FMEA AI Implementation: RFP and Acceptance Tests for Thai Plants

FMEA AI implementation is now on the agenda at many Japanese automotive and electronic parts plants in Thailand. QA managers and production engineering managers often tell us the same story: “New part numbers keep launching at the same time, and we cannot keep up with creating and revising PFMEAs. The veteran who could identify failure modes has retired, and a customer audit pointed out inconsistencies between our FMEA and control plan.” That is where bringing generative AI into FMEA comes up. Here is the conclusion first. What you should buy through FMEA AI implementation is not a tool that writes faster, but a tool that finds the failure modes you have missed.

Research and vendor products alike are designed on the principle that “AI proposes, people decide.” So whether implementation succeeds depends less on how clever the AI is and more on whether you have decided the following three things before you place the order.

• Where to draw the line between the steps you leave to AI and the steps people decide (in particular, only people assign Severity S)
• Preparing the data the AI will reference (foundation FMEAs, past 8D reports and customer complaints, 4M change records, control plans)
• How to measure “gaps” and “fabrications” in acceptance testing

And one more conclusion up front: labor-hour savings alone will not pay back the investment. Whether you recover it depends on how many customer complaints caused by FMEA gaps you can prevent. We show this in the model calculation later in the article.

Note that all amounts, hours and counts in this article relating to “Model Plant R” (described below) are original estimates and assumptions created for this article. They are neither industry averages nor survey results. Please read them as a calculation template and replace them with your own plant’s actual figures.

What Is Happening with FMEA and Generative AI in 2026

Generative AI arrived in major FMEA tools one after another

2026 was the year generative AI was built into dedicated FMEA tools one after another.

In its introduction of “Teamcenter Quality 2512,” published on January 12, 2026, Siemens explains that Teamcenter Copilot uses AI to reduce the effort of building the FMEA structure. According to the company, it generates “suggestions” for system elements, functions and failures, as well as for descriptions of quality actions of the prevention and detection types, and users accept or edit them as needed. Siemens says it also added the ability to jump directly from the AP, RPN and difference dashboards to the relevant location in the FMEA tree, and a feature to compare FMEAs side by side across variants.

Next, Relyence announced “Relyence 2026” on March 20, 2026. According to the company, it added “SmartSuggest,” in which AI generates functions, failure modes, effects, causes and more inside the FMEA software, using ChatGPT. It offers “full prompt transparency,” showing the entire prompt, along with a display of token usage, and users decide whether to accept, edit or discard each suggestion. Relyence says that by default only the item names in the analysis tree are included in the prompt.

Then on June 18, 2026, Siemens explained that “Teamcenter Quality 2606” extends Teamcenter Copilot’s AI assistance to all Teamcenter Quality modules. According to the company, on the control plan and inspection plan side it also generates suggestions for inspection definitions, control methods and reaction plans to speed up their creation. In other words, AI suggestions have started to enter not only the FMEA itself but also the link from the FMEA to the control plan.

Before that, in July 2025, Omnex Systems announced the addition of agentic AI to “AQuA Pro” in a published webinar. According to the company, it reuses Foundation FMEAs to generate context-appropriate PFMEAs and DFMEAs, recommends failure modes, causes, controls and actions, and verifies FMEA completeness, consistency and conformance to standards. Omnex says it supports AIAG-VDA 1st edition, AIAG 4th edition and SAE J1739, and has BOM integration and structural links to control plans and process flows.

In Japan, Things Inc. (PRISM) announced in a press release on May 18, 2026 that its CEO Atsuya Suzuki contributed an article on how generative AI is changing preventive activities, covering the front line of AI use in DRBFM and FMEA, to the June 2026 issue of “Kikai Sekkei” (Machine Design), published by Nikkan Kogyo Shimbun. The company points out that in DRBFM and FMEA practice it is not unusual for “the number of concerns raised by junior and veteran engineers to differ by a factor of 2 or more,” and cites as challenges the loss of tacit knowledge when veterans retire and the fact that internal documents are scattered across multiple systems. According to the company, PRISM is a workflow that presents concerns spanning multiple domains at the time of a design change, and turns DRBFM creation that used to take weeks into “minutes.”

What stands out here is that every product positions the AI output as a “suggestion” and leaves the decision to adopt it to people. All of the companies’ claims about effectiveness are manufacturer claims, and within the scope of research for this article we found no independent comparative evaluation of accuracy.

IATF 16949 2nd edition is planned for mid-2027 (indicative)

In July 2026, IATF Global Oversight published the revision status of the IATF 16949 2nd edition in Stakeholder Communiqué SC-2026-005. It lists 5 priority themes: simplification, clarification and efficiency (avoiding duplication with ISO 9001), software quality assurance, Tier N supply chain management, launch management, and customer-specific requirements (CSR). The schedule is described as indicative: the working draft and external feedback in 2026, final validation and translation and supporting documents in 2027, and publication planned for mid-2027.

It is worth noting that this document makes no direct reference to FMEA or AI. This is not a story of “the 2nd edition will require AI for FMEA.” On the other hand, the certification body Smithers, based on IAOB data, lists among the frequent IATF 16949 nonconformities 8.3.5.2 (manufacturing process design output: missing specifications, deficiencies in FMEA documents, etc.) and 8.5.1.1 (control plan: not reflecting all steps of the actual manufacturing process, insufficient alignment with the FMEA, etc.). It is natural to see the period before the revision as a good opportunity to firm up the alignment between your FMEA and control plan. We cover overall preparation for the 2nd edition in “Preparing for IATF 16949 2nd Edition.”

The research conclusion: human verification is a prerequisite

Research points in the same direction as the vendors. An SAE technical paper (2026-01-0106) published on April 7, 2026 evaluated a system that uses an LLM to create FMEAs from architecture diagrams combined with human-in-the-loop verification (the subjects were functional-safety system diagrams), and reported that the accuracy of S/O/D scoring was 70-90%. A paper by El Hassani et al. in the journal Design Science states explicitly that LLMs can extrapolate failure modes beyond the input data, and that the output requires verification and correction by subject matter experts (SMEs). We discuss these in more detail in a later section.

The context of Thai production and investment

Turning to Thailand, according to an FTI (Federation of Thai Industries) announcement reported by Trading Economics on September 28, 2026, vehicle production in August 2026 was 124,646 units, up 10.93% year on year. Exports, however, fell 2.04%, and the FTI projects full-year 2026 production to decline 3.33%.

On the investment side, in the first-half 2026 investment applications announced by the Thailand Board of Investment (BOI) on July 23, 2026, electrical and electronics accounted for 179 projects worth USD 3.56 billion, automotive for 122 projects, and applications from Japan numbered 123 (all on an application basis). Separately, according to a BOI explanation reported by New Straits Times in August 2026, of the 880 semiconductor and electronics-related investment applications from 2023 to mid-2026, printed circuit boards (PCBs) were the largest category at 224 projects worth USD 9.85 billion. As new plants and new part numbers increase, so does the work of creating a PFMEA for every launch and aligning it with the control plan. It is a natural development that plants want AI’s help to cope with a limited number of quality engineers.

FMEA AI Implementation: What to Leave to AI and What People Decide

The AIAG-VDA FMEA Handbook is a reference manual for automotive industry suppliers whose 1st edition was published in June 2019. It supports the creation of Design FMEA, Process FMEA and the Supplemental FMEA for Monitoring and System Response (FMEA-MSR), and is described as having been aligned with SAE J1739. According to Quality-One’s explanation, AIAG-VDA FMEA proceeds in 7 steps, and the former RPN (Risk Priority Number) has been abolished and replaced by AP (Action Priority: H/M/L).

Following these 7 steps, the parts that are easier to leave to AI and the parts people decide can be organized as follows.

AI “proposes” candidates, people “decide”

As the table shows, AI’s role is concentrated in “proposing candidates” in steps 2-4 and “cross-checking” in step 7. Generating a broad range of candidate structures, functions and failures is divergent work at which AI is comparatively good. Steps 5 and 6, on the other hand, namely risk evaluation and deciding actions, are judgments that carry responsibility, and people make them.

Only people assign Severity S

Above all, we recommend excluding Severity S even from what the AI is allowed to propose (apart from letting it propose values during acceptance testing so you can measure the deviation, keep the suggestion function turned off in production use). Severity is the evaluation of “how serious an impact the failure has on the customer or end user,” and it strongly influences whether subsequent actions are required. If the AI proposes a lower severity and a busy engineer adopts it as is, a failure mode that actually needed action could end up treated as “no action required.” We show the size of that loss in the calculation later (Key Point 3).

Think in terms of AP, not RPN

Another point to note is that the idea of having AI “calculate the RPN” belongs to the old method. Under the AIAG-VDA method, RPN has been abolished and action priority is determined by AP. When choosing a tool, check whether it supports AP under the AIAG-VDA method and how it can handle existing FMEAs made under the AIAG 4th edition method. Note that the evaluation criteria for S/O/D and the AP combination tables should be confirmed in the original handbook, and this article does not reproduce them.

How AI Features in FMEA Software Differ from General-Purpose Generative AI

When people hear “using generative AI for FMEA,” many imagine describing a process to a general-purpose generative AI such as ChatGPT or Claude and having it write an FMEA table. However, AI features built into dedicated FMEA tools and having a general-purpose generative AI write an FMEA are different things.

General-purpose generative AI is useful as a brainstorming partner. However, because it lacks version control, audit trails and control plan linkage, in practice it is hard to use as the means of creating the FMEA itself that you operate as an IATF 16949 document. Also, personal ChatGPT and Claude.ai plans have different terms of use from commercial APIs, so pasting information that includes customer drawings or process conditions into personal accounts should be avoided.

Tools are not all the same, either. Relyence’s user guide states that it uses ChatGPT or Copilot, that it consumes tokens, and that even the on-premises version requires a maintenance contract and communication with the Relyence application server. Being on-premises does not necessarily mean that no data leaves your site, so this is a point that should always be checked in the RFP.

The Data the AI References: Foundation FMEAs, 8D Reports, Complaints and 4M Change Records

What determines the quality of an FMEA support AI’s output is what you have the AI reference, more than the model’s performance. Generic failure modes can be produced even by general-purpose generative AI. What is valuable on the shop floor is being shown “the failure modes that actually occurred in the past, in this process, at this company.”

Foundation FMEAs (family FMEAs)

As Omnex describes it, “reusing Foundation FMEAs to generate context-appropriate PFMEAs,” foundation FMEAs (family FMEAs) are one of the most important inputs for the AI. If you have foundation FMEAs that organize standard structures, functions and failure modes for each process such as stamping, welding and plastic molding, new part number PFMEAs can start from there. Conversely, at plants where foundation FMEAs do not exist, or have been left outdated, because each part number has been handled by repeated copying and editing, the AI will only reproduce the old content.

Past 8D reports and customer complaints

The most valuable inputs are past 8D reports and customer complaint records. Defects that actually escaped show failure modes that the FMEA had not anticipated, or failure modes that were anticipated but whose controls did not work. If the AI can search these and propose candidates such as “this welding process has had complaints like this in the past,” that is exactly the work of preventing gaps. We explain the idea of using AI to create the 8D reports themselves in “AI for Corrective Action Reports and 8D.”

4M change records and control plans

4M changes in man, machine, material and method are the entry point for new failure modes. If the practice of reviewing the PFMEA with every change is working, change records and the FMEA revision history should correspond. Having the AI reference 4M change records can also provide clues for finding “processes where a change occurred but the FMEA was not revised.” For systems to record 4M changes, see “4M Change Management System.” If you also need to track changes on the design side, “AI for Design Change Management” is relevant as well.

Data preparation is the first big hurdle

In reality, past 8D reports are scattered across paper, Excel and email, complaint records have different formats for each customer, and Thai, Japanese and English are mixed together. Things also cites the scattering of internal documents across multiple systems as a challenge. Before introducing AI, getting at least the foundation FMEAs for the main processes and recent 8D reports and complaints into a searchable form is the first big hurdle of the implementation project. The work of documenting the knowledge in veterans’ heads also overlaps with “AI for Skill Transfer.”

The Limits of Generative AI Shown by Research: Gaps, Fabrications and S/O/D Deviation

Several studies on creating FMEAs with LLMs have been published. Lining up their results shows the range you can leave to AI and what you should measure in acceptance testing.

From these, 3 limits can be read.

The first is “fabrication.” El Hassani et al. state explicitly that LLMs can extrapolate failure modes beyond the input data. This can be a strength in terms of broadening ideas, but if failure modes with no basis in your own process creep into the FMEA, they take up the CFT’s review time and lower the document’s credibility.

The second is “gaps.” The IBM study says it was able to correctly generate over half of the key content, which conversely means a little under half was not correct. Collier et al. also assess the output as too generic and not reflecting the depth of experts’ knowledge. If you rely entirely on the AI’s candidates, failure modes specific to that plant, which do not emerge from generic knowledge, will be missed.

The third is “S/O/D deviation.” In the SAE paper, the accuracy of S/O/D scoring was 70-90%. The subjects were functional-safety system diagrams, and the number of diagrams used in the evaluation is not clear from the abstract, so this figure cannot be generalized. Even so, as an example showing that results “can deviate from expert evaluation at a certain rate,” it provides grounds for not letting AI decide S/O/D.

None of these studies measured accuracy on process FMEAs in a plant. That is exactly why it is essential to actually measure gaps and fabrications for your own processes in acceptance testing using your own past FMEAs.

FMEA Control Plan Linkage: Using AI to Check Consistency

As mentioned earlier, the frequent nonconformities cited by Smithers include 8.3.5.2 (deficiencies in FMEA documents, etc.) and 8.5.1.1 (insufficient alignment between the control plan and the FMEA, etc.). As an example of FMEA deficiencies weakening the control plan, the same article cites the situation in which “Controls specified in the plan do not adequately address the high-risk failure modes identified in the Process FMEA.”

An explanation from 16949store.com also says that the control plan takes into account the outputs of DFMEA and PFMEA, and lists as review triggers the shipment of nonconforming product, product and process changes, customer complaints, and a frequency based on risk analysis. In other words, FMEAs and control plans are not finished once created. Both need to be kept updated together with every complaint and change.

This is where tools, including AI, become useful. What we recommend is using them for mechanical consistency checks before using them to write text.

• For every failure mode with AP of H, is a prevention or detection control linked on the control plan side?
• Are all items designated as special characteristics in the FMEA included in the control plan without omission?
• When an FMEA line has been revised due to a 4M change or complaint, has the corresponding control plan version also been updated?
• Do the process numbers and process names match across the process flow, PFMEA and control plan?

The first of these checks can be run mechanically, even without AI, in a tool where the FMEA and control plan are structurally linked. AI’s role is to present candidate control methods and reaction plans for failure modes that have no linked control. Siemens adding suggestions for control plan reaction plans in 2606 is a move directly related to this area. However, it is people who decide whether a proposed control method can actually be carried out with the existing equipment and measuring instruments. For how to demonstrate FMEA and control plan alignment in customer audits, “Customer Audits and Traceability” may also be helpful.

FMEA AI Implementation Costs and ROI: Labor Savings Alone Do Not Pay Back

From here on, we present a calculation for Model Plant R. To repeat, all the figures below are original assumptions for this article and are not industry averages.

Assumptions for Model Plant R

Model Plant R is a Japanese automotive parts manufacturer (stamping, welding, plastic molding) in Rayong Province, Thailand. It is a Tier 1/Tier 2 supplier with 450 employees and IATF 16949 certification.

The loss of 250,000 THB per complaint includes sorting, 8D response, expedited shipping and customer chargebacks.

Assumed configuration and costs

The configuration is “an FMEA tool with AI features plus search of past FMEAs, 8D reports and 4M change records, with S/O/D and actions decided by people.”

Note that nearly half of the initial investment goes not to the tool but to data preparation. As described in the earlier section, if you cut this, the AI can only produce generic content.

Key Point 1: Labor savings alone produce a loss

Assume that thanks to AI candidate generation, new creation becomes 30% faster and revisions 25% faster.

In monetary terms, 672 hours × 450 THB = 302,400 THB/year. However, annual operating costs are 450,000 THB, so the annual net benefit is 302,400 − 450,000 = −147,600 THB/year. It does not even cover operating costs, and the initial investment cannot be recovered. The 5-year cumulative total is −147,600 × 5 − 850,000 = −738,000 − 850,000 = −1,588,000 THB.

This is what happens if “writing FMEAs faster with AI” is the only reason for implementation. Assuming a larger percentage of labor savings would change the numbers, but the time spent discussing in the CFT, evaluating S/O/D and deciding actions does not shrink much even with AI. If anything, it is time that should not be shortened.

Key Point 2: Preventing 2 complaints per year makes it pay back

Next, assume that by searching past 8D reports, complaints and 4M change records, the AI presents “failure modes that occurred in this process in the past,” FMEA gaps decrease, and as a result 2 of the 6 complaints per year are prevented.

• Complaints avoided: 2 × 250,000 THB = 500,000 THB/year
• Annual net benefit: 302,400 + 500,000 − 450,000 = 352,400 THB/year
• Simple payback period: 850,000 ÷ 352,400 = about 2.4 years
• 5-year cumulative: 352,400 × 5 − 850,000 = 1,762,000 − 850,000 = 912,000 THB

However, if only 1 complaint is prevented, the annual net benefit is 302,400 + 250,000 − 450,000 = 102,400 THB/year, and the simple payback period is 850,000 ÷ 102,400 = about 8.3 years.

Whether it is 1 complaint or 2 changes the picture dramatically. That is why you should decide at the very start, before implementation, on a design that classifies and records the root cause of each complaint as either “anticipated in the FMEA but the control did not work” or “missed in the FMEA.” Without this classification, even if complaints decrease after implementation, you cannot explain whether that is the effect of AI reducing gaps or of some other factor. We also cover how to decide when a PoC should end in detail in “Exit Criteria for AI PoCs.”

Key Point 3: The cost of getting 1 severity rating wrong on the low side

Let us also put a number on the risk in the opposite direction. Suppose an AI suggestion is adopted as is, Severity S is set too low, a failure mode treated as not requiring action escapes once, and a loss of 1,000,000 THB occurs. This is equivalent to about 3.3 years’ worth (1,000,000 ÷ 302,400) of the labor savings of 302,400 THB in Key Point 1.

Several years of labor savings can be wiped out by a single severity rating set too low. The conclusion is clear. Only people assign Severity S. AI S/O/D suggestions are something to measure for “deviation from human evaluation” in acceptance testing, not something to adopt. In the SAE paper introduced earlier, too, there is an example where the accuracy of S/O/D scoring stayed at 70-90%. Note that this 1,000,000 THB is a yardstick for comparison, and is not meant to be subtracted from the net benefit above to recalculate the payback period.

A note on avoiding double counting

The loss of 250,000 THB per complaint includes the labor hours for 8D response. Meanwhile, the 960 hours of FMEA revisions also include revisions triggered by complaints. In this calculation, the effect of complaint prevention is counted as “avoided loss,” and the same hours are not counted twice as revision labor savings. Also, fewer complaints could reduce the number of complaint-triggered revisions, but in this calculation the number of revisions is fixed at 120 per year (a conservative assumption).

FMEA Software Comparison: 12 Items to Include in the RFP for AI Features

When comparing FMEA software, we recommend writing the following 12 items into your request for proposal (RFP) and having each vendor answer in the same format, rather than relying on impressions from demos.

As a supplementary point, whether a vendor holds certification to ISO/IEC 42001 (the AI management system standard published in December 2023) can be one factor in your selection. However, it is not something required in IATF 16949 audits. For how to evaluate development companies and implementation partners, see also “How to Choose an AI Development Company.”

What to Check in FAT/SAT: Measuring “Gaps” and “Fabrications”

In acceptance testing (factory acceptance test, FAT, and site acceptance test, SAT), confirming in a demo that “a plausible FMEA comes out” is meaningless. What you should measure is gaps and fabrications for your own processes.

As a way to build a test FMEA set, we recommend selecting several FMEAs for part numbers that have had complaints or 8D reports in the past, hiding the failure modes that actually occurred, and having the AI propose candidates. Because you know the answers, you can objectively measure how many known failure modes the AI picks up. However, if the relevant 8D reports and complaints remain in the reference data, the test becomes a test of “whether it could search them,” so if you want to see generalization, also measure under conditions where they are removed from the reference data.

The example pass criteria are only examples from this article. Please decide them after estimating with your own past data and reaching agreement between the CFT and the quality assurance department. What matters is not to measure once at FAT and stop, but to repeat the same tests at SAT with your own data and in your own language environment.

Thailand-Specific Issues: Confidentiality, PDPA, the Draft AI Act and Multilingual Teams

Customer confidentiality and cloud AI

FMEAs contain customer drawings, process conditions, defect histories and customer names. First, check whether your NDAs and quality agreements with customers restrict storage in the cloud or provision to third parties.

On the LLM provider side, Anthropic’s commercial terms state that customer content will not be used to train models. OpenAI’s API documentation says that data sent via the API is not used for training unless you explicitly opt in, that abuse monitoring logs are retained for 30 days by default, and that zero data retention (ZDR) is available with prior approval. However, “not used for training” and “not stored” are different matters. In addition, when data is sent to an external LLM via a tool, as with Relyence, you need to check the terms of both the tool vendor and the LLM provider.

PDPA (Personal Data Protection Act)

Failure modes and process conditions themselves are usually not personal data. However, the names of FMEA CFT members, persons in charge in 8D reports and customer contacts can be personal data. It is safer to design the process so that names are removed before data is passed to the AI. For specific handling, please check individually with your legal department or experts.

Thailand’s draft AI Act

On July 2, 2026, Thailand’s Electronic Transactions Development Agency (ETDA) published a new draft Artificial Intelligence Act and held a public hearing. According to an explanation by the law firm Tilleke & Gibbins, the draft is risk-based and includes classifications of prohibited AI, high-risk AI and AI subject to registration, extraterritorial application to acts affecting people in Thailand, strict liability and more. Because it is still at the draft stage and has not been enacted, it is not known at present how in-house FMEA support tools will be treated. Keep watching future developments and, where necessary, check individually with the competent authorities or experts.

Multilingual CFTs

The CFT at a Japanese plant in Thailand involves Thai process engineers, Japanese expatriates and the design and quality departments at the head office in Japan. It is normal for past 8D reports and complaint records to mix Thai, Japanese and English. Unless you standardize the terminology for failure modes and causes with a dictionary, the same “burr” or “welding defect” will be treated as different things because of variations in language and wording, and the AI’s search will come up empty. Like data preparation, building the dictionary is work that should start in the first 30 days.

A 90-Day FMEA AI Implementation Plan

Finally, here is an example plan for taking implementation through to a decision in 90 days.

The key is to build a “counting mechanism” in the first 30 days, before touching the AI tool. Without records classifying the root causes of complaints, you will have no material to explain an effect like the one in Key Point 2 when making the decision in days 61-90. Also, the reason for limiting parallel operation to 1 process is to keep the comparison conditions the same. If you start with several processes at once, you will not be able to tell whether the difference is the effect of AI or the difference in process difficulty.

Frequently Asked Questions (FAQ)

Q1. What does using generative AI for FMEA mean? How is the AI feature in FMEA software different from having ChatGPT write it?

Using generative AI for FMEA means having AI propose candidates for structure, functions, failure modes, controls and so on, while people decide whether to adopt them. With the AI features in FMEA software, suggestions go into the tool’s FMEA structure and are covered by version control, approval, audit trails and control plan linkage. When a general-purpose generative AI writes it, those mechanisms are absent, and the work of transcription and alignment remains. Also, pasting confidential information into personal plans should be avoided.

Q2. Will an AI-created PFMEA pass an IATF 16949 audit?

What audits are likely to examine is not whether a tool created it, but the records of CFT review, alignment with the control plan, and whether it is up to date, reflecting complaints and changes. The IATF 16949 2nd edition is planned for publication in mid-2027 (indicative), but the documents published so far make no direct reference to FMEA or AI. For specific judgments, please check individually with your certification body.

Q3. Can we leave the S/O/D (Severity, Occurrence, Detection) evaluation to AI?

We recommend not leaving it to AI. Research includes an example where the accuracy of S/O/D scoring stayed at 70-90%, and in particular, if Severity S is mistakenly set too low, necessary actions will be missed. In this article’s model calculation, the loss from 1 severity rating set too low was equivalent to about 3.3 years of labor savings. It is safer to treat AI S/O/D suggestions as reference values for measuring deviation from human evaluation in acceptance testing.

Q4. What are the key points when comparing FMEA software?

Rather than relying on impressions from demos, write the 12 items in this article into your RFP and collect answers in the same format. Items where differences tend to appear include whether S/O/D suggestions can be turned off, whether the past records behind a suggestion can be displayed, whether the data sent to the LLM can be made visible and restricted, whether external communication occurs even with the on-premises version, and whether there are structural links with the control plan.

Q5. What are the costs and benefits of FMEA AI implementation?

In this article’s calculation for Model Plant R (original assumptions), against an initial investment of 850,000 THB and annual operating costs of 450,000 THB, labor savings alone produced an annual net benefit of −147,600 THB and could not be recovered. If 2 complaints per year caused by FMEA gaps are prevented, the annual net benefit is 352,400 THB and the simple payback is about 2.4 years. The benefit depends on how many complaints you can prevent.

Q6. How long does implementation take?

This article gave as an example a 90-day plan in total: 30 days for record-keeping and data preparation, 30 days for parallel operation with 1 product family and 1 process, and 30 days for FAT/SAT and the expansion decision. It is common for the first 30 days to run longer depending on how scattered your data is and whether foundation FMEAs exist.

Summary

• What you should buy through FMEA AI implementation is not “a tool that writes faster” but “a tool that finds gaps.” Both the AI features that appeared in tools one after another in 2026 and the research are consistent in the design principle “AI proposes, people decide.”
• What you leave to AI is generating candidate structures, functions and failures, and cross-checking against the control plan. People evaluate S/O/D and decide actions, and in particular only people assign Severity S.
• The quality of the output is determined by reference data such as foundation FMEAs, past 8D reports and complaints, and 4M change records. Data preparation is the first big hurdle.
• In acceptance testing, measure the recall of hidden known failure modes, the rate of unsupported candidates, and the control linkage check for AP=H.
• In the model calculation, labor savings alone produced a loss; preventing 1 complaint per year gave a payback of about 8.3 years, and preventing 2 per year gave a payback of about 2.4 years. Classifying and recording the root causes of complaints as “anticipated / missed” before implementation is the prerequisite for explaining the effect.
• Confidentiality, PDPA, Thailand’s draft AI Act and standardizing multilingual terminology are issues particular to plants in Thailand. Where judgment is needed, please check individually with experts or your certification body.

At TOMAS TECH, we are happy to help from the stage before tool selection, such as how to classify the root causes of complaints and which past records to have the AI reference. If you are considering how to firm up the alignment between your FMEA and control plan, please feel free to reach us through our contact form.

References

• Siemens (Teamcenter Quality 2512): https://blogs.sw.siemens.com/teamcenter/whats-new-teamcenter-quality-2512/
• Siemens (Teamcenter Quality 2606): https://blogs.sw.siemens.com/teamcenter/whats-new-teamcenter-quality-2606/
• Relyence (Relyence 2026 announcement): https://relyence.com/?p=17326
• Relyence (SmartSuggest user guide): https://relyence.com/help/user-guide/smartsuggest-generate-suggestions.html
• Omnex Systems (AQuA Pro agentic AI webinar): https://my.omnex.com/webinars/transforming-fmea-launch-of-agentic-ai-in-aqua-pro
• PR TIMES (Things Inc. contribution to the June 2026 issue of Kikai Sekkei): https://prtimes.jp/main/html/rd/p/000000018.000089483.html
• VDA (AIAG & VDA FMEA Handbook): https://webshop.vda.de/QMC/en/aiag-vda-fmea-handbook_eng
• Quality-One (Explanation of AIAG-VDA FMEA): https://quality-one.com/aiag-vda-fmea/
• IATF Global Oversight (SC-2026-005 notice): https://www.iatfglobaloversight.org/?p=6589
• Stakeholder Communiqué SC-2026-005 (IATF 16949 2nd edition status): https://media-cdn.smithers.com/mediacontainer/medialibraries/smithersb4f/resources/2026/downloads/stakeholder-communique-sc-2026-005_iatf-16949-2nd-edition-status.pdf
• Smithers (IATF 16949 top 5 nonconformances): https://prd-b4f.smithers.com/resources/2025/december/iatf-16949-top-5-non-conformances-via-iaob
• 16949store.com (Explanation of the IATF 16949 control plan): https://16949store.com/articles/iatf-16949-control-plan/
• SAE 2026-01-0106 (AI Powered Failure Mode and Effect Analysis in Safety Critical Industry): https://saemobilus.sae.org/papers/ai-powered-failure-mode-effect-analysis-safety-critical-industry-2026-01-0106
• El Hassani et al. (Design Science): https://www.cambridge.org/core/product/22F110A2BF0DB4D01A69472CF17A0B43/core-reader
• Collier et al. (How good are large language models at product risk assessment?): https://vtechworks.lib.vt.edu/items/01506cca-0075-4ffe-8553-49ede6fb6d90
• IBM Research (FMEA Builder, arXiv 2411.05054): https://arxiv.org/abs/2411.05054
• Younus et al. (arXiv 2511.17743): https://arxiv.org/abs/2511.17743
• ISO/IEC 42001:2023: https://committee.iso.org/standard/81230.html
• Anthropic (Commercial Terms of Service): https://www.anthropic.com/legal/commercial-terms
• OpenAI (Your data): https://developers.openai.com/api/docs/guides/your-data
• Tilleke & Gibbins (Explanation of Thailand’s draft AI Act): https://www.tilleke.com/insights/thailand-releases-new-draft-artificial-intelligence-act/
• Trading Economics (Thai vehicle production, FTI announcement): https://tradingeconomics.com/thailand/total-vehicle-sales/news/587114
• Thailand Board of Investment (BOI) first-half 2026 announcement: https://www.boi.go.th/upload/content/PR117_2569EN.pdf
• New Straits Times (Thailand’s semiconductor and electronics investment applications): https://www.nst.com.my/amp/business/economy/2026/08/1503841/thailand-attracts-us268bil-chip-electronics-investment-bids-supply