Anomaly Detection AI 2026 | Where to Start Among Five Areas

Search for “anomaly detection AI” and you get predictive maintenance, visual inspection, process monitoring, abnormal sound detection and motion analysis all filed under one phrase. Those five need different data, different budgets and different payback periods. On top of that, October 2026 sees two of the best-known cloud anomaly detection services shut down within a week of each other. This article scores the five areas against four axes to decide where to start, then models what getting the order wrong costs, using a Japanese-owned plant in Thailand.

The phrase anomaly detection AI covers five different things

The main reason an anomaly detection project stalls before it starts is that everyone in the room is picturing something different. Production engineering is thinking about equipment failure prediction. Quality assurance is thinking about catching defective parts. IT is thinking about which cloud service to select. The phrase is the same, but the data required, the budget and the time to first result are not.

An explainer article from AI Souken sets out five main areas of anomaly detection AI in manufacturing. This article uses that framing as its starting point.

One note on naming. The source labels Area 5 as work analysis. This article uses motion analysis throughout, to make clear that the area is about watching how people and objects move on video.

The same article carries examples. Daikin is reported to have shortened its product improvement cycle by more than a year against its previous pace by using operating data from its products for anomaly detection. Bridgestone is described as analysing 480 items of sensor data per tyre in real time on its tyre forming equipment, improving roundness by more than 15 percent against the conventional method, and reaching roughly twice the previous productivity when combined with its multi-drum forming method. JFE Steel is reported to have developed J-mAIster, a system that supports recovery from control faults, and to have completed its rollout to six steelworks and manufacturing sites nationwide.

All three are large-scale cases, and none of them transfers directly to your own plant. Put them side by side, though, and one thing is clear. Anomaly detection AI is not a single technology. It becomes an entirely different project depending on the kind of data involved. Daikin used product operating logs, Bridgestone used time-series sensor data from forming equipment, JFE used fault records from control systems. What they share is that each of them settled where its data would come from before starting. Daikin and JFE began from records they already held. Bridgestone began by building the measurement into the forming system itself. The entry point differs, existing records in one case and new measurement in the other, but in both cases the data question was answered before any model or service was chosen.

In October 2026, two cloud anomaly detection services close within a week

Until now, the argument that you should start from your own data has been brushed aside with one objection. That may be so, but why not just try a cloud service first? From October 2026, the general-purpose services the two major vendors offered stop being an option.

Azure AI Anomaly Detector retires on 1 October 2026

Microsoft’s official documentation states that the Anomaly Detector service is being retired on 1 October 2026, and that from 20 September 2023 you can no longer create new Anomaly Detector resources. The migration paths offered are Microsoft Fabric or the open-source anomaly-detector project.

In other words, the service stopped accepting new users almost three years ago. What ends on 1 October 2026 is the part that has been kept running for existing users.

AWS Lookout for Equipment is discontinued on 7 October 2026

An official AWS blog post states that new customers will not be able to access Amazon Lookout for Equipment effective 7 October 2025, while existing customers can use the service as normal until 7 October 2026, and announces that the service is discontinued on that date. The migration path offered is AWS IoT SiteWise, whose multivariate anomaly detection capability added in July 2025 supports Lookout for Equipment modelling, with migration scripts for existing models provided on GitHub.

What to read from this is that the platform stays and the finished product goes

The two ending six days apart is a coincidence. What they mean is the same. At least the layer described as a general-purpose API that returns anomalies when you throw data at it, offered by hyperscalers, is no longer being maintained as a standalone product. Not every service from every vendor is disappearing, but the fact that the two best-known ones were folded at the same time is reason enough to re-examine any plan that bets on this layer.

What matters is the nature of the migration targets. Neither Microsoft Fabric nor AWS IoT SiteWise is a finished anomaly detection product. Both are platforms that assume you bring your own data and assemble the solution yourself. The amount of work you can hand off has clearly gone down.

The practical conclusion follows directly. An anomaly detection project should not start from the question of which service to buy. Services end. Your data, and your definition of what you want to detect in it, do not. So the first thing to decide is which of the five areas your data currently sits in.

Scoring the five areas against four axes

Decide where to start by scoring four axes. Each axis is scored from 0 to 3 points, giving each area a total between 0 and 12. Start with the area that scores highest. That is the whole model.

The scoring rules for each axis follow. You score four axes for each of five areas, so twenty judgements in total. An hour is enough.

Axis A, existing data

Score the data you are capturing at this moment, not the data you plan to capture. “We could get it if we fitted sensors” is 0 points. The moment installation work and a budget request are required, ease of starting has fallen.

The gap between 2 and 3 points is the largest step in practice. Between data existing inside a machine and data you can train on sit three pieces of work, adding communications, decoding the format, and providing somewhere to store it. Plans that treat this lightly almost always slip.

Axis B, history of anomaly events

Anomaly detection AI comes in two forms, one that learns only the normal state and flags outliers, and one that learns both normal and abnormal and classifies. The first appears to need no labels, but you still need records of past anomalies at the point where you verify whether a flagged outlier was actually a problem. Skip that and the shop floor will say “this is not an anomaly” every time an alert fires, and within three months nobody will be looking at it.

Axis C, annual loss

State the amount in THB, on an annual basis. What matters here is never counting the same loss in two areas. An unplanned stop caused by bearing noise is counted once, as a loss in Area 1, not in Area 4. Allow double counting and every area scores high, and nothing can be ranked.

There is a reason the amounts are held as bands. A loss estimate moves 20 percent on assumptions alone. Comparing to the nearest THB is meaningless, so the bands are cut at 1,000,000, 3,000,000 and 10,000,000, steps of three times or more, and only crossing a band changes the ranking.

Axis D, action after detection

This is the axis people overlook, and it is the one that actually kills projects. Detect an anomaly with no spare equipment, no spare parts and nobody on site authorised to stop the line, and all you are left with is the fact of having detected it.

How to read the score, and the one override

Start with the area that has the highest four-axis total. On a tie, prefer the area with the higher Axis D score, because an area where you can act the same day makes results visible inside the company faster.

There is exactly one override. An area scoring 0 on Axis D is not started, whatever its total. Even with full marks on the other three axes, an investment does not pay back if nothing can be done after detection. A common case in Thai plants is a single specialised machine imported from Japan, with a two-week lead time on parts and no alternative line. For that machine, the first step is holding spare parts or securing an alternative. AI comes after that.

Scoring a model plant

Abstractions only go so far, so we will set up a plant that could plausibly exist and score it.

A Japanese-owned electronic components maker in Rayong, moulding, pressing and plating connectors and terminals for automotive use. 450 employees, two shifts.

Start with the contribution lost per hour of stoppage. 780,000,000 THB divided by 250 days is 3,120,000 THB a day, and divided by 16 hours that is 195,000 THB an hour. Apply the 28 percent contribution margin and you get 54,600 THB an hour.

How much each of the five areas loses a year

Build up the loss area by area. To avoid double counting, decide which area each loss belongs to before counting it.

Area 1, equipment monitoring and predictive maintenance. Of the 340 hours of unplanned stoppage, assume 25 percent could not be recovered through overtime or weekend working and hit shipments. 340 times 25 percent is 85 hours, and 85 times 54,600 THB is 4,641,000 THB. The remaining 255 hours are recovered through overtime, which carries a premium. With 8 people on overtime at a wage-based hourly rate of 210 THB including the premium (this figure carries no overhead loading), 255 times 8 times 210 is 428,400 THB. Add 620,000 THB a year of outsourced emergency repairs and expedited parts. The total is 5,689,400 THB.

Area 2, visual inspection and quality. Against annual shipment volume of 42,000,000 pieces, assume 0.05 percent escapes visual inspection and reaches the customer, giving 21,000 pieces. Assume those surface as 24 complaints a year, each costing on average 165,000 THB for sorting, recovery, transport, reporting and expedited replacement production, so 24 times 165,000 is 3,960,000 THB. On top of that, 100 percent visual inspection costs 14 people times 2,000 hours times 165 THB, or 4,620,000 THB (165 THB is the internal loaded rate, wages plus statutory benefits and overhead, not the hourly value of salary itself), of which perhaps 30 percent could be displaced by automated detection, giving 1,386,000 THB. The total is 5,346,000 THB.

Area 3, process monitoring. Poor control of plating bath concentration and temperature causes 18 lots a year to be scrapped, at an average of 96,000 THB per lot for material, labour and re-plating. 18 times 96,000 is 1,728,000 THB.

Area 4, abnormal sound detection. This is the wasted electricity from air leaks in the compressor system. Annual electricity cost attributable to compressors is 2,400,000 THB, and with an estimated air leak rate of 20 percent, 2,400,000 times 20 percent is 480,000 THB. Unplanned stops caused by bearing noise are already counted in Area 1 and are not counted again here.

Area 5, motion analysis. Rework caused by deviation from the work procedure runs at 310 cases a year at 2,900 THB each. 310 times 2,900 is 899,000 THB.

The five areas add up to 14,142,400 THB. Checking the arithmetic, 5,689,400 plus 5,346,000 plus 1,728,000 plus 480,000 plus 899,000 is 14,142,400, which matches.

The scores

Apply the plant’s current state to the four axes.

Here is the reasoning behind each score.

Area 1 scores 2 on Axis A because the injection moulding machines hold temperature and pressure inside the controller, but there is no way to get the data out and it is overwritten. Axis B is 1 point because stoppages are written up in the maintenance log book but without a time to the minute. Axis D is 2 points because the site can decide, but folding work into a planned stoppage goes through a monthly meeting.

Area 2 scores 3 on both Axis A and Axis B because an automated visual inspection machine is already installed, a daily routine writes both the judgement results and the captured images to a file server, and more than 12 months of time-stamped images of good and defective parts are held, far more than 1,000 of each. Axis D is 3 points because a rejected part is simply removed from the line on the spot, requiring no parts and no procedure.

Area 3 scores 1 on Axis A because plating bath concentration is measured by hand twice a day and recorded on paper, while temperature and current are only read off local indicators. Axis B is 1 point because the date a lot was scrapped is known but the time at the causing process cannot be identified. Axis D is 3 points because the bath can be corrected on site the same day.

Area 4 scores 0 on Axis A because not a single microphone is installed, and 0 on Axis B because no record exists of when abnormal noise or an air leak occurred. Axis D is 3 points because an air leak can be repaired on the pipework the same day. Area 5 scores 1 on Axis A because the ceiling security cameras only display to a monitor, with nothing kept on a recorder, and 0 on Axis B because rework counts are tallied but carry no timestamp that ties to the video. Corrections to procedure can be made on site, but retraining runs monthly, so Axis D stops at 2 points.

The verdict is unambiguous. Start with Area 2, visual inspection and quality.

What deserves attention is the structure of the ranking. The largest annual loss is Area 1 at 5,689,400 THB, above Area 2 at 5,346,000 THB. Area 2 still wins because Axis C is scored in bands, both land on 2 points, and the 343,400 THB difference in loss never reaches the ranking. The gap opens on the other three axes, 3 points across Axis A and Axis B, a further 1 point on Axis D, for a total of 11 against 7, a 4-point margin. Decide by size of loss alone and equipment wins almost every time. And equipment is, in most plants, the area where the data is thinnest.

Going deeper into the equipment area is outside the scope of this article. Which conditions make predictive maintenance work and which do not is covered in the article sorting predictive maintenance cases into what works and what does not, and what vibration as a signal can and cannot measure is covered in the article on equipment diagnosis with vibration sensors. Scoring Axis A and Axis B correctly assumes you have read both.

What getting the order wrong costs

A decision model is worth whatever it saves you when you get it wrong. Three scenarios, compared. The benefit estimates are re-stated conservatively as only what can genuinely be captured, separately from the rough loss figures used for the Axis C band judgement.

Scenario A, following the verdict and starting with Area 2

Annual running cost is 290,000 THB for maintenance and retraining. The benefits are set out as follows.

Visual inspection headcount falls from 14 to 10, but booking the full 1,320,000 THB of labour cost for those 4 people would not be honest. In Thailand, freed-up headcount is normally redeployed to other processes rather than laid off, and what actually reaches the financials is the reduction in overtime and the avoided hiring. Here that is taken at 60 percent, or 792,000 THB.

Annual net benefit is 2,442,000 minus 290,000, or 2,152,000 THB. Simple payback is 2,230,000 divided by 2,152,000, or 1.04 years.

Year one has a ramp, though, so put benefits at 60 percent of the full figure and running cost at half a year. Year one comes to 1,465,200 minus 145,000 minus 2,230,000, or −909,800 THB. Year two carries the full 2,152,000 THB, so the cumulative position turns positive at around 1 year 5 months.

Cumulatively that is −909,800 in year one, 1,242,200 in year two and 3,394,200 THB in year three.

Scenario B, starting with Area 1 because the loss is bigger

This is what happens when the argument that equipment is where we lose the most wins the internal debate.

Annual running cost is 380,000 THB. The decisive point here is that almost nothing lands in year one. Axis B scores 1, meaning there is no record that pins down when an anomaly occurred to the minute, so a failure prediction model cannot be trained. All year one can do is accumulate data and raise threshold alerts on obvious deviations.

Assume year one prevents 8 percent of the 340 hours of unplanned stoppage. 340 times 8 percent is 27.2 hours. The 25 percent of that which hits shipments, 6.8 hours, is 6.8 times 54,600, or 371,280 THB. The 20.4 hours previously recovered through overtime work out at 20.4 times 8 people times 210, or 34,272 THB. Repair cost falls by 620,000 times 8 percent, or 49,600 THB. That totals 455,152 THB. Subtract running cost and only 75,152 THB is left.

From year two, 12 months of time-stamped history has accumulated and a predictive model becomes viable. Raise the reduction rate to 22 percent and 340 times 22 percent is 74.8 hours. The 25 percent hitting shipments, 18.7 hours, is 18.7 times 54,600, or 1,021,020 THB. The 56.1 hours previously recovered through overtime are 56.1 times 8 people times 210, or 94,248 THB. Repair cost falls by 620,000 times 22 percent, or 136,400 THB. That totals 1,251,668 THB, and 871,668 THB after running cost.

Cumulatively that is −2,714,848 in year one, −1,843,180 in year two and −971,512 THB in year three. Three years in, it has still not paid back. The cumulative position turns positive around 4 years 1 month, just after the start of year five.

The three-year gap against Scenario A is 3,394,200 minus −971,512, or 4,365,712 THB.

One thing needs to be said plainly here. Scenario B is not a technical failure. It does pay back from year five onward, and predictive maintenance on equipment is a sound investment in itself. The problem sits elsewhere. A project that produces nothing in year one does not survive internally. Japanese-owned plants in Thailand report annually to the head office in Japan, and a project that cannot show a number in its first-year benefits review loses its year-two budget. Stopping just before the training data was complete is the most common way projects in this area end.

Scenario C, delivering results in Area 2 first, then moving to Area 1

This is the realistic path, following the verdict while also starting on the equipment area in year two.

When Area 1 starts in year two, the 620,000 THB data platform built for Area 2 can be shared, so the initial cost falls to 2,790,000 minus 620,000, or 2,170,000 THB. Anomaly detection areas differ, but the place time-series data is stored and the way it is visualised are common to all of them.

Checking the arithmetic, year two is 2,152,000 minus 2,170,000 plus 455,152 minus 380,000, or 57,152. Year three is 2,152,000 plus 871,668, or 3,023,668. The cumulative position is −909,800 plus 57,152 plus 3,023,668, or 2,171,020.

Comparing the three scenarios

Three things can be read from this.

First. At the three-year mark, the gap between A and B is 4,365,712 THB. Same plant, same order of budget, and the only difference is which area you started with.

Second. C is behind A right through to year four. At the end of year four the gap is 351,512 THB, and C overtakes at around 4 years 5 months. Additional investment in the equipment area takes more than four years to catch up on a cumulative basis even when done in the right order. That means Area 1 is something you should do, but not something to rush. If the single-year budget is tight, stopping at Area 2 is a perfectly defensible choice.

Third. Even so, C has the highest annual net benefit in a normal year. If you are running the plant on a horizon longer than five years, C is the right answer. The model is not saying do not do Area 1. It is saying do not put Area 1 first.

Sensitivity analysis, does the conclusion survive halving the complaint reduction

Of the 2,442,000 THB of benefit in Scenario A, 1,650,000 THB is complaint reduction. A single assumption carries 68 percent of the total, so it is worth flexing.

Take the case where complaints fall by 5 cases rather than 10. The 792,000 THB reduction in visual inspection effort does not depend on that assumption, so it is left as it is with no factor applied.

Annual net benefit is 1,617,000 minus 290,000, or 1,327,000 THB. Year one is 970,200 minus 145,000 minus 2,230,000, or −1,404,800, the year-two cumulative is −77,800, and the year-three cumulative is 1,249,200 THB. The turn to positive slips back to around 2 years 1 month.

The gap against Scenario B’s three-year cumulative of −971,512 THB is 2,220,712 THB. The conclusion does not flip. Even with complaint reduction halved, starting with Area 2 is more than 2,000,000 THB better over three years.

Read the other way, the persuasive weight of this case does not rest on complaint reduction, an uncertain item, but on the 792,000 THB of visual inspection effort, which occurs every single day. In an approval paper, leading with the effort reduction should get you further.

There is a second sensitivity worth running, on the verdict itself. If Area 1 moves from 1 to 3 points on Axis B and from 2 to 3 points on Axis A, its total becomes 3 plus 3 plus 2 plus 2, or 10 points. That is still 1 point short of Area 2’s 11, but the gap to Areas 3, 4 and 5 becomes decisive. How to lift those two axes is the subject of the next section.

The fastest way to raise your score is not buying AI

This continues from the point at the end of the previous section about lifting Axis A and Axis B for Area 1. This is the most practical part of the article.

What it takes to move Axis B from 1 point to 3 is neither AI nor sensors. It is adding four columns to the maintenance log book, stop start time, recovery time, symptom and action taken. Add one operating rule saying they are written to the minute, and a year later you will have 10 or more time-stamped anomaly events. The cost is close to zero.

Moving Axis A from 2 points to 3 does need installation work, but it does not need to be done all at once. Of the 12 injection moulding machines, take only the 3 with the longest downtime over the past two years, and pull data out of their controllers and store it. At that scale a few hundred thousand THB covers it. There is no reason at all to connect all 12 machines simultaneously.

So the right first-year investment in the equipment area is not anomaly detection AI itself. It is the work of raising your score. Start the area in year two with a higher score and the period where nothing lands in year one, the one that hurt Scenario B, disappears. Note that Scenario C above is a conservative calculation that does not build in this head start and leaves Area 1’s first year at an 8 percent reduction rate. Add the four log book columns first and year two of Scenario C looks somewhat better.

The same thinking applies on the quality side. A plant without an automated visual inspection machine is not thereby barred from starting with Area 2. It starts by adopting a routine of keeping, rather than discarding, the images captured at the inspection process. The argument that judgement accuracy is set by the proportion of data you can cross-reference is covered in detail in the article on AI analysis of quality inspection data.

Points specific to running this in Thailand

BOI technology upgrading incentives may be available

The Thailand Board of Investment has a category called Activity 10.1, technology upgrading incentives. Where a manufacturer already receiving BOI privileges invests in predictive maintenance systems such as IoT sensors, data platforms and equipment failure prediction AI models, or in visual inspection AI using image recognition and deep learning, that investment can qualify for an additional 3 years of corporate income tax exemption plus import duty exemption on new machinery and AI-related equipment. There is no minimum investment threshold for the technology upgrade itself. The only precondition is that the base activity is already receiving privileges.

Applied to Scenario A in this article, of the 2,230,000 THB initial cost, the 340,000 THB for the industrial PC and edge inference units could fall within the import duty exemption. Rates vary by item classification, so confirm the actual amount case by case.

There are people-side incentives too. AI training expenses qualify for a 200 percent deduction, and a company investing between 1 percent and 3 percent of its annual payroll in AI training can receive an additional 1 to 3 years of corporate income tax exemption depending on the proportion invested. 1 percent buys 1 year, 2 percent buys 2 years, 3 percent buys 3 years.

The amounts involved are not small. With the model plant’s 450 employees and an average annual salary per head of 240,000 THB including employer contributions, the annual payroll used to judge the incentive is 108,000,000 THB. 1 percent of that is 1,080,000 THB. That means committing close to half of Scenario A’s initial cost to training every single year. This is not a figure to jump at because of the extra years of exemption. Build it up from the number of people you actually want to develop and the courses they would take.

Whether the incentives apply depends on your business activity and the application category. Always confirm case by case.

Make sure the detection decision can be made on site

The difference between 1 and 2 points on Axis D, meaning whether head office approval takes days or the site can decide, weighs more heavily in Thailand than in Japan. An anomaly detection alert loses its meaning if it is not acted on within hours. If the authority to stop equipment sits with the Japanese parent company, and a time difference plus an approval circuit means three days, anomaly detection in that area will not work however much you invest.

This is an organisational problem, not a technical one. But if the organisation is not going to change, Axis D stays at 1 point, and the model correctly pushes that area down the ranking. The conclusion it produces is simply not to force the investment.

Decide where the data goes before you start

The point about cloud anomaly detection services closing has a consequence here too. Where you put training data and where inference runs is a design decision that is hard to change later. Build a design that sends images to an overseas region and the cost of changing it grows sharply the day a customer audit asks you to explain it.

A configuration that runs inference at the edge and does training in a separate environment makes this issue much lighter. That is why Scenario A includes edge inference units in the initial cost.

Frequently asked questions

What is anomaly detection AI

It is the collective term for systems that learn the pattern of a normal state and automatically find states that depart from it. In the framing used here, it splits by the kind of data involved into five areas, equipment monitoring and predictive maintenance, visual inspection and quality, process monitoring, abnormal sound detection, and motion analysis. It is not a single product category, so a statement like “we will implement anomaly detection AI” does not define any requirements. Settle which area you are talking about first.

What is the difference between anomaly detection AI and predictive maintenance AI

Predictive maintenance AI is the name for the part of anomaly detection AI that corresponds to Area 1, equipment monitoring. One contains the other. That said, the phrase predictive maintenance implies a time axis, predicting when something will break and replacing the part before it does, which is a step harder than simply detecting that something is wrong now. Estimating remaining time to failure needs data from several actual failures. In a plant with a low Axis B score, the sensible route is to start with anomaly detection and move to prediction once cases have accumulated.

How much data do you need to build an equipment failure prediction AI

There is no single answer, but the scoring rules here put full marks on Axis B at 10 or more time-stamped anomaly events a year. Treat that as one rule of thumb. Read the other way, you cannot build a failure prediction model from one year of data for a machine that only breaks once a year. If you run several units of the same model, cases accumulate across the fleet, so starting with the equipment type you have most of is an advantage.

Can a defect detection AI completely replace visual inspection

The model plant here assumes a reduction from 14 people to 10, not to zero. People remain for ambiguous judgements, for first articles and the period straight after a specification change, and for sampling checks on the AI’s own judgements. The realistic substitution is moving from 100 percent visual inspection to sampling. An investment plan built on complete removal of people almost never delivers.

What should we do if we were using Azure AI Anomaly Detector or AWS Lookout for Equipment

Azure AI Anomaly Detector retires on 1 October 2026, with migration offered to Microsoft Fabric or the open-source anomaly-detector project. AWS Lookout for Equipment is discontinued on 7 October 2026, with migration offered to AWS IoT SiteWise and migration scripts provided for existing models. In both cases the destination is a platform rather than a finished product, so we recommend using the migration as an opportunity to re-score whether that area was ever a high scorer for you. It is not unusual to find that keeping the thing running has become the objective in itself.

Is there any reason not to start all five areas at once

If you have the budget and the people, we will not stop you. In practice, though, the part of an anomaly detection launch that takes the most time is the dialogue with the shop floor. Confirming the validity of alerts case by case with the people on the line is not something one production engineer can do for five areas in parallel. Delivering a result in one area and earning the shop floor’s trust before expanding is faster in the end.

What is different about doing this in Thailand compared with Japan

The technical requirements barely change. Three things differ. Whether decision authority sits locally, which feeds straight into Axis D. Whether recording routines become established. And whether you can show a first-year number in the annual report to the Japanese head office. The third of those bears directly on which area you choose. Start in an area that produces nothing in year one and the project will not survive, however sound the technology.

Summary

Anomaly detection AI is not a single technology. It is the collective term for five areas, equipment monitoring and predictive maintenance, visual inspection and quality, process monitoring, abnormal sound detection, and motion analysis. Where to start is decided by scoring four axes from 0 to 3 points, existing data, history of anomaly events, annual loss, and action after detection, and taking the area with the highest total. An area scoring 0 on Axis D is not started, whatever its total.

At the model plant the largest annual loss was equipment monitoring at 5,689,400 THB, but the area to start with was visual inspection, on 11 points. Decide by size of loss alone and equipment wins, and equipment is usually the area where the data is thinnest. Getting the order wrong costs 4,365,712 THB over three years. Even a sensitivity analysis halving the complaint reduction assumption left a gap of 2,220,712 THB, and the conclusion held.

Azure AI Anomaly Detector ends on 1 October 2026 and AWS Lookout for Equipment on 7 October 2026. The option of buying a general-purpose API from a hyperscaler that returns anomalies when you throw data at it disappears, at least for these two. What remains is your own data. So the first question is not which service to buy, but which of the five areas your data currently sits in. And the fastest way to raise your score is not buying AI. It is adding a stoppage time column to the maintenance log book.

Try the four-axis scoring on your own plant, and it is fine if the scores come out close with no clear ranking, or if you cannot even establish the current state of Axis A. Give us four things, your equipment count, the structure of your inspection processes, unplanned stoppage hours over the last twelve months and the number of customer complaints, and we will run a first-pass score and come back with the area to start with and the reasoning behind it. For an informal discussion well before any quotation, please get in touch through our contact page. We will tell you what can realistically be delivered in year one, based on what we have seen implementing these systems in Thailand.

References

• Microsoft Learn, What is Anomaly Detector? (the Anomaly Detector service is being retired on 1 October 2026, and new resources cannot be created from 20 September 2023), learn.microsoft.com
• Microsoft Azure Updates, AI services Anomaly Detector will be retired on 1 October 2026, azure.microsoft.com
• AWS Machine Learning Blog, Preserve access and explore alternatives for Amazon Lookout for Equipment (new customers cannot access the service from 7 October 2025, existing customers until 7 October 2026), aws.amazon.com
• AI Souken, Anomaly Detection AI in Manufacturing (the five-area framing, and the Daikin, Bridgestone and JFE Steel examples), ai-souken.com
• Pertama Partners, Thailand BOI Manufacturing (Activity 10.1 technology upgrading, Last Updated February 9, 2026), pertamapartners.com