AI-OCR Comparison 2026 — What 99% Accuracy Means in Thailand

Line up AI-OCR product comparison tables and you see a row of similar numbers — 99.8%, 99.6%, 99.2%. Yet when the same products are tried in a factory in Thailand, those numbers usually do not reproduce. The cause is not the quality of the products but the way the comparison is framed. What an AI-OCR comparison should really line up is not product names but four things — which document, in which language, measured at which unit of accuracy, and at what cost. This article is written for people selecting a product to reduce back-office data entry at a Japanese-affiliated factory in Thailand or elsewhere in ASEAN. It lays out the published accuracy, price and market-share figures as they are, and then works through, one by one, how far those figures actually carry on the ground here.

If you are still at the earlier stage of asking whether to adopt AI-OCR at all and where to start, that broader picture is covered in our article on AI-OCR and back-office automation. This article is the next step — the one you use to decide which product to pick. If the adoption decision itself is still open, reading that one first and then coming back makes the selection criteria here easier to follow.

The First Thing to Look At in an AI-OCR Comparison Is Not the Product Name

Let me start with the conclusion. The first thing to settle in an AI-OCR comparison is not a product name but these four points.

First, which documents will be read. Supplier invoices? Incoming inspection reports? Handwritten daily work logs? Change the document and you change the functions you need, the accuracy that is achievable, and the number of billable fields. If you begin the comparison at the granularity of “all the paper in the company”, you will inevitably reach a point partway through where you can no longer decide.

Second, which language they are written in. Printed Japanese? Printed Thai? Thai handwriting produced by Thai staff? Alphanumerics only? The accuracy figures published in Japanese comparison articles assume Japanese and alphanumeric text.

Third, at which unit the accuracy was measured. Character level, field level, or document level? The same “99%” means very different things depending on the unit, and the human time required for checking changes several-fold. This is the core of this article.

Fourth, what it costs at your own volume and field count. AI-OCR pricing uses different billing units from product to product. Monthly flat fees, per-field metering and per-sheet metering are all mixed together, so reading the published prices side by side and sorting from cheapest to most expensive leads to the wrong decision.

Once these four are fixed, narrowing down products proceeds almost mechanically. Without them, staring at a product comparison table only feels like comparing — you are looking at the appearance of numbers. The rest of this article takes the four points in order.

What Is the 99% in 99% Accuracy — Character, Field and Document Level Change the Review Workload

This chapter is the heart of the article.

AI-OCR accuracy can be measured at three different units at minimum. And published product comparison articles very often do not state which unit they used.

What the floor actually wants to know is the third one, the document-level figure, because what determines an operator’s working hours is how many documents a person has to re-check. What products find easiest to publish, however, is the first one, character level. Character level makes the number look best, and it is also a natural performance indicator for an engine.

The problem is that the three cannot simply be converted into one another. 99% at character level is not 99% at document level. In fact, the more characters a single document contains, the further the two drift apart.

A Thought Experiment — How Far the Review Workload Moves When the Unit Changes

Let us put some concrete numbers down. All of the values below are assumed values, not measured results for any specific product. They are calculated on the simplifying assumption that each character error and each field error occurs independently of the others, as a thought experiment (real errors cluster in particular places such as amount columns and handwritten fields, so these are not predictions of reality). Even so, they are enough to show that mistaking the unit multiplies the review workload required.

Suppose one document carries 12 fields, with an average of 8 characters per field, for 96 characters in total. When 1,200 such documents are processed, the share that passes without a single character or field error looks like this for each unit of accuracy.

For the same “99.0%”, field level means fixing a little over one document in ten, while character level works out to some correction on around six documents in ten. That is a difference of more than fivefold in the number of documents to fix. Compare the smallest value in the table (2.4% at 99.8% field level) with the largest (61.9% at 99.0% character level) and the gap exceeds 25 times. The displayed numbers may look similar, but the workload on the floor moves this much. That is why you should not select a product without confirming the unit.

The gap also widens on documents with many fields. At the same 99.0% field level, extracting 12 fields gives a no-correction rate of approximately 88.6%, but a document such as a line-item-heavy invoice from which 30 fields are extracted drops to approximately 74.0%. A situation where “the accuracy is the same, yet complaints started once we switched documents” arises quite naturally from this structure.

To repeat, this is an explanation of structure using assumed values. What the 99.8% or 99.6% published by real products would correspond to at field level is something nobody can state, because the sources do not say. If you find an article claiming that the figure “actually drops to a certain percentage at field level”, check the basis for that conversion. At least in the sources this article referenced, no such basis exists.

Three Questions to Ask in the Sales Meeting

What to confirm with a sales representative during evaluation follows automatically from this structure.

First, “what is the denominator behind that accuracy figure?” Characters, fields, or documents? If they cannot answer immediately, the measurement conditions may not be shared internally at the vendor.

Second, “may we see samples of the documents used for that measurement?” Whether the figure came from clean printed forms or from real working paper with smudges and stamps changes what it means.

Third, “would you run the same measurement on our own documents?” That question is the design of a PoC in itself. The metric to watch in a PoC is not whether the text could be read, but how many documents were left requiring visual review by a person.

Published Accuracy, Price and Japanese Market Share for the Major AI-OCR Products

From here, the published figures are presented as they are. Having set them out, we will then confirm what this way of presenting them cannot settle.

The Published Accuracy Figures

The published accuracy figures for the major products circulating in Japan are as follows. The source is the AI-OCR comparison article published by OptiMax.

Three things should be stated honestly about this table.

First. The source does not state the unit of measurement for any of the products. As the previous chapter showed, character level, field level and document level mean very different things, and here we do not know which applies. Consequently, how much practical difference there is between 99.8% and 99.2% cannot be calculated from this table.

Second. Tegaki’s “approx. 99%” is somewhat different in character from the rest. The source explicitly attaches the conditions “fully specialised in handwriting” and “as a recognition rate in the financial sector”. In other words, it is a figure under the limited conditions of handwritten documents in finance. Whether the same figure would appear when the product is used on factory daily work logs cannot be determined from the source. One could equally say that stating the conditions makes it the more honest presentation.

Third. ABBYY Vantage and LINE WORKS OCR are not disclosed. Non-disclosure does not mean low performance. If anything, the vendors most aware that numbers shift with measurement conditions tend to be the ones least willing to put out a single figure. Treating the two undisclosed products as “excluded because there is no number” is a sloppy way to compare.

The Price List, and Why It Must Not Be Sorted from Cheapest

Next, AI-OCR pricing, quoted from the same source.

Scanning this table top to bottom and reading it as “eas is the cheapest” or “SmartRead is cheap because it is JPY 25” is a mistake, because the billing unit differs by product. Monthly flat fees, metering against the number of fields, and an annual fee shown alongside a per-sheet rate are all mixed together. Numbers whose units are not aligned cannot be ranked against one another.

Let us look at this concretely. SmartRead’s “JPY 25 per 10 fields” means JPY 2.5 per field (25 divided by 10 equals 2.5). Extracting 12 fields from one invoice gives JPY 30 per sheet, and at 1,200 sheets a month that is JPY 36,000 per month. The “JPY 25” display, which looks small as a unit price, in practice turns into spending in the JPY 30,000-per-month range. The smallness of the displayed number and the size of the amount paid do not correspond.

Tegaki’s “from JPY 360,000 per year (approx. JPY 20 per sheet)” also reveals its premise when the two numbers are set against each other. Dividing 360,000 by 20 gives 18,000 sheets, or 1,500 sheets a month. That is, the “approx. JPY 20 per sheet” unit price corresponds to the effective rate when roughly 18,000 sheets a year are processed. This premise is not stated explicitly in the source. Please read it as the result of simply dividing the two figures presented together. Put the other way round, a site processing only 1,200 sheets a month (14,400 a year) divides JPY 360,000 by 14,400 sheets and arrives at JPY 25 per sheet — higher than the published “approx. JPY 20 per sheet”.

One more thing. DX Suite and LINE WORKS OCR are annotated “includes metered charges”. This means the displayed JPY 30,000 and JPY 55,000 per month are a floor, not a total. As your volume grows, metered charges are added on top of that amount. The actual figure for these two products is therefore not fixed until you give the vendor your processing volume and obtain a quotation.

How It Looks Once Converted to Your Own Volume (Assumed Values)

The only way to align the billing units is to fix your own processing volume first. Here, as an assumed value, we place “1,200 sheets per month, an average of 12 fields per sheet, so 14,400 fields per month” and mechanically divide the published floor amounts alone.

Let me be clear about how to use this table. It is not a ranking from cheapest to most expensive. The row order is the same as the price table above and is not sorted by amount. With metered charges undetermined for two products and a volume cap unknown for one, this column cannot settle which product is better. All the table shows is that applying the same processing volume produces a picture different from the appearance of the list prices. SmartRead, displayed as “JPY 25 per 10 fields” and looking small as a unit price, converts to JPY 30 per sheet and lands second-highest among these five.

Note that this article does not build a table converting the yen-denominated prices into baht, because the source referenced carries no exchange-rate premise. If you set a conversion rate yourself, every number in the article becomes false the moment the rate moves. When putting an approval through at a Thai site, either ask the vendor for a baht-denominated quotation at the point of obtaining the estimate, or convert while stating your internal period rate explicitly.

How to Read the Japanese Market Share Figures

Share figures are available too. They come from a survey conducted by BOXIL in 2025 (n equals 1,588, targeting people responsible for AI-OCR adoption), quoted in the OptiMax article above.

These four products add up to 68.22%. They do not reach 100%. The remaining 31.78% is held by products that do not appear in this table. So the reading that “the market is consolidated among the top four” does not hold, and a little over 30% of it goes to other options. There is a real possibility that the product which happens to fit your documents sits inside that 31.78%.

And one more point, which matters more for a Thai site. This is a survey conducted in Japan. The respondents are people who adopted AI-OCR within Japan, and this is not Thai market share. A strong adoption record in Japan means a track record built on Japanese-language documents. It does not mean a track record on Thai-language documents. Judging that “it is the number one in Japan, so it is safe” may be reasonable at a site inside Japan, but it does not transfer as it stands to a site in Bangkok or Chonburi.

Four Japanese Assumptions That Break at a Thai Site

What happens when selection criteria built from Japanese comparison articles are carried into Thailand unchanged? Four points break.

Of these four, the two that generally do not appear in comparison articles are the last two. Anyone can think of language and document format, but regulation and operating structure stay out of view as long as you are reading Japanese articles.

A little more concretely on operating structure. The work that remains after AI-OCR is adopted is checking and correcting the recognition results. In most cases the people doing that work are Thai accounting and purchasing staff. If the admin screen is Japanese only, the only person able to do the checking is the Japanese expatriate, and that one person becomes the bottleneck. If the admin screen supports English, on the other hand, the work can be left to local staff. If you select a product with a Japanese-only UI, you can end up with a result where the data entry shrank but the expatriate’s checking workload grew.

Support structure works the same way. Whether you can raise an enquiry within business hours in Thai time, or only through a Japanese desk, changes how long you are stopped when something goes wrong. This information is not written in the accuracy column of a product catalogue, so the only way to get it is to ask during the sales meeting.

Where Thai and Vietnamese Document OCR Actually Stands Today

Plenty of products answer “yes, it is supported” to the question “can it read Thai?” What needs examining is not whether it is supported but at what level it reads.

What the Thai Benchmark Shows

ThaiOCRBench is a benchmark specialised in Thai document understanding. It consists of 2,808 human-verified samples across 13 tasks. What deserves attention is the fact itself that before it, no comprehensive benchmark specialised in Thai document understanding existed. In other words, a common yardstick for measuring Thai document recognition performance side by side has only just come into being.

The scores on this benchmark’s handwritten text recognition task are as follows.

These values need careful handling. They are normalised scores, not accuracy percentages. Placing them on the same footing as the 99.8% and 99.2% from the published-accuracy chapter and rereading them as “71.4% accuracy” is wrong. The definitions of the metrics are entirely different.

So what can be said? Three things. First, that even the top model sits far from a perfect score of 1.0. Second, that the gap between top and bottom is large (the difference between 0.714 and 0.393 is 0.321), so model choice still strongly determines the outcome at this stage. Third, that printed Japanese and handwritten Thai have not reached a maturity where they can be discussed on the same yardstick in the first place.

The “99.8%” in a Japanese product comparison table and the “0.714” in this benchmark cannot be compared side by side. The very fact that they cannot be placed side by side is what you need to know when running an AI-OCR comparison at a Thai site.

Why Thai Is Difficult

ThaiOCRBench cites, as factors making Thai document recognition difficult, that it is a non-Latin script, that there are no explicit separators (spaces) between words, and that real-world documents are highly unstructured.

Of these, the one that matters for document OCR is the second. Even if the characters can be read one by one, the invisibility of word boundaries starts to hurt at the stage of dividing them into fields such as item name, quantity and unit price. Japanese does not use spaces between words either, but the switching between kanji and kana works as a de facto separator. Thai has few such cues, and segmenting a character string into meaningful units becomes a technical problem in its own right.

This means that on Thai-language documents, the gap between “characters read” and “fields correctly extracted” may be larger than on Japanese-language documents. The discussion of the unit of accuracy from the opening of this article comes back into play here. Even where character-level accuracy is high, field-level accuracy does not necessarily follow it.

Thai-Specialised OCR Keeps Improving

Work specialised in Thai is progressing too. Typhoon OCR, published by SCB 10X, reports improved handwritten form recognition in version 1.5.

Here too the nature of the metrics needs care. BLEU and ROUGE-L were originally created to evaluate machine translation and summarisation, and they are not a correctness rate for extracting fields from a document. You cannot read them as “BLEU 0.522, therefore 52.2% is readable”.

What these numbers do say is that Thai handwriting recognition is an area in the middle of improving. Turned around, that also means that if you lock in a product on today’s evaluation, the premises may have changed a year from now. When choosing a product with a long contract term, this point is worth factoring in.

What Can and Cannot Be Said About Vietnamese

An honest note for those whose scope includes a Vietnamese site. Among the sources this article was able to reference, there is no verified benchmark for Vietnamese document OCR equivalent to ThaiOCRBench for Thai. This article therefore cannot produce figures on Vietnamese recognition performance.

What can be said goes only this far — it is a writing system in which Latin characters carry diacritical marks (including tone marks), and it presents recognition challenges different again from Japanese or Thai. If you intend to use it at a Vietnamese site, the only route is to hand the vendor your real documents and measure through a PoC. Please do not decide on the strength of the phrase “Southeast Asia supported” alone.

For how to think about product selection when using generative AI or multimodal models in a multilingual working environment, we have organised that separately in our comparison of enterprise generative AI. If you are weighing the option of reading documents with a general-purpose multimodal model rather than a product sold as AI-OCR, please look at that alongside this one.

Regulatory Differences — Japan’s Electronic Books Preservation Act and Thailand’s e-Tax Invoice

As a criterion in an AI-OCR comparison, regulatory compliance is easy to overlook relative to how much it affects the amount you pay. This is where the premises differ completely between Japan and Thailand.

The “Electronic Books Preservation Act Compliance” Baked into Japanese Selection Criteria

Within Japan, from 2024 onward the requirements for storing electronic transaction data became in effect mandatory, and compliance with the Electronic Books Preservation Act was added to the selection criteria for AI-OCR products. The standard view in Japan is that if AI-OCR is used on invoices and receipts, points such as timestamp functionality, support for the search requirements, and whether qualified-invoice fields can be extracted need to be confirmed.

That background is why Japanese comparison articles carry a column marked “Electronic Books Preservation Act compliant”. And it is not unusual for that column to be reflected in the product price.

This is a Japanese domestic regulation, however, and it does not apply to a Thai legal entity. It typically reaches a Thai site only when a Japanese parent company brings its own selection criteria with it.

Thailand’s e-Tax Invoice and e-Receipt Are Voluntary

Now the Thai side. Responsibility is split between two bodies — the Revenue Department handles taxation, and ETDA handles the technical standards.

And here is the most important point. For both 2026 and 2027, mandatory B2B electronic invoicing has not been legislated in Thailand. It is voluntary. If you come across an explanation that “Thailand has made it mandatory too”, check the source. At least in the information as of July 2026 that this article referenced, it remains voluntary.

If you do join the scheme, there are two routes.

What to watch in this table is the standing of the THB 30 million threshold. It is the condition for being able to use the simplified route, not a threshold for whether to adopt AI-OCR. A manufacturing site with annual revenue above THB 30 million cannot use the simplified route, but that is a separate matter from AI-OCR selection.

While participation is voluntary, tax incentives are available to those who join. They are a 200% deduction for qualifying investment, and a reduction of the e-Withholding tax rate to 1%. In June 2026 the cabinet approved an extension of these incentives to the end of 2027. At the point of approval, however, the royal decree and ministerial regulations were still pending. Please confirm the latest gazette status with your tax adviser before relying on the incentive. This part of the article is the most likely to change over time.

What the Regulatory Difference Concretely Means for Selection

Taking all of the above together, the problem with carrying a Japanese parent company’s selection criteria into Thailand unchanged becomes clear.

You end up paying for features you do not need. Timestamp functionality for the Electronic Books Preservation Act, field extraction for Japan’s qualified invoice system, support for Japan’s search requirements. None of these are used in the document processing of a Thai entity. If you pick a higher plan that bundles them as standard features, that portion of the cost is wasted.

At the same time, you overlook the features you do need. Field extraction accuracy in Thai, an English admin screen, whether integration is possible with an accounting system that handles XML under ETDA standard 3-2560. There are no columns for these in Japanese comparison tables. What has no column cannot be compared.

So when running an AI-OCR comparison at a Thai site, you need to add your own columns to the Japanese comparison table. The columns to add are at least these four — whether Thai-language documents can have fields extracted, whether the admin screen supports English, whether there is a local support desk in Thailand, and the integration method with your existing accounting or ERP system. On how to design integration with an ERP and its surrounding systems, we have organised the cost structure in our article on business system development and ERP integration.

Judging AI-OCR Fit by Document Type

An approach of “all the paper in the company, through AI-OCR” almost always stalls partway. Difficulty and payoff differ by document type. Here are the points to judge on for five types commonly targeted at Japanese-affiliated factories.

Of these five, the one where AI-OCR most readily returns on investment is the invoice. Volume is high, formats differ by trading partner so human pattern learning does not help, and entry errors lead directly to payment errors. Starting with invoice processing automation is a reasonable judgement at many sites.

Handwritten daily work logs, on the other hand, are the hardest area. Here the difficult conditions stack up, starting with handwritten Thai, where — as the benchmark chapter showed — even the top model sits far from a perfect score. Taking this on from the start means cost is added while the review workload does not fall. If you are going to tackle handwriting OCR, the first step is to revisit the form design itself — ruling boxes for entries, replacing free text with selections, separating numeric fields from text fields — as preparatory work.

On incoming inspection reports, a short addition from the angle of automating goods receipt work. What characterises this document is that there is room for you to specify the format. If you can design the format handed to suppliers yourself, you can shift it toward something AI-OCR finds easy to read. Rule the fields with lines, fix the entry positions, standardise the digit count in numeric fields. That alone changes the recognition results. It is worth examining whether the document design can be changed before selecting an AI-OCR product.

Documents you issue yourself, such as purchase orders, should not be made OCR targets at all. When the data is already in hand from your own system, turning it into paper and reading it back is a round trip. This angle comes up again in the later chapter on the option of not choosing AI-OCR.

How to Convert AI-OCR Pricing to Your Own Situation

Now we translate all of the above into the actual work of getting a quotation. The order matters.

Step 1 — Fix Your Processing Volume First

Before looking at products, fix your own numbers. Three are needed — the monthly volume of the target documents, the number of fields extracted per document, and the time that entry currently takes. Obtain a quotation without these three and you get back amounts calculated on different premises by each vendor, and comparison becomes impossible.

There is a caveat in how you count sheets. If one transaction comes with one invoice and two line-item sheets, that is three sheets. With per-field metered products the number of fields in line items feeds straight into the cost, so please also count the average number of line-item rows.

Step 2 — Write the Shared Premises into a Single Table

Since numbers are involved, gathering the premises in one place and stating them explicitly is essential. Below is an example of the assumed values placed for the purposes of this article. Please replace them with your own numbers.

Step 3 — Fix the Baseline to a Single One

This is where accidents happen most easily. When converting the effect of AI-OCR into money, the following two must not be added together.

One is the effect that “the working hours of in-house data-entry staff go down”. The other is the effect that “the outsourcing fees paid to an external data-entry provider go down”. For the same entry work on the same documents, it is the former if you do it in-house and the latter if you outsource it, and both cannot hold at once. An estimate that adds both is booking an effect that does not exist.

In this article the baseline is fixed to operator working hours only. Reductions in outsourcing fees are not counted.

Step 4 — Express the Effect in Hours

With the premises in place, calculate the hours saved. Everything below is an estimate based on assumed values.

With the same product and the same volume, the reduction rate splits between 70.0% and 37.5%. What produces this gap is the unit of accuracy from the second chapter and the language question from the fifth. Documents with handwritten Thai mixed in tend toward the lower row, while standardised invoices centred on alphanumerics tend toward the upper. Which way yours falls is not knowable until you run a PoC on your own documents. If the estimate a vendor presents assumes only the upper row, ask them to produce the lower case as well.

Step 5 — Convert to Money Using Your Own Rate

This article does not produce a payback period converting the saved hours into money. There are two reasons.

First, the source for product pricing is denominated in Japanese yen, while personnel costs at a Thai site are denominated in baht. Since the source referenced carries no exchange-rate premise, setting a conversion rate ourselves would put the numbers beyond this article’s responsibility from that moment on.

Second, whether the saved time actually disappears as cost depends on how the site is run. Even if 56 hours a month are freed up, personnel cost does not change while that member of staff remains on the payroll. Whether to redirect the freed time to other value-adding work, to cut overtime, or to forgo an additional hire — deciding which form the return takes is a management judgement, not a formula.

So please carry out the conversion to money yourself, stating your own labour rate (baht per hour) explicitly. On how to design effect measurement for AI adoption and what to place as indicators, we have organised a framework in our article on measuring AI effect and ROI. Reading it before converting to money helps you arrange things in a form that is easy to explain internally.

The Option of Not Choosing AI-OCR — Capturing Data Digitally from the Start

It may look odd to write this in an article about product comparison, but there is something to check at the very start of the selection work. Does that document need to be paper at all?

AI-OCR is a technology for turning information written on paper back into data. Put the other way round, if the data can be entered digitally from the start, the recognition step itself is unnecessary. No worry about recognition accuracy, no review workload, no per-field metering — none of it arises.

For documents that fall in the right-hand column — those originating on your own floor — electronic forms are often the more natural answer than AI-OCR. Having a daily work log written by hand in Thai, reading it with AI-OCR, and having a person fix the recognition errors is a flow that disappears entirely if entry is done directly on a tablet. On how to switch shop-floor documents to digital entry, we have set this out concretely in our article on paperless factories with electronic form systems.

For the left-hand column, on the other hand — documents received from outside — you cannot decide the counterparty’s format, so this is where AI-OCR comes in. In practice, most sites settle into using the two together. Sorting the target documents into left and right before starting the AI-OCR comparison work reduces the number of products to compare and speeds up the decision.

Selection Checklist for AI-OCR Adoption

Here are the items to confirm in sales meetings and internal review, arranged along the flow of this article. There are ten.

Of these ten, numbers 5 and 6 matter most. As we have seen, the published accuracy figures come without a stated unit of measurement, and the benchmarks available for Thai-language documents are measured on indicators different from those in Japanese product comparison tables. In the end, there is no way to obtain a figure that applies to you other than measuring on your own documents.

A note on PoC design as well. The common failure is handing over ten clean sample documents and finishing at “it read them”. What should be handed over is the real thing from the last month. Smudged ones, ones with overlapping stamps, ones with handwritten corrections, ones that arrived by fax at reduced resolution. Hand over 100 or so, including these, and count the number of documents a person had to check visually. That is a PoC.

Frequently Asked Questions

What is the 99% in AI-OCR’s 99% accuracy

Because the source does not state the unit of measurement, it cannot be determined from the published figures alone. Three are possible — character level (what percentage of the characters read are correct), field level (what percentage of the extracted fields are correct), and document level (what percentage of processed documents pass without correction). The one that carries meaning in practice is document level, but the one most readily published is character level, where the number looks best. Ask in the sales meeting, “what is the denominator behind that 99%?” A vendor who can answer immediately is also evidence that measurement conditions are shared internally.

How should AI-OCR pricing be compared

Fix your monthly volume and fields per sheet first, then apply each product’s billing unit to convert. You cannot line up the published prices and read them from cheapest, because monthly flat fees, per-field metering and per-sheet metering are mixed together. Also, some products with a monthly display carry the note “includes metered charges”, and in that case the displayed amount is a floor, not a total. Give the vendor your processing volume and obtain a quotation for the total.

Is choosing a product with high market share in Japan a safe bet

A track record in Japan is a track record on Japanese-language documents. The survey referenced targeted 1,588 people responsible for AI-OCR adoption within Japan, and it is not Thai market share. In addition, the top four products total 68.22%, and the remaining 31.78% is held by other products. There is a real possibility that a product outside the top group fits your documents.

Can handwritten Thai documents be read with AI-OCR

We are not yet at a stage where this can be answered as a binary of readable or unreadable. On the benchmark for Thai documents, the handwritten text recognition score is 0.714 even for the top model, a distance from a perfect score (this is a normalised score, not an accuracy percentage). Thai-specialised OCR also keeps improving, but the metrics reported derive from machine translation and differ from a correctness rate for field extraction from documents. In practice, for documents such as handwritten daily logs, revisiting the form design or switching to digital entry is a surer path than reading them with OCR.

Is compliance with Japan’s Electronic Books Preservation Act needed when adopting document OCR

Japan’s Electronic Books Preservation Act does not apply to a Thai legal entity. Within Japan, from 2024 onward the requirements for storing electronic transaction data became in effect mandatory and compliance was added to AI-OCR selection criteria, but that is a Japanese domestic regulation. Whether you need to satisfy Japanese requirements for the sake of reporting to a Japanese parent company is something to confirm with your accounting department. For work confined to the Thai site, there is no need to pay for compliance features under that Act.

Has electronic invoicing been made mandatory in Thailand

It has not. For both 2026 and 2027, mandatory B2B electronic invoicing has not been legislated, and it is voluntary. If you join the scheme, there is a full route using XML under ETDA standard 3-2560 with an electronic signature by NRCA certificate, and a simplified route called e-Tax Invoice by Email for small businesses with annual revenue of THB 30 million or less. Incentives exist in the form of a 200% deduction for qualifying investment and a reduction of the e-Withholding tax rate to 1%, and in June 2026 the cabinet approved an extension to the end of 2027, though at that point the royal decree and ministerial regulations were still pending. Confirm the latest status with your tax adviser before relying on it.

Where should invoice processing automation start

Invoices received are an area where AI-OCR readily returns on investment, because volume is high, formats differ by trading partner, and entry errors lead directly to payment errors. As an order of attack, first measure the volume of the last three months and the number of fields extracted per sheet, then run a PoC on 100 or so real documents and measure the number needing visual review. Narrowing down products once those two are in hand is the shortest path.

Is AI-OCR suited to automating goods receipt inspection work

Incoming inspection reports differ in character from other received documents in that there is room for you to specify the format handed to suppliers. If you can work on the form design — ruling fields with lines, fixing entry positions, standardising the digit count in numeric fields — the recognition results change substantially. Before starting a comparison of AI-OCR products, first examine whether the document design can be changed. If you can decide the format entirely yourself, having suppliers enter the data through a web form is sometimes the surer option to begin with.

How should the effect of AI-OCR adoption be measured

Measure it in hours saved. Compare the pre-adoption figure of entry time per sheet multiplied by monthly volume with the post-adoption figure of checking and correction time per sheet multiplied by monthly volume. When doing so, fix the baseline for the effect to a single one. A reduction in in-house operator hours and a reduction in outsourced entry fees cannot both hold for the same work at the same time. An estimate that adds both is booking an effect that does not exist. Carry out the conversion to money with your own labour rate stated explicitly.

Summary — Compare the Combination of Four Factors, Not Product Names

This article makes one claim. The “99% accuracy” lined up in Japanese product comparison tables does not reproduce as it stands on the documents of a Thai site. What should be compared is not product names but four things — which document, in which language, measured at which unit of accuracy, and at what cost.

On accuracy, the sources do not state whether the published figures were measured at character, field or document level. Change the unit and the number of documents a person has to check visually changes several-fold.

On price, monthly flat fees, per-field metering and per-sheet metering are mixed together, so you cannot line up the published prices and read them from cheapest. There is no method of comparison other than fixing your own monthly volume and field count first and then converting.

On market share, the figures available come from a survey conducted in Japan, and they are not Thai market share. In addition the top four products total 68.22%, so a little over 30% of the market uses other products.

On language, the comprehensive benchmark for Thai documents has only just come into being, and handwriting recognition sits far from a perfect score even for the top model. The 99.8% for printed Japanese and the 0.714 for handwritten Thai are not figures that can be compared side by side in the first place.

On regulation, Japan’s Electronic Books Preservation Act does not apply to a Thai entity, and Thailand’s e-Tax Invoice is voluntary. Carry Japanese selection criteria over unchanged and you pay for features you do not need while overlooking the ones you do.

And the conclusion that remains once all of this is taken into account is that there is no way to obtain a figure that applies to you other than running a PoC on your own real documents and counting the number needing visual review. Catalogue figures are a starting point, not an answer.

Selecting an AI-OCR product is not the work of staring at a comparison table. It is the work of classifying your own documents, counting your processing volume, and trying the real thing. TOMAS TECH is based in Bangkok and builds back-office and shop-floor business systems for Japanese-affiliated factories in Thailand and across ASEAN, and we take enquiries from the stage before a product is chosen, including document classification and taking stock of processing volume. Whether you are still at the stage before drawing up a shortlist, or have a single question such as whether a particular document suits OCR or should be switched to digital entry, please get in touch through our contact form.

References

• AI-OCR Comparison of 8 Products, 2026 Edition (OptiMax, in Japanese) — published accuracy, pricing and market share for the major products
• AI OCR Comparison of 41 Products (Digital-ka no Madoguchi, in Japanese) — points on Electronic Books Preservation Act compliance
• ThaiOCRBench — A Task-Diverse Benchmark for Vision-Language Understanding in Thai — benchmark for Thai document understanding
• Typhoon OCR 1.5 Release Notes (SCB 10X) — improvements in Thai-specialised OCR
• E-Invoicing Remains Voluntary — 2026/2027 Updates and Tax Incentives (VATupdate) — Thailand’s e-Tax Invoice scheme and tax incentives
• Electronic Invoicing in Thailand (EDICOM) — overview of Thailand’s electronic invoicing scheme