Drawing Search AI 2026 | The 3 Layers and the Cost of Factories That Cannot Find Past Drawings

Enquiries about drawing search AI have become noticeably more frequent among Japanese-owned factories in Thailand. Yet when you actually walk into a plant that is genuinely struggling because nobody can find its past drawings, very few of them are in a state where bolting AI on top would solve anything. The reason is that the single sentence “we cannot find our drawings” contains several problems of completely different natures, folded on top of one another. This article separates that sentence into three layers, and sets out which factory is stuck at which layer, in what order it should act, and how much it should expect to spend, including cost ranges.

Before you evaluate drawing search AI, break down what “we cannot find it” really means

When a design or production engineering manager says “we cannot find our old drawings, so we want AI to search them for us”, the first thing to check is not a product feature list. The first thing to check is what is actually happening inside that particular plant.

The phrase is identical from plant to plant, but the reality behind it is not. In one factory the drawings sit on shelves as paper and no electronic data exists at all. In another, the PDFs are on a server, but the file is called new_drawing_final_2.pdf and nobody can tell what it is without opening it. In a third, drawing numbers and revision marks are properly held in a register, and the only thing that cannot be done is the search that starts from “I am fairly sure we made a part shaped like this before”.

These three need completely different remedies. At the point where the evaluation starts, however, all three are reported upward using the same words. And when someone goes to look at a product catalogue, the catalogue says “our AI searches across all your drawings”. So the same product ends up on the shortlist regardless of which of the three the factory actually has.

Here is the conclusion up front. Drawing search AI delivers real value only at the topmost of those layers. Introduce it into a plant where the two lower layers are not in place and the AI has nothing to search in the first place. And in most cases the outcome is recorded internally as a false verdict – “the AI was not accurate enough”. It is not an accuracy problem. It is a structural problem, in that the information that should have been indexed does not exist.

So the first decision on the buying side is not which product to choose. It is to identify which layer your own version of “we cannot find it” is happening at.

“We cannot find the drawing” splits into 3 layers – location, identification, content

Break the complaint apart and you get the following three layers. Each layer can only sit on top of the one below it, and the order cannot be swapped.

L1 Location is the state where nobody knows where the drawing itself physically is. Paper cabinets, individual designers’ PCs, the departmental file server, local CAD working folders, the archive area of a decommissioned server, attachments in email threads with subcontractors. Instances of the same drawing number are scattered across several places and nobody can say with confidence which one is authoritative. At this layer the act of searching does not even become possible, because the scope of what is being searched cannot be defined.

L2 Identification is the state where the files are gathered in one place, but a machine cannot read what each individual sheet is. Drawing number, revision, part name, customer, material, projection method – all of that is written inside the frame called the title block. In a scanned PDF, however, it is nothing more than a collection of lines inside an image. A human eye can read it; from the system’s point of view it is no different from a blank file.

L3 Content is the state where drawing numbers, revisions and attributes are all present and condition-based search works, but you cannot search by content. Wanting to find parts of a similar shape, or parts that carry a Ø50 H7 bore, is a way of searching that exact-match attribute queries never reach.

The column that matters is the last one. L3 AI uses the attributes produced by L2 as its search keys. Buy L3 AI for a plant with no L2 and the AI cannot build an index. That is precisely why the order cannot be skipped.

L1 Location – start by counting how many instances actually exist

Whether you are stuck at L1 can be decided with one simple question. Ask “for the drawings released last year, in how many places does the master copy currently sit?” If nobody can answer immediately, you are at L1.

Investigating this layer is not a systems discussion; it is a counting exercise. The drawing folder on the departmental file server, designers’ PC desktops, local CAD working folders, the archive area of the decommissioned server, attachments in email threads with subcontractors, and the paper cabinets. Count how many sheets sit in each of those six places. When the counting is finished, the total very often turns out to be far larger than the original assumption, because instances of the same drawing exist in duplicate across several locations.

The duplicates that surface here come in two kinds – exact copies, and versions that somebody modified along the way. The former can simply be deleted. The latter is dangerous, because which of them is correct cannot be judged on the spot. Push that judgement down the road, upload everything to the cloud in one go, and every piece of work from L2 onward is contaminated.

So the work at L1 is not “collect”, it is “decide the master copy”. For each drawing number, decide which file is the controlled original and move everything else into an archive area that cannot be referenced. This judgement cannot be automated. Somebody in the design department has to make it, drawing number by drawing number. In a plant holding on the order of 20,000 drawings, that judgement work alone takes anywhere from several weeks to several months.

For paper drawings, do not make a blanket decision about digitisation. Scanning every historical drawing is not always the right answer. Drawings for discontinued products, parts superseded by successor models, and projects whose customer contract has ended may well have to be retained for record-keeping reasons, but very often do not have to be inside the search scope. Because scanning cost is proportional to sheet count, the decision about scope translates directly into money.

L2 Identification – if the title block is still an image, AI can index nothing

L2 is the layer that effectively decides whether drawing search succeeds or fails. It is also the layer most likely to be dismissed as trivial.

A drawing’s title block carries the drawing number, drawing name, revision mark, revision date, drafter, checker, approver, scale, projection method, material, surface treatment and customer name. All of that is information that can be used directly as a search index. In a scanned PDF, or a drawing output as a raster image, that title block is nothing but a set of lines and dots. It holds no information as character codes.

The work that removes that state is L2. Concretely – crop the title block region, run OCR to lift the characters, map the result onto fields such as drawing number, revision and part name, then match it against the existing part master and customer master and normalise it. The mapping and the matching consume far more effort than OCR accuracy itself, because title block layouts differ by era, by designer, and by whether the drawing was supplied by the customer. On choosing the tool that reads the drawings, our comparison of AI-OCR engines sets out what each reading method is and is not suited to.

The problem that matching almost always exposes is notation drift. The same customer appears three different ways as “ABC Co., Ltd.”, “ABC Co Ltd” and “ABC”. The material drifts across “SS400”, “ss400” and “SS-400”. A human reads all of those as the same thing; a search system treats them as different entities. Feed data into a search platform without passing it through this normalisation and the symptom is simple – you search and nothing is returned. And that symptom tends to get reported as insufficient accuracy in the AI further downstream.

How far you go with attributes at L2 is directly tied to cost. Lifting every field at high accuracy is expensive. Realistically, narrow it first to three – drawing number, revision and part name – and make those reliably machine-readable. Customer and material can be added once you have seen how the search is actually used. Completing three fields properly makes search work, whereas aiming at every field from day one tends to exhaust the budget first.

L3 Content – this is where drawing search AI finally earns its place

Once L1 and L2 are in place, drawings can be pulled by drawing number, by part name or by customer. Only when you get this far does the remaining frustration turn out to be L3.

The L3 frustration is specific. “Have we made a part shaped like this before?” “I want to see the drawing of a bracket with a similar hole pattern.” “I want to list every drawing whose notes specify heat treatment.” None of these can be solved by exact-match attribute search, because the request is to trace by shape or by the content of the description, without knowing either the drawing number or the part name.

What AI provides at this layer falls broadly into two functions. One is shape similarity search, which converts the geometry itself into feature vectors and ranks results by similarity. The other is a response to near-natural-language queries against the notes and dimension text inside the drawing. The way the market frames it is much the same – a configuration that combines shape similarity search, OCR of dimensions and notes, and attribute text search is the common pattern, with a movement to extend from there into natural language queries.

The point to hold onto is that L3 AI is a consumer of what L2 produced. Shape similarity search looks at geometric data, so at first glance it appears not to depend on L2, but presenting a result as “which drawing number, at which revision” ultimately requires L2 attributes anyway. Returning ten similar-looking drawings is useless if nobody can tell which product they belong to or which version they are, because the engineer cannot then make a reuse decision.

As AI search over documents, this overlaps with the issues involved in building RAG over internal documents. Drawings, however, are not the same as text documents, because much of the information rides on geometry and layout and only a limited portion can be extracted as characters. Our article on building RAG for factory knowledge sets out the cost and the approach on the text document side, but read it with the difference in mind – with drawings, the weight of L2 pre-processing is far greater.

Drawing search AI has 4 technology components

The phrase “drawing search AI” in practice refers to a bundle of several technology components. Whoever receives the proposal needs to read it apart into which of those components it is built from, because the prerequisites differ from component to component.

Read the right-hand column of that table vertically and you can see that the weight of the prerequisites differs by component. Attribute text search and natural language search take extracted attributes as a direct prerequisite, and shape similarity search also needs the output of L2 at the point where it presents candidates as a drawing number and a revision. In other words, three of the four components are not usable until L2 is finished. The only one that can precede L2 is title block OCR, and that is the tool used to build L2 in the first place.

When you evaluate a proposal, tabulate which of these four are included and which are not. There is no need to buy all four. If your own “we cannot find it” is L2, what you need is the first component, not shape similarity search. Sign a contract with those mixed together and you will keep paying an annual fee for functions you never use.

What is “search accuracy” the accuracy of – are you buying recall or precision

If the word “accuracy” comes up during a proposal meeting, ask for the breakdown. That single word packs together two metrics of quite different character.

One is recall – of the drawings that should have been found, what proportion was picked up without being missed. The other is precision – of the results that came back, what proportion was genuinely what you were looking for. These two are hard to raise at the same time. Loosen the conditions to reduce misses and unrelated drawings creep into the results. Tighten the conditions to exclude the unrelated and the drawings you needed drop out.

Which one to prioritise is decided by what the search is for. When looking for a similar part for design reuse, recall is what counts. Even if the candidate set is somewhat large, the engineer can narrow it down by eye. Miss something here and the same part gets designed twice, which means the tooling and the inspection are also incurred twice.

By contrast, when pulling up “which is the current drawing for this product” in order to answer a customer enquiry, precision is what counts. Ten candidates coming back is a problem. Unless exactly one correct result is returned, the search is meaningless.

This distinction matters because it feeds directly into how the evaluation is designed. A common approach during a trial is for someone to search for ten drawings they happened to think of and judge the product on whether each came back. That measures neither recall nor precision. If you are going to evaluate, fix the correct answer set in advance. Build a mapping table stating that for this search condition, this group of drawings should be returned, and only then run the searches. The work of building that mapping table is itself the work of putting into words what your own organisation wants to find.

Note also that accuracy figures depend heavily on the state of the drawings at each plant. Deploy the same product and the outcome differs between a factory whose title block layouts are consistent and a factory carrying twenty years’ worth of mixed formats. It is better not to transplant a figure quoted in another company’s case study directly into your own expectations.

The boundary between a drawing management system (PDM) and drawing search AI

Take the evaluation of drawing search AI far enough and the question of where the boundary with a drawing management system lies will come up. The two are not competitors; they cover different layers.

What a drawing management system covers is centralised control of the master copy, revision history, approval workflow, access rights, and search by attribute. That is, it covers L1, and it covers the attribute search that only becomes possible once L2 is in place. What it does not do is L2 itself, namely lifting attributes out of the title block and matching them against master data. Attributes are produced either by people or by title block OCR. Drawing search AI, on the other hand, covers the L3 that attribute search cannot reach – shape similarity search and natural language search.

The question “which one do we introduce first” is therefore in practice the same question as “which layer are we stuck at”.

Do not take the fourth row lightly. It is not unusual for the reason search goes unused to lie in the flow of work rather than in a missing function. Behind a designer drawing something new instead of looking for the old drawing is a judgement that drawing it is less effort than finding it. Unless that judgement changes, making the search platform more sophisticated will not increase usage.

Drop revision control and the drawing you found will be the wrong one

Concentrate on the search discussion and the thing that gets left out is revision control. Yet this is far more directly connected to incidents than search accuracy is.

What happens when the drawing you found turns out to be a superseded revision is entirely predictable. The part is machined to the old dimensions, inspected against the old specification, and rejected at the customer. The faster the search, the faster the wrong drawing reaches the shop floor. If current and superseded revisions sit in the search platform with no distinction between them, the probability of that incident can be higher after the deployment than before it.

The requirement for revision control also comes from the quality management system side. Control of documented information under ISO 9001 requires approval, review, updating, identification of revision status, and control of access. What is called identification of revision status there is exactly drawing revision control. That requirement is already present in the current edition, ISO 9001:2015.

A word on the revision in progress. The official announcement from ISO/TC 176/SC 2 states that the ballot on ISO/FDIS 9001 closed on 9 July 2026. However, as of 10 August 2026 ISO 9001:2026 has not yet been published. ISO itself has not stated a publication date, and it is certification bodies that are indicating an expected publication in September 2026. It should therefore be treated for now as expected rather than issued, and it is too early to build a transition plan on the assumption that it already exists. In any case, the requirement to control documented information is already in the current edition, so the revision is not a reason to wait.

In practical terms, put into your requirements that the revision mark and revision date are always displayed on the search results screen, that superseded revisions cannot be downloaded, and that opening a superseded revision redirects the user to the current one. On how to digitise document versions and approval flows, our article on electronic forms and the paperless factory covers the thinking on the forms side. Drawings and forms share the same problem where versions and approvals are concerned.

5 assumptions that break at Thai and ASEAN sites

A deployment procedure for drawing search AI written for domestic use in Japan cannot be applied as-is at a site in Thailand or elsewhere in ASEAN. Here are five places where the assumptions break.

The drawings are released from the head office in Japan. The Thai site is often the recipient of drawings, and the master copy sits on the Japanese side. In that case, building a search platform in Thailand does not give the Thai side control over updates to the master. Without a mechanism that tells the Thai side that a revision has been made in Japan, the Thai search platform quietly goes stale. The synchronisation mechanism has to be decided first.

The remaining volume of paper drawings cannot be read off. Paper is not only left on the shop floor; paper has also been handed to partner companies and outside processing subcontractors. Draw the boundary of the inventory exercise at the factory walls and paper from outside the company will surface later.

Title blocks are written in mixed languages. Japanese title blocks, English title blocks and Thai notes all coexist within the same drawing population. How far you widen the OCR target languages changes both cost and accuracy. In practice, a workable split is to restrict the title block to Japanese and English and to treat Thai as part of the notes.

Staff turnover is fast. The state where “that drawing is something only that person knows about” is lost faster than it is in Japan. The more a factory’s knowledge of where drawings are depends on individual memory, the more reason there is to hurry the L1 inventory. On how to retain knowledge tied to individuals, the points covered in our article on AI for skills transfer overlap with this.

The assumptions about network and cloud are different. In a configuration where the files stay on the head office server in Japan and are referenced from Thailand, handling large files such as drawings becomes slow. But keep a copy on the Thai side and the “multiple instances” problem that L1 was supposed to have solved comes straight back. Make it explicit at design time whether a given copy is a reference cache or the master.

Personal data and trade secrets inside drawings – 2 checks before moving them to the cloud

Before putting drawings on the cloud, there are two points to check – personal data, and trade secrets.

The first is personal data. A drawing itself is not normally personal data. The title block, however, contains the drafter’s name, the checker’s name and the approver’s name. Those identify individuals and can therefore constitute personal data. In Thailand, the Personal Data Protection Committee published notifications based on Sections 28 and 29 of the PDPA in December 2023, and they came into force on 24 March 2024. Cross-border transfers fall under that framework.

There is an important practical distinction here. A transfer whose sole purpose is storage and which no third party accesses is treated as being outside the scope of a regulated transfer. In other words, the issue is not “do we put it on the cloud” in itself, but “who can access it”. Is the data merely being stored in overseas-region storage, or will an overseas development vendor view the drawings in order to tune search accuracy? If it is the latter, the treatment changes. Make the accessing parties explicit in both the contract and the system configuration.

The second is trade secrets. Thailand has a trade secrets statute, the Trade Secrets Act B.E. 2545, enacted in 2002. That protection presupposes that the information is kept under management as a secret. Having access restrictions in place is what substantiates that secrecy management.

This collides head-on with a drawing search deployment. In the effort to make search convenient, there is a pull towards loosening permissions so that everyone in the company can retrieve anything. But leave drawings and know-how sitting in a shared folder that anyone can see and the claim that the information is managed as a secret becomes weak. Ease of finding and demonstrable secrecy management are in a trade-off relationship.

The realistic landing point is a two-stage design – allow the fact of a hit in search broadly, and restrict the right to open the drawing itself. Keeping a log of who opened which drawing and when also serves as material demonstrating that management is real.

Break the cost into 5 layers

Estimate cost the same way, by splitting it into layers. The layers here are different from the three layers of location, identification and content; these are five categories by type of spend. Receive a quotation for drawing search AI as a single figure and you lose sight of what you are paying for. The following is a guide assuming a Japanese-owned plant in Thailand holding on the order of 20,000 drawings.

That table needs to be read with care. There is no requirement to stack all five layers. A plant that already has a drawing management system may not need layer 4, and a plant whose CAD originals survive and which has no paper does not need layer 2. Conversely, you cannot skip layers 1 and 3 and buy layer 5 alone. Some layers can be skipped and some cannot.

The range for layer 4 includes the first year’s annual fee, but once the first 12 months are behind you, budget the annual running cost at 15 to 25 % of the total initial build cost across layers 1 through 5. That covers annual licence fees, adding and correcting attributes, ingesting new drawings, and re-tuning accuracy. Layer 5 in particular loses its effect if nobody touches it after go-live, because the character of the drawings shifts as the product mix changes.

What is most often overlooked is your own internal effort. Deciding the master copy at layer 1 and checking the matching results at layer 3 cannot be completed by an outside vendor. Somebody in the design department has to make the call. Push an approval through without putting that internal effort into the estimate and the project stalls partway because of the load on that person. The 2026 edition of the Japanese manufacturing white paper likewise lists shortages of knowledge and know-how, and difficulty securing people, among the issues in applying AI and digital technology. In a drawing search deployment, the place where that issue shows up is precisely this judgement step.

Note also that digital sector businesses in Thailand – software development, cloud services, data centres and the like – are among the activities eligible for BOI promotion. Whether your own investment qualifies depends on the nature of the business, so it has to be judged case by case after checking the primary source for the scheme.

How to measure the effect – putting “time spent searching” at the centre will fail

The first metric anyone proposes for deployment benefit is “reduction in the time spent looking for drawings”. Make that the primary metric, however, and in most cases it fails.

There are two reasons. First, the pre-deployment “time spent searching” was never measured. Ask about it afterwards in a survey and you get self-reported figures based on memory, which move depending on how the question is worded. Second, shortening search time is not in itself connected to a business number. You can build a formula that converts saved time into money, but no evidence remains that the time was actually redirected into some other value-adding activity.

What to put in its place is the concrete loss that occurred because something could not be found. For example – the number of cases where an equivalent part existed but a new design was produced anyway. The number of cases where a part was machined from a superseded revision. The number of days taken to respond when a customer asked for a drawing to be submitted. The number of times a subcontractor had to resend a drawing. These can be counted as incidents and they leave a record when they occur.

The other effective source is the search usage log. Who searched for what, and did they actually open a drawing from the results. A query that was searched but from which nothing was ever opened is a documented instance of not being able to find something. Reviewing those failed queries regularly is a reliable input into improving accuracy. The words the shop floor actually typed show what needs fixing far more clearly than the search conditions imagined in advance.

The thing to watch in metric design is to decide what you are measuring before deployment. Start thinking about “what shall we report as the benefit” only after go-live and the only numbers selected will be the ones that can be reported. That is not evaluation; it is retrospective justification.

6 patterns in which drawing search AI fails

Factories where deployment goes badly share common shapes. Here are six. None of them is a technology problem; all of them are problems of order and prerequisites.

Of these, the first pattern is the largest loss in money terms, because layer 3 has to be redone after an investment equivalent to layer 5 has already been made. Work that would have been done once if the order had been respected gets paid for twice.

A word more on the sixth. A drawing search platform loses value if it is left alone. If newly released drawings are not ingested, only historical material is searchable, and users conclude that the new material is not in there anyway and stop using it. Automating the ingest, and monitoring it, is as important as the initial build.

What to do in the first 90 days

Here is what a factory that has just started evaluating should do in the first 90 days. There is no need to select a product in that window. It is better positioned as the period in which you assemble the material needed to select one.

The counting work in the first 30 days is unglamorous but cannot be skipped. Take a proposal without those numbers and the quotation gets built on the vendor’s assumed sheet count, and is revised upward later.

The mapping table built in the middle stretch pays off well beyond the project. It is the yardstick for the evaluation, and it can be used unchanged for accuracy measurement after go-live. The trick when building it is to write down the searches that recently caused real trouble, not idealised searches. Ask a designer to “remember a drawing that was a pain to find last month” and concrete conditions come out.

Getting quotations broken down layer by layer in the final 30 days is the purpose of the whole 90 days. A quotation that arrives as a single total should be split by layer on request. A proposal that cannot be split, or whose author is reluctant to split it, may well contain uncertainty in the breakdown.

Frequently asked questions (FAQ)

Do we have to scan every paper drawing before drawing search AI can be used?

No. Only the drawings you want inside the search scope need to be digitised. Drawings for discontinued products and closed contracts can usually be excluded from the search scope even where a retention obligation applies. Deciding the scope first is an effective way of holding the cost down.

What level of accuracy does title block OCR reach?

It depends on the state of the drawings at each plant, so no general figure can be given. Where title block layouts are consistent and scan quality is uniform, the result is good. Where decades of formats are mixed together and the sheets are folded and faded, it falls. If you want to evaluate it, pull samples of each format from your own drawings and have them read for real.

We already have a drawing management system. Do we need drawing search AI on top of it?

If you can already pull by attribute, an addition is only needed where you want shape similarity search or natural language search. If the management system’s search is not being used, first separate whether the cause is missing attributes or the users’ workflow. If it is missing attributes, L2 work comes before layering AI on top.

The head office in Japan holds the master copies. Can we deploy at the Thai site alone?

The deployment itself is possible, but the mechanism for synchronising revisions has to be decided at the same time. Without synchronisation, the search platform on the Thai side ends up returning superseded revisions. In that case, making the search faster raises the probability of an incident in direct proportion.

Is there any problem with putting drawings on the cloud?

The issue lies less in the storage location itself than in who accesses it. A transfer whose sole purpose is storage and which no third party accesses is treated as falling outside Thailand’s cross-border transfer rules. If the configuration has an overseas vendor viewing drawings to carry out tuning work, the treatment changes. Fix the accessing parties in both the contract and the system settings.

We have nobody dedicated to this internally. Can it all be outsourced?

Scanning at layer 2 and platform build at layer 4 can be run by an outside vendor. But deciding the master copy at layer 1 and checking the matching results at layer 3 require internal judgement. Hand that over wholesale and the platform gets built on incorrectly designated master copies. The person does not have to be dedicated, but their time has to be secured.

Summary

Drawing search AI is not a tool that solves every form of the symptom “we cannot find our drawings”. What it solves is the topmost of the three layers, the layer of searching by content. In a plant without the consolidation of location and the assignment of identification information beneath it, the AI has nothing to search.

The order of evaluation is therefore fixed. First identify which layer your own “we cannot find it” is happening at. If it is L1, decide the master copies and consolidate. If it is L2, make the title block machine-readable and match it against master data. Only when those two are finished does L3 AI make sense as an investment.

Break the cost down along the same structure. Split it into the 5 layers of current state survey, scanning, attribute assignment, search platform and AI functions, and stack only the layers your organisation needs. Budget annual running cost at 15 to 25 % of the initial spend. And do not drop revision control. If the drawing you found is a superseded revision, making the search faster is the same thing as making the incident faster. Control of documented information under ISO 9001 already requires identification of revision status as of the current edition, ISO 9001:2015. ISO 9001:2026 was not yet published as of August 2026 and is at the stage of expected publication, but that is no reason to defer revision control.

At sites in Thailand and ASEAN, the assumptions are changed by the fact that drawings are released from the head office in Japan, that title block languages are mixed, and that staff turnover is fast. On top of that, the fact that names in the title block can constitute personal data, and the fact that leaving drawings in a shared folder weakens the demonstrable management of trade secrets, both have to be handled at the system design stage.

TOMAS TECH supports Japanese-owned factories in Thailand end to end, from the inventory of where drawings are, through attribute assignment, to the design of the search platform. It is entirely fine to come to us before you are at the product selection stage, when what you want is simply to work out which layer your own “we cannot find it” sits at. We are happy to take enquiries that go no further than organising the current state, so please get in touch through our contact page.

References

• ISO 9001 revision update (official announcement from ISO/TC 176/SC 2)
• White Paper on Manufacturing Industries 2026, published by METI, MHLW and MEXT
• Manual on trade secret management in Thailand, commissioned by METI and prepared by JETRO Bangkok
• Commentary on Thailand’s cross-border personal data transfer rules, by Tilleke and Gibbins
• Overview of the technology components of drawing search AI, by AI Market
• Outline of Thailand’s BOI investment promotion scheme, by JETRO