People selling AI rarely talk about limitations. But if you plan to use it for bookkeeping or system integration, what you need to know most is where it will fail. This article uses numbers from our actual tests.
Our receipt recognition project ThatReceipt used 76 real receipts for testing: accuracy for amounts was 94.7%, currency accuracy was 98.7%, and it took about 3 seconds per receipt.
The most useful aspect of these numbers is not how "high" they are, but that they are "measurable". Whenever anyone tells you their OCR is very accurate, you should ask: which documents were used for testing? How many? Which fields? Before we start work, we will do the same thing with your documents to give you numbers specific to your batch of documents, not a marketing number.
The extracted fields can be directly entered into your backend, eliminating the need for manual re-entry — this is where the time-saving comes in. Simply "recognizing the text" but still requiring manual copying and pasting doesn't save much time.
"Document recognition" sounds like a single feature, but in reality, each type of document requires separate calibration.
Our OCR demo currently supports three types of documents: marriage certificates, identity cards, and birth certificates, with proof of address still in development. Just for Hong Kong identity cards, it extracts 9 fields: Chinese name, English name, name code, date of birth, gender, symbol, first registration number, issue date, and identity card number.
Birth certificates and marriage certificates are even more complicated: the text order read by the recognition engine is unreliable, and the arrangement can differ each time the same document is read. Therefore, we don't rely on OCR but use spatial coordinates instead — that is, looking at the actual position of each word on the paper and then matching it to the corresponding field based on its position. This method is much more cumbersome but extracts much more reliably.
Watermarked fields cannot be extracted. Scanning that is too light, taken at an angle, with reflections, or with creases will affect the results. These are not due to poor software but are physical limitations of recognition itself — the machine cannot see clearly what the human eye cannot see clearly either.
Therefore, recognition will never be 100% accurate. Critical fields (amount, document number, date) should retain a manual review step, or set a confidence threshold, and if it falls below the threshold, switch to manual. When designing the process, we will directly tell you which fields can be safely automated and which need to be kept manual.
Also, don't use the demo results for direct bookkeeping. The demo is to show you the capability, not a system validated through your company's process.
This is the question we get asked the most.
In our OCR demo, the image you upload is only processed in memory: it is sent to Google Cloud Vision for text recognition, and the extracted fields are returned to your browser; we do not write to any database, bucket, or file. You can read the same statement written in more detail in our privacy policy.
The arrangement for formal production use is different: the system runs in your own cloud account, and the documents and extracted data remain in your own environment; we do not handle them. This is not just a privacy issue but also an ownership issue — you can revoke access at any time without losing the system.
By the way, the demo is free and doesn't require an email; you can try it as many times as you like (within reasonable use).
The cheapest way to start is at zero cost: take a real document of your own and go to the demo to see how it performs on your type of document. After trying, you will immediately know if this is useful for you.
For a formal system, AI document recognition starts from HK$32,000, including calibration for one document type, accuracy testing with your real documents, and integration with your backend. Each additional document type starts from HK$12,000. If your pain point is not just recognizing text but the entire process (such as receipt entry, reconciliation), you can check out automation modules, starting from HK$18,000.
We recommend starting with the most painful document type — usually the one with the highest monthly volume that is currently manually typed. Do one, verify one, save time, and then move on to the next.
Tell us in one email what you want to do online and we will come back with a written scope and a fixed quote.
Talk to us →