Artificial Intelligence in a Document Management System: from archiving documents to understanding information

How AI turns a DMS from an electronic archive into a platform that understands documents: OCR, classification, metadata, semantic search, chat, workflow and human-in-the-loop.

Document digitisation has gone through several stages. At first, the main goal was simple: eliminate as much paper as possible and keep documents in an IT system. Scanning, electronic archives and the first Document Management Systems — DMS — appeared.

Later, organisations realised that electronic storage alone does not solve the whole problem. A directory with tens of thousands or millions of files remains hard to manage if every document must be classified manually, if information from it must be typed in by an operator, and if finding something requires knowing the exact file name or location.

Today a third stage begins: the intelligent DMS.

Artificial Intelligence turns the document management system from an electronic warehouse into a platform that can, to some extent, understand document content, extract information, identify relationships and take an active part in an organisation’s processes.

A modern DMS must no longer answer only “Where is the document?”, but also questions such as:

• “What does the document contain?”
• “Which other documents is it related to?”
• “When does this contract expire?”
• “Which invoices are unpaid?”
• “Which of these documents should go to Legal?”
• “What are the annexes of document number 1542?”
• “Does the file contain all required documents?”

This shift is fundamental: from the file as something to store, to information as a resource the organisation can use.

From file to information

In a classic DMS, the document is primarily a file.

It may be a PDF, a Word document, a PowerPoint presentation, an Excel file, a scanned image or an email. The system stores it, assigns metadata and lets the user find it later.

Artificial Intelligence changes the perspective.

For AI, the document becomes first and foremost a source of information.

Suppose a PDF invoice is uploaded into a DMS.

In a traditional system, the operator may have to fill in manually: supplier, invoice number, date, amount, VAT, payment term and the responsible department.

In an AI-assisted system, the process can look different.

The document is uploaded, text is extracted automatically, the system identifies that it is an invoice and extracts the main fields. The operator no longer types every value — they verify what the system found.

The difference seems small for one document.

It becomes enormous when you process tens or hundreds of thousands of documents a year.

OCR is the foundation, but it is not enough

OCR — Optical Character Recognition — is one of the foundational technologies of a DMS.

A scanned document is, from the computer’s point of view, initially just an image. OCR turns that image into searchable, processable text.

But OCR must not be confused with Artificial Intelligence.

OCR can identify the text:

“Contract no. 254 of 12.09.2026”.

AI can interpret its meaning:

• contract number = 254
• contract date = 12 September 2026
• document type = contract

Moreover, it can analyse the whole document and identify the contracting parties, the contractual period, the value, obligations, termination conditions or important deadlines.

In practice, OCR gives the system the ability to “read”, and Artificial Intelligence gives it the ability to interpret what it reads.

Automatic document classification

One of the most important AI functions in a DMS is automatic classification.

A company may have invoices, contracts, addenda, requests, official correspondence, minutes, certificates, offers, HR documents, fiscal documents, reports, presentations and many other categories.

Manual classification consumes time and inevitably produces errors.

AI can analyse the document right after upload and propose:

• Document type: Contract
• Document type: Supplier invoice
• Document type: Addendum

Based on that classification, the DMS can apply different rules automatically.

An invoice can be sent to Accounting. A contract can be routed to Legal. A CV can enter an HR flow. A complaint can go to Customer Relations.

So Artificial Intelligence does not classify a document only to organise it more neatly. Classification can trigger the entire organisational process.

AI and electronic registry

A particularly interesting case is integrating Artificial Intelligence with the registry.

Suppose an institution receives an official document.

The registration may be: 1542 / 08.10.2026

The main document may be a PDF letter, and three annexes may be added later: an Excel file, a technical PDF and a PPTX presentation.

The same file may also contain supporting documents that are not registered separately: a copy of a certificate, a proof, an internal note or a working document.

A well-designed DMS must understand these differences.

AI can help identify them. When a file is uploaded, the system might suggest:

• “This document appears to be an annex to registration 1542/2026.”
• “Supporting document — no registry number required.”
• “The document contains reference 1542/2026. Do you want to link it to the existing registration?”

That is an important difference between a simple file system and a true intelligent document management system.

Automatic metadata extraction

Metadata is information about the document.

For an invoice it may be the supplier, tax ID, invoice number, date and amount. For a contract it may be the contractual partner, number, signing date, value, period and expiry date. For official correspondence it may be sender, recipient, number and subject.

AI can extract this information directly from the content.

Modern document management platforms are already moving in this direction. Box, for example, uses AI to identify and extract metadata from content and to integrate the resulting information into organisational flows.

Mayan EDMS version 4.12 offers integration with both OpenAI and Ollama, enabling structured data extraction from documents and sending results to the workflow engine.

This is a strong example of how AI can be integrated into a DMS without abandoning the traditional document management system. AI becomes an additional component of it.

Semantic search changes how we find documents

Traditional search relies largely on exact words.

If the user searches for “server maintenance contract”, the system tries to find documents that contain those words.

Semantic search works differently. It tries to understand the meaning of the question.

The user might write:

• “Show me IT contracts that expire in the next three months.”
• “Find documents about company server maintenance.”
• “Which contract makes the supplier responsible for backup?”

The document does not necessarily need to contain the exact phrase the user typed. AI looks for semantically related information.

In Mayan EDMS, for example, the current AI integration explicitly includes semantic search, alongside summarisation and structured information extraction.

This technology can radically change the user experience. Instead of learning the archive’s complicated structure, the user can simply ask the system.

Chat with the company’s documents

From here comes the next natural step: conversation with the archive.

Instead of opening ten contracts and reading each one, a director can ask:

• “Which contracts have a value above 100,000 lei?”
• “Which contracts expire by the end of the year?”
• “What are our obligations towards company X?”
• “What penalties are provided in contract Y?”
• “Summarise this documentation in five paragraphs.”

M-Files Aino, for example, lets users generate summaries and ask questions about document content in the system. M-Files documentation notes that Aino uses content available in the vault and can work with textual documents and with scanned documents after they are turned into searchable PDFs through OCR.

This approach makes the direction of DMS evolution very clear.

Automatic summarisation

Documents can have tens or hundreds of pages.

A complex contract, technical documentation, a financial report or a requirements document can take a long time for a first analysis.

AI can generate a summary that presents the main points.

What matters, however, is that the summary must not replace the original document.

In a serious DMS there must always be a difference between:

• the original document, which is the official source;
• the AI interpretation, which is an automatically generated result.

This separation is essential.

Artificial Intelligence can make mistakes. It may omit an important paragraph or misinterpret a legal formulation.

That is why, in areas such as legal, financial, medical or public administration, AI must be treated primarily as an assistant, not as the final authority.

AI in workflow: the document starts to move by itself

One of the greatest advantages appears when AI is connected to the workflow engine.

Suppose an invoice arrives. AI automatically identifies the supplier, amount and due date. The DMS can apply a rule:

• if the amount is under 5,000 lei, send the document to the department manager;
• if the amount exceeds 5,000 lei, also request approval from the finance director;
• if the supplier is new, also send the document to Legal;
• if the invoice contains an IBAN different from the one in the database, raise an alert.

Here AI is no longer just a chatbot. It becomes part of the process.

In 2026, Box describes this evolution through AI flows in which information is extracted from documents and used for automatic routing of tasks and approvals.

Mayan, in turn, offers the ability to send AI model results to the workflow engine.

This combination — document + AI + workflow — is probably one of the most important directions for the modern DMS.

Identifying missing documents

AI can also play an interesting role in case-file management.

Suppose a certain type of file requires six mandatory documents. The DMS can know the required structure: application, ID copy, certificate, proof of payment, declaration and supporting document.

Instead of the operator checking the file manually, the system can display:

File complete: 5 of 6 documents. Missing: proof of payment.

Moreover, if the user uploads a document named “scan000134.pdf”, AI can recognise from the content that it is, in fact, the proof of payment.

In this way, the DMS starts to understand not only individual documents, but also the context of the file.

Detecting sensitive information

Artificial Intelligence can also contribute to security.

A document may contain personal ID numbers, bank details, medical information, salaries, passwords, confidential commercial information or other sensitive categories.

AI can identify such information and help the system apply rules. For example:

• “The document contains personal data.”
• “Access restricted to the HR department.”
• “The document must be anonymised before external transmission.”
• “Confidential financial information was identified.”

AI can even be used to anonymise copies intended for publication, automatically hiding certain personal information.

But this function must be connected to the classic DMS mechanisms: permissions, ACLs, audit and retention policies.

AI must not replace security. It must complement it.

AI must respect DMS permissions

This is probably one of the most important rules when designing an intelligent DMS.

If a user is not allowed to open a document, AI must not give them information extracted from that document.

It seems obvious, but implementation can be difficult.

A chatbot improperly connected to the entire archive can become a path through which a user learns information they would normally not have access to.

For that reason, modern solutions place growing importance on “permission-aware” AI.

M-Files emphasises that results provided to Microsoft 365 Copilot are grounded in the M-Files permission model, precisely so that information provided by AI respects the user’s rights.

Box is also developing controls and policies specific to AI agents that access company content.

In a well-built DMS, AI must see exactly what the user it works for is allowed to see. No more.

Cloud AI or local AI?

Another important decision is where the model runs.

A DMS can use cloud AI services. The advantage is access to high-performing, up-to-date models.

But some organisations do not want their documents sent to an external provider.

That is why the local AI option is becoming increasingly relevant.

Mayan EDMS 4.12, for example, supports integration with OpenAI as well as connection to Ollama, which allows models to run on the organisation’s own infrastructure or on infrastructure it controls.

For companies that manage sensitive information, public administrations, law firms or organisations in regulated sectors, this option can be extremely important.

We can thus have a DMS where documents are stored locally, OCR runs locally, the search engine runs locally, the AI model runs locally, and information does not need to leave the organisation’s infrastructure.

This is one of the directions where AI and the concept of digital sovereignty can meet. For documents and collaboration on controlled infrastructure, see also INITWIN Cloud.

Human-in-the-loop: people must stay in the process

Artificial Intelligence must not be designed as a mechanism that completely removes people from processes.

Especially for important documents, the healthy approach is “human-in-the-loop”.

AI proposes. The person confirms.

For example:

• AI: Document type – Contract.
• AI: Partner – Example LLC.
• AI: Value – 250,000 lei.
• Operator: Confirms.

If there is an error, the operator corrects the information. Over time, those corrections can improve the process.

Such a model keeps automation fast without turning AI into an opaque mechanism that makes unverifiable decisions.

In Europe, the legislative framework for Artificial Intelligence must also be taken into account. The EU AI Act introduces a risk-based approach, and certain transparency obligations under Article 50 apply from 2 August 2026.

For DMS developers, the message is clear: AI must be introduced together with transparency, accountability and control mechanisms.

The DMS of the future will no longer be just an archive

The role of Artificial Intelligence in a Document Management System is not to add another button labelled “AI”.

Its role is much deeper. AI can transform the entire document lifecycle.

The document enters the system. It is read. It is classified. Metadata is extracted. It is linked to a registration or a file. Annexes are identified. Mandatory documents are checked. Relevant information is extracted. The document enters a workflow automatically. Users can find it through semantic search. They can ask questions about the content. They can receive summaries. The system can identify deadlines, obligations or risks.

And all of this must respect permissions, audit and organisational rules.

Instead of “Which folder is the contract in?” we will ask “When does the contract expire?”. Instead of “Where are the annexes?” we will ask “Is the file complete?”. Instead of opening twenty documents to find one piece of information, we will be able to ask the system to identify it and point to the document it comes from.

This is the shift from Document Management to Intelligent Document Management.

And it will probably be one of the most important transformations of DMS systems in the coming years.

Artificial Intelligence does not eliminate the need for organisation, registry, archiving, permissions, audit or workflow. On the contrary. It needs all of them.

AI connected to a disorganised collection of files will produce unreliable results. AI connected to a well-structured DMS — where we know what each document represents, who has access, what relationships exist between documents and which version is official — can become an extraordinary productivity tool.

That is why the future is not simply “AI for documents”.

The future is AI integrated into a solid information management system.