Document processing with AI is a pipeline: the system works out the type of file, reads the text, pulls out the fields it needs, cross-checks them and writes them into a CRM, an ERP or 1C. Nobody retypes the data; an employee confirms the values the system is unsure about. Accuracy here is measured on a sample of your own documents rather than in percentages from a brochure.
Key takeaways
The path is the same for a contract, an invoice and a form, and it has six steps, each of which can be checked on its own.
Which part of this a model takes on and which is covered by ordinary rules is set out in our article on business process automation with AI. What matters here is different: recognition only makes sense when the result lands in a system rather than in yet another folder.
It is worth recognising whatever arrives as an image or an unstructured file. A document that already exists in electronic form with fields is taken from the system that holds it — recognition there is an extra step and an extra source of errors.
| Document | Fields extracted | Where they go |
|---|---|---|
| Contract | Number, date, parties, subject, amount, term, bank details | The counterparty record and the deal, plus a reminder before the term ends |
| Supplier invoice | Supplier, number, date, line items, total, VAT, payment due date | A payment request, an expense record, a task for accounting |
| Act of completed work | Number, date, parties, amount, reference to the contract | Closing a deal stage and handing it to accounting |
| Paper form or application | Name, contacts, address, chosen terms, consents | A customer record in the CRM and a task for the manager |
| Electronic invoice (ESF) | No recognition needed: the fields are already filled in | Collected from the electronic document exchange operator's system |
That last row is a specific feature of Uzbekistan. Invoices go through the ESF information system, and the procedure for issuing, submitting and accepting them is set by Cabinet of Ministers resolution No. 489 of 14 August 2020, as the Tax Committee reminds businesses. Such documents are already structured, so they are taken from the operator's system, and recognition is kept for incoming scans, paper forms and contracts.
Accuracy is measured on your documents: a sample is taken, the extracted fields are compared with a reference, and you get the share of values filled in correctly with no edits. A general claim that "accuracy is high for standard documents" means nothing, because it differs between a printed PDF, a phone photo and a handwritten form.
What helps is that recognition services return not only the value of a field but a confidence score as well. In Google Document AI, for instance, every extracted entity comes with a number between 0 and 1. That is what the threshold is built on: above it the record moves on by itself, below it the document opens for an employee with the field highlighted.
The threshold is set by the cost of a mistake. For a form with contact details it is lowered, because an error costs one phone call. For the total on an invoice it is raised: a wrong figure goes off to be paid.
Documents that create money, obligations or rights are not posted without an employee's confirmation: contracts, invoices for payment, acts, HR orders, customer applications. The system prepares the record, a person confirms it.
That is more than caution. Law No. ZRU-1115 of 21 January 2026 prohibits relying exclusively on the conclusions of systems built on artificial intelligence technologies when taking legally significant decisions that affect a person's rights and freedoms. In document processing that comes down to a simple rule: a refusal, a charge or an obligation cannot arise from a single recognised field without an employee's confirmation.
Important A human check is a step in the system, not "having a look". The document has its own status, fields with low confidence are highlighted, and the confirmation is recorded: who, when and what they corrected. Otherwise, six months later nobody can reconstruct where the amount in the record came from.
A passport scan, a customer form and an HR document contain personal data, and handling them is governed by law No. ZRU-547 of 2 July 2019. For a recognition project that means the question of where the files sit and who opens them is settled alongside the question of how to pull the fields out of them.
Biometric data is a separate category: the law defines it as information describing a person's anatomical and physiological features. Databases holding biometric and genetic data, and the data of users of telecommunications operators, are kept on the territory of Uzbekistan under article 27-1 as amended by law No. ZRU-1125 of 26 March 2026. Other personal data may be processed abroad if the conditions of the same article are met.
Where careless storage of such files leads is covered in our article on the risks of a personal data leak.
Most of the problems are visible during preparation if you go through the list before the first launch.
If half of these are open, a pilot will show the readiness of the process rather than the quality of recognition.
One document type is rolled out at a time, starting with the one that arrives most often.
After launch the system does not get more accurate on its own. Accuracy grows because employees' corrections are collected in one place and used to adjust the set of fields, the validation rules and the wording of requests to the model. If documents come from an ERP or go into one, the sequence is built into the overall plan — the stages are covered in our article on rolling out an ERP system.
The price depends on the number of document types, the quality of the source files and where the data has to go. One type with a stable structure and decent scans costs noticeably less than five types mixed in with photos.
Starting prices and the scope of work are listed on our AI and automation page. On top of the one-off part there is a charge for requests to the model, which depends on how many pages pass through recognition each month. Where documents are processed — in the cloud or on your own server — is covered in our article on integrating AI into CRM and ERP.
Syntra Systems starts with your files: we collect examples of every type from recent months, run them through and show the share of fields recognised without edits, type by type. That figure is the basis for a decision, not a demo on somebody else's samples.
Mistake Starting with the hardest document — a handwritten form or a forty-page contract. Accuracy on it will be low, and the conclusion will be drawn about the whole project. The first type to take is the one with the biggest volume and a stable structure: an invoice, an act or a form.
From there we configure the set of fields and the validation rules, connect the write-back into your system and keep a route to a human with a log of confirmations. The retention period for files, the access rights to the folder of scans and the data that goes to an external service are agreed before launch.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
The system accepts phone photos, but accuracy on them is lower than on a printed PDF and is measured separately. The whole page needs to be in frame and the text needs to be legible. Fields the system isn't confident about aren't processed automatically — the document goes to an employee with that field highlighted.
Handwritten fields are recognised noticeably worse than printed ones, and legibility varies from form to form. Accuracy on such documents is measured separately and mandatory human confirmation is usually kept. If the flow of handwritten forms is large, changing the form itself is cheaper than fighting the recognition.
Contracts, invoices, acceptance certificates, HR orders and client applications: the system prepares the record, an employee confirms it, and the confirmation is logged. For everything else a confidence threshold is enough — disputed fields go to a human. Law No. ZRU-1115 also bars legally significant decisions about a person's rights from resting on an AI's output alone.
There is no technical limit, but you start with one or two: every type needs its own list of fields, its own validation rules and its own sample for measuring accuracy. Once the first type runs on the live stream without intervention, the next one by volume is added.
The quality of Uzbek text recognition differs between services, especially when Latin and Cyrillic script appear in the same document. This is checked on a pilot with your own files: if the share of fields needing no edits is low, another service is tried or that type stays with an employee.
The duplicate rule is set before launch. A document is usually identified by the pair of number and counterparty or by a checksum of the file: a repeated upload attaches to the existing record instead of creating a second one. Without such a rule the same invoice gets paid twice.
Recognition falls under routine automation, from $2,500; if you first need to work out what to automate, an AI audit costs $400. A monthly charge for model requests is added on top, and it grows with the number of pages that go through recognition.
Sources
