates missing mandatory fields and equipment ID | Defines the target work | Request number, equipment ID |
| Documents | Suggests candidate revisions and applicable equipment | Checks documents against the actual equipment | Revision, reference timestamp |
| JSA | Drafts steps, hazards, and controls | Decides hazards and priority controls | Revision history, responsible persons |
| Site | Presents verification items | Confirms conditions and simultaneous work | Site check record |
| Controls | Indicates missing evidence of implementation | Implements and verifies isolation and controls | Isolation record, verifier |
| Permit | Displays inconsistencies and missing approvals | Authorized person decides whether work may start | Permit issuer, time |
| Handover | Summarizes incomplete items | Confirms restoration, trial run, and handover | Completion record, shift handover |
What a JSA Draft Should Read – and What It Should Not Be Allowed to Read
The minimum set of materials to provide to AI includes the equipment asset register, approved work procedures, the applicable JSA form, permit conditions, revised drawings, and records of past corrective actions. If documents are stored in different locations, materials for old and new machine versions may become mixed even under the same equipment name. Search keys should therefore be equipment asset ID, installation location, model, modification revision, and procedure revision, not common nicknames. Text without a cited source should not be treated as evidence; it should be shown as pending confirmation. It is also important not to end with only a fixed PDF draft. Inputs must be updateable whenever the assumptions behind the work change.
Items entered from the shop floor should include not only “what will be done,” but also “what will not be done.” Clearly stating wiring outside the scope, covers that will not be removed, and equipment that will not be shut down at the same time reduces unnecessary guessing. However, whether a part entered as “not touched” is truly unrelated must still be confirmed by people on site. Even if AI detects contradictions in the text, it cannot grasp temporary hoses or makeshift wiring that are not shown in drawings. Photos and notes added by workers should be linked to the date/time taken, equipment, location, and before/after work status to prevent reuse for the wrong case.
Personal names, face photos, health information, and contractor contract documents should only be handled after defining the purpose of use and access rights. If images or free-text entries are not necessary, use only equipment IDs and work conditions. If documents are sent to an external AI service, confirm through the contract and settings the conditions under which inputs are stored or reused, where they are stored, how they are deleted, who can view logs, and who the subcontractors are. Specific legal compliance regarding personal data protection and occupational safety in Thailand should be judged by the relevant local department or specialist. To avoid later mixing safety records with employee evaluation, separate purposes and permissions as well.
Priorities When Listing Hazards and Controls
Simply making a JSA look tidy does not reduce risk. For each hazard, first consider whether the work itself can be eliminated, whether it can be changed to another method, or whether contact can be prevented through physical guarding, isolation, or jigs. After that, consider procedures, training, supervision, and PPE. Controls should not end with “be careful”; they should state who will implement them and how they will be verified. Even if AI frequently finds PPE in past documents, that does not mean it is the most effective control. The frequency with which wording appears in existing documents must be evaluated separately from its effectiveness for the current risk.
For example, in conveyor jam clearing, “stop before reaching in” may be insufficient. You must confirm whether new workpieces can be fed from upstream, whether conveyed material could fall, whether residual energy remains after emergency stop, whether someone else can release the stop, and from which position on the machine access occurs. AI can generate questions from equipment drawings and past near-miss reports. However, labels identifying the actual isolation points, application of locks, and verification of zero energy cannot be replaced by the output of a language model. Site records should capture not only checkboxes but also the observed condition and who confirmed it.
Abnormal conditions should be written into the JSA from the start. If the work content changes midway, another contractor enters the same equipment, an alarm is issued, the planned isolation cannot be achieved, or the shift changes, the permit should be paused and reassessed. A design in which the draft can be updated while the status remains “approved” is dangerous. If the revision changes, the permit status should be invalidated, related persons notified, and work restarted only after the necessary reconfirmation. AI can help summarize changes, but the authority to accept a change belongs to people.

Required Authority, Status Control, and Audit Logs for AI-Supported Permit-to-Work
Permit-to-work status should be divided into stages such as “requested,” “draft,” “awaiting site confirmation,” “awaiting control implementation,” “awaiting approval,” “active,” “suspended,” and “completed.” Define these to fit the actual factory workflow, and make the on-screen status transitions correspond one-to-one with actions in the field. Operations allowed to AI should be limited to document retrieval, drafting, candidate suggestions, inconsistency notifications, and summaries of change differences. It must not be allowed to transition a permit to “active,” substitute for isolation verification, sign on behalf of people, or send startup commands to equipment. The prohibition on auto-approval must be enforced not only in prompt wording, but also through permissions and API implementation.
At minimum, roles should be divided into requester, work owner, worker, equipment-owning department, permit authorizer, safety officer, and system administrator. Even where one person holds multiple roles on site, the system should make clear what that person did in each case. Contractors should access only the information necessary for their own jobs and should not be able to view records for other companies or other lines. During shift change, the incoming responsible person must receive outstanding items and equipment status; the permit authority must not simply carry over via login handover.
Audit logs should retain input values, referenced document IDs and revisions, the revision of the AI model or rules, generation time, the draft itself, before/after edits, approvers, reasons for rejection, and reasons for suspension/restart. Logs are not for proving that the model was correct; they are for checking what information was used, who made which decision, and what can later be improved. If there is any chance the site will be offline, define how to operate when the permit screen cannot be displayed, how records will be synchronized, and how to prevent duplicate approvals or resurrection of stale status. Avoid designs in which AI infers that work is permitted during communication failure.
NIST’s AI Risk Management Framework can serve as a reference for organizations designing roles, evaluation, and monitoring for AI use. Materials dealing with human-AI interaction state the need to clarify who decides and who supervises. This is a reference framework for AI product governance; it does not guarantee certification of equipment safety functions or compliance with Thai law. The design and verification of safety circuits, interlocks, emergency stops, and similar functions must be carried out separately by engineers familiar with the relevant equipment and applicable standards.
Multilingual Operation at Thai Sites: Do Not Confuse Translation With Approval
At Thai factories, procedures from Japanese headquarters, English manufacturer documents, and Thai on-site explanations may all coexist. If AI generates multilingual drafts, verify for each target language that equipment names, isolation points, operation prohibitions, severity of warnings, work sequence, and numbers/units are consistent. In particular, “stop,” “isolate,” “measure for absence of voltage,” and “prevent restart” are different actions. Even if a translation sounds smooth, reducing these into one term destroys the meaning of the controls. Establish a mechanism that refers to approved terminology lists in each language and to the actual signs and labels used on site.
Screens that workers sign should be displayed in a language they understand, and where necessary, explanation and Q&A should be conducted in their native language. AI translation can help prepare explanations, but it is not evidence that understanding has been confirmed. If site instructions differ from the permit, work should not begin until an authorized person resolves which is correct. Changes to the terminology list should carry revision control and approval so that the meaning of past permits does not change later.
Legal requirements in Thailand, explanations to workers, allocation of contractor responsibilities, and handling of personal information vary depending on the factory location, industry, and contracts. Overseas OSHA, HSE, ILO, and NIST materials are design references and are not a substitute for compliance confirmation in Thailand. In the RFP, include tasks such as “designate the person responsible for confirming local laws and internal rules” and “determine the standards applicable to the target equipment and work.” Rather than having an AI vendor broadly guarantee legality, it is more practical to clarify the necessary evidence and boundaries of responsibility in the contract.
What to Write in the RFP: Inputs, Outputs, Exceptions, and Boundaries of Responsibility
At the beginning of the RFP, state: “Support the drafting of JSAs and permits for target work, and detection of missing items; permit decisions are made by designated personnel.” Also explicitly state exclusions such as automatic approval, equipment control, and automatic legal compliance determination. The input specification should list equipment ID, work category, document revision, site conditions, related parties, simultaneous work, isolation points, and attachments. The output specification should distinguish work steps, hazards, candidate controls, supporting documents, applicable conditions, unconfirmed items, differences, and permit status. If the deliverable is only a single “risk score,” it cannot explain important omissions.
The exception specification should separately define cases such as document not found, equipment ID mismatch, outdated procedure revision, contradictory inputs, out-of-scope work, communication failure, contractor change, unverified isolation, conditions changing during work, and an increase in simultaneous work. For each exception, decide who is notified, whether the draft stops, whether the permit is invalidated, and whether the process returns to a manual procedure. Do not decide permit approval solely from the confidence score of AI output. It may show high confidence even when the input is incomplete.
The boundary of responsibility also affects post-procurement operating costs. The factory retains ownership of the master versions of equipment and work procedures, hazard assessment in the field, and approval authority. The vendor handles, within the contracted scope, revision control of searchable documents, system permissions, logs, testing, change management, and incident response. Create a table for the issues that remain between the two parties: who updates documents, who closes out incorrect AI suggestions, and who performs retesting after model updates. Include requirements for fallback to paper or the existing system during service outage, as well as the format for data export.
| RFP item | Minimum description example | Evidence reviewed at acceptance |
|---|---|---|
| Scope boundary | Equipment group, work category, exclusions | Stop screen for out-of-scope cases |
| Documents | ID, revision, applicable model, update responsibility | Source links and revisions |
| Authority | AI drafts only; humans issue permits | Permission test records |
| Exceptions | Outdated revision, inconsistency, communication failure, condition change | Negative test results |
| Audit | History from input through approval | Reproducible logs |
| Multilingual | Approved vocabulary and site confirmation | Review records for each language |
| Changes | Model, prompt, and document updates | Retest and rollback records |
FAT and SAT: Prioritize Tests That Stop Dangerous Passage, Not Just “Correct Drafts”
In FAT (Factory Acceptance Testing), use fixed document revisions and work scenarios to reproduce search, drafting, permissions, logs, and exception handling in the vendor environment. In SAT (Site Acceptance Testing), test using the local equipment IDs, network, terminals, languages, existing maintenance and permit workflows, and actual approvers. PoC is the stage for checking applicability, FAT is functional confirmation against the contract specification, and SAT is confirmation that operation is workable on site. None of these are tests in which AI guarantees safety functions.
As a positive case, prepare planned maintenance where the approved procedure matches the site conditions. That alone is not enough. For negative cases, include outdated drawings, isolation points from another line, a contractor added during the work, missing lock records, unverified residual pressure, handover to the night shift, simultaneous work, offline terminals, and out-of-scope live electrical work. The expected result is not just “output the right text.” It is to hold the draft, indicate insufficient basis, reject actions by persons without permit authority, and stop the existing permit. Failed cases should be retested after correction.
Do not compress evaluation metrics into one overall accuracy rate. Count separately missing hazards, incorrect candidate controls, unsupported wording, equipment mix-ups, mistaken acceptance of out-of-scope work, bypasses of approval gates, mistranslations of display language, and missing required logs. For scenarios involving severe hazards, do not let them disappear into averages; set pass/fail criteria for each individually. Acceptance criteria should be defined by the factory’s own risk assessment and the contract, not by applying the model provider’s general benchmark values as-is. After acceptance, repeat the same tests when equipment is modified, procedures change, or the model is updated.
Link Test Scenarios as One Continuous Job
Using a planned shutdown and die change on a packaging machine as an example, the requester enters the machine number and planned stoppage, and the AI indicates the applicable approved procedure and drawings. The site leader checks the actual machine model and modification revision, then confirms transfer equipment, heaters, pneumatic sources, moving parts, tool storage locations, and so on for each work step. If, at this point, the drawing given to AI is replaced with an outdated revision, the expected result is not “continue using a past JSA,” but “show revision mismatch and stop the draft.” The test sheet should specify the input, expected warning, roles allowed to operate, logs that remain, and conditions for recovery.
Next, add a wiring task on the adjacent line during on-site confirmation. If the work area or control panel is shared, the assumptions behind the original JSA have changed. AI may indicate the possibility of simultaneous work, but whether separation distances and schedule adjustments are workable must be confirmed between the responsible persons on site. SAT should verify that this change in conditions does not leave the existing approval active. Further, if a shift change occurs after permit approval, test whether the incoming shift can confirm and take over the remaining work, isolation status, tools, and trial-run conditions. Pass/fail should include not only whether the AI summary is readable, but also whether the user can trace back to the original records.
Negative testing cannot be verified by the abstract goal of “do not output wrong text.” For example, mix in drawings whose machine numbers differ only at the end, enter the names of isolation points using old naming, change pressure units, operate worker and permit-authorizer roles under the same account, or have a contractor attach photos from another case. Change each failure trigger one by one. In each case, observe what the system rejects, who is notified of what, and what evidence is retained. If the issue is not detected, do not settle for model tuning alone; investigate whether the correction should be made in input constraints, equipment ID matching, permission design, or manual gates.

Do Not Judge Cost and Effect Only by “Document Creation Time”
Costs include document inventory and revision control, equipment ID preparation, workflow design, terminals, integration with existing CMMS and permit systems, multilingual reviews, FAT/SAT, training, operational audits, model updates, and manual operation during failures. If you compare only the usage fee of generative AI, you will overlook the preparation and maintenance costs required on site. In the RFP, separate initial cost, monthly cost, usage-based cost, additional equipment, modifications, and data return upon exit, and ask vendors to show the costs that increase when deploying across multiple sites.
Effects should also not be limited to simple reduction of document creation time. Track the number of draft revisions, differences found during on-site verification, references to outdated procedures, permit rejections, delays in job start, missed handovers, and the implementation rate of reassessment after changes. However, do not jump to the conclusion that an increase in detected issues means more incidents. It may simply mean that problems previously unseen are now being recorded. Conversely, do not treat a reduction in warnings as proof of improved safety; evaluate it together with field observations, near-miss data, and implementation of controls. Also avoid a design that tries to causally prove accident reduction from a short-term PoC alone.
Before implementation, measure the current state of the target work over at least multiple shifts, and compare AI-assisted and non-assisted work under the same conditions. For example, record the time from JSA preparation to approval, equipment mix-ups, when unconfirmed items are discovered, and the confirmation workload on supervisors. Do not compare cases with different hazard levels by simple averaging; separate them by equipment and work category. Even if efficiency improves, it is a failure if important confirmations are omitted. In procurement decisions, review both the time to work start and whether safety-related stops and rejections functioned appropriately.
When assessing the burden on the site, count any increase in input work as well. If existing permits are simply photographed and fed to AI, it may become unclear who is responsible afterward for correcting equipment IDs and document revisions. If the extra work needed to improve input quality falls mainly on maintenance personnel, total time may not decrease even if draft generation is fast. Conversely, once the equipment register and procedures are well maintained, searches and handovers in existing non-AI operations may also improve. Do not attribute effects solely to AI; report separately the changes brought by document maintenance, authority clarification, and shop-floor training.
If there are too many false alerts, the shop floor will ignore them; if there are too few, omissions will be missed. Rather than uniformly reducing alerts, distinguish on screen between high-risk stop conditions, waiting for document confirmation, and simple input assistance. Show permit authorizers the “differences that require immediate judgment,” and show workers the “points that must be checked in the field.” In operational review meetings, look not only at the number of alerts, but also the reasons for rejections, the number of cases actually corrected, and concrete examples of notifications that were not useful. Also confirm through observation and interviews whether people are becoming overly dependent on AI output.
Phased Introduction and Change Management: An Example of Dividing 90 Days Into Three Parts
Use the first period for work observation and document organization. Cross-check equipment IDs, drawings, permit forms, actual isolation procedures, and past corrective actions, and confirm the Thai-language expressions used on site. In the next period, run AI drafting in parallel with the existing procedure for one work category. The AI draft should not yet become the formal permit; the site leader records differences and errors. In the final period, carry out the defined FAT/SAT and move into formal operation with a limited scope. The “90 days” here is only a planning example; it is not a legal deadline and does not guarantee safety.
After operations begin, put document updates and model updates under the same change-management process. If there are manufacturer procedure revisions, jig changes, line relocation, additional isolation points, or contractor changes, retest the relevant scenarios. Review output differences before and after updates to confirm that hazards or controls have not disappeared. If a serious error appears, stop AI support and revert to the existing human permit procedure. When restoring service, record the cause, scope of impact, correction, retest, and approval. Do not make a system the only permit path on site if stopping and rollback have never been tested.
Do not let the safety department carry the operation alone. Maintenance, manufacturing, contractors, IT, and procurement should periodically review incorrect suggestions and changes in the field. Provide a reporting channel where workers can say, “the AI suggestion is wrong,” and ensure reports are handled without disadvantage. The idea of continuous measurement and management shown in the NIST AI RMF can be used in designing such reviews. The party being evaluated is not the model, but the factory’s operating team.
Five Failure Modes to Identify Before Placing an Order
The first is searching documents using only the equipment’s common name. Even if both are called “Packaging Machine No. 2,” the equipment register number, control panel modifications, or protective cover specifications may differ. If a procedure from a similar but different line is found, the resulting draft may read naturally while specifying different isolation points. Before placing an order, decide how equipment IDs will be cross-checked against the physical label, the register, drawings, maintenance history, and permits. At a minimum, do not search critical items using only freely entered names; mechanically verify consistency between the selected asset ID and revision. If the drawing is unclear, do not let users choose a “candidate latest revision”; instead, hold it pending until it is confirmed in the field.
The second is training it on past permits as if they were an answer key. Past documents contain temporary measures used only at that time, omissions, and procedures that have already been discontinued. Even if AI recommends “frequently used controls” based on wording frequency, those controls are not necessarily effective for the current equipment condition. First, make approved procedures and current drawings the higher-priority reference materials, and limit past JSAs to supporting material for finding “what was checked for similar work.” If a serious deficiency is found in a past case, do not make only that single case a prohibited phrase for the model; instead, correct the controlled copy of the procedure and field training.
The third is substituting attached photographs for digitized field verification. If photos taken before isolation, photos of different equipment, and photos taken after the work are mixed into the same case, judgment worsens even though the amount of evidence increases. When using photos, link them to the equipment ID, location, photographer, time, work stage, and the condition to be verified. Even then, the range shown by an image is limited. Do not assume that absence of voltage inside a panel, residual pressure in piping, or interlocking with adjacent equipment can be proven by photographs alone; keep actual measurements and prescribed verification procedures separately. It is also necessary not to mistake AI-generated image descriptions for verification records.
The fourth is approvers treating a green display on the screen as proof of safety. A screen may show input consistency, but it does not necessarily know automatically that conditions in the field have changed. On the permit screen, clearly distinguish “verified,” “not verified,” “not applicable,” and “re-verification required after change,” and avoid making the completeness of a draft look like a safety certification. Make it possible for approvers to move in one action to the source documents and field items they must verify, and disable the button if there is any unverified isolation or out-of-scope work. It is not enough merely to create a button that an authorized person can press; the approval must be tied to the document revision and field conditions at that point in time.
The fifth is updating documents and the model separately after go-live. Differences arise such as procedure revisions not being reflected in the search index, a model update changing the way hazards are expressed, or only the Thai glossary remaining outdated. Updates must include the reason for change, the affected equipment and work, test records, approver, and rollback destination. When incomplete permits created under an old revision are loaded into a new screen, do not treat them as automatically valid; have a person review the differences. Do not leave update work to the vendor alone; the factory’s document owner should manage the controlled copy. If you periodically test whether one permit can be traced all the way through in the field, you can discover breaks that do not appear from a normal screen display alone.
These issues arise not only from software defects, but at the boundaries between the equipment register, document control, authority, and field verification. Therefore, during selection, in addition to the quality of the text shown in demonstrations, give the vendor actual outdated drawings or contradictory inputs and observe where the system stops. Even if it can produce a polished JSA in a short time, a product that hides uncertain source material behind it is not suitable for this use. On the other hand, behavior that explicitly shows insufficient basis and returns the matter to a person is worth implementing, even if it slows the work slightly.
Pre-Implementation Checks Shared by Site and Procurement
In implementation review meetings, maintenance should be able to explain the equipment and energy sources, production should be able to explain the conditions for shutdown and restart, safety personnel should be able to explain hazard assessment and training, IT should be able to explain permissions and logs, and procurement should be able to explain the contract and exit conditions. Any item for which someone answers, “the AI vendor will decide,” indicates that the division of responsibility has not been finalized. In particular, deputy coverage for permit authorities, participation of contractors, handover on night shift, and switching to emergency repairs should be checked using actual field scenarios. Rather than making AI promise complete coverage of hazards, define an operating method that can stop, investigate, and correct when an omission is suspected.
For documentation, list the storage location, updater, revision, and applicable equipment for approved procedures. Also specify the difference between a permit-to-work and a JSA, who can revise each at what point, and the conditions under which signatures must be redone after a revision. For the system, confirm what data the AI can use, where it is sent, retention periods, access logs, and operation during failures. For testing, include normal, abnormal, out-of-scope, communications loss, multilingual, and shift-change cases. Putting these into a single requirements sheet makes visible the site-side preparation needed before expanding AI functions.
For the contract, confirm ownership of documents and logs, notice of changes to models or services, deadlines for corrections, re-testing, and return of data. In addition to charges when equipment or work is added after implementation, also decide who will perform the hazard assessment before new equipment is brought into scope. When copying to another site, equipment numbers, language, contractors, and permit routes differ, so simply duplicating settings does not constitute completion of acceptance. Plan for physical verification and SAT at each factory. Only after these conditions are met can reduced draft preparation time be evaluated as a sustainable benefit.
FAQ: Common Questions About Introducing AI Risk Assessment for Non-Routine Work
Can AI alone complete risk assessment for non-routine work?
No. AI can propose JSA drafts and check items from past materials, but the actual state of isolation, temporary items, adjacent work, and worker understanding must be confirmed in the field. Responsibility for issuing permits is not transferred to AI either. If materials are insufficient, the system should hold the draft and return to human investigation.
Is AI support for permit-to-work the same as automating electronic signatures?
No. An electronic signature is a means of recording who approved something and when. AI support helps with document retrieval, JSA drafting, inconsistency detection, and organizing change differences. Who has authority to sign, how implementation of controls is verified, and how a permit is stopped when conditions change must be designed separately.
How is this different from AI camera systems for factory safety management?
Cameras can be used to observe actions and conditions occurring on site. JSA support for non-routine work deals with planning and permit approval before work starts, changes in conditions during work, and handover at completion. The two can complement each other, but general AI camera systems should not be used in place of equipment safety circuits or permit decisions. For differences in operation, see also Thinking About AI for Factory Safety Management.
For the RFP, FAT, and SAT of risk assessment, what should be decided first?
First decide the target equipment and work categories, exclusions, permit authority, master copies of input documents, and stop conditions. In FAT/SAT, test not only normal cases but also negative cases such as outdated materials, lack of isolation, equipment mix-ups, communication failure, and work changes. When using past site examples as materials, AI Analysis of Near-Miss Reports may also be helpful.
Can introducing an AI product prove compliance with Thai law?
No. The mere fact that an AI product has been introduced does not prove compliance. The Thai laws, internal rules, equipment requirements, and contracts applicable to the target work must be confirmed by local responsible personnel and specialists, and implementation records must be retained. The OSHA and HSE materials in this article are overseas practical references and are not legal advice for Thailand.
Summary: The Value of AI Is Making Gaps Visible Before Permit Approval
When introducing AI risk assessment for non-routine work in Thai factories, first limit the target work and equipment, and observe the site’s actual JSA and permit workflow. Next, define the master documents, roles, isolation verification, exceptions, and approval states, and allow AI only to draft and indicate gaps and inconsistencies. Make the boundaries and evidence explicit in the RFP, and use FAT/SAT to test whether dangerous passage can be stopped. Even after implementation, if procedures, equipment, or models change, reassess them, and if necessary stop AI and return to the human procedure. More important than producing documents quickly is making visible what must be checked before approval, so that responsible personnel can make informed decisions.
Even if you are still at the stage of organizing existing JSAs or permits-to-work, please contact TOMAS TECH. We can review the target equipment, current approval workflow, and document locations together, and consider how to define the scope for AI support and the acceptance criteria.
Reference Information
- OSHA, Safety Management: Hazard Identification and Assessment — Hazard identification including non-routine and low-frequency work.
- OSHA, Job Hazard Analysis (OSHA 3071) — A method for examining work steps, hazards, and controls.
- OSHA, Control of Hazardous Energy — U.S. materials on hazardous energy control during maintenance.
- ILO, A 5 step guide on workplace risk assessments — Hazards, controls, responsibilities, and review.
- HSE, Permit to work systems — Concepts for permits, roles, handover, and simultaneous work.
- NIST, AI Risk Management Framework — AI governance, evaluation, and operational management.
- NIST, AI RMF Appendix C: Human-AI Interaction — Responsibilities for human and AI judgment and oversight.