This page explains how DocAI protects the documents you entrust to it: where they are stored, who can reach them, what gets logged, and what happens to the data during processing. We wrote it for managing directors, finance leads, IT and data protection officers, in enough detail to fill in a supplier security questionnaire. Every statement here is backed by a working function: what you read on this page can be verified in the code. What we cannot verify, we do not write down, and we list those gaps explicitly at the end of the page.
The key facts in one place. If it is not here, we do not claim it.
Last reviewed:
There are two operating models, and responsibility is split differently in each.
DocAI is the same system in both models, with the same security functions. The difference is where the machines run and who operates them.
The entire system runs on the customer's infrastructure: document files, database, search index and the AI models as well. Documents never leave the organisation, and processing requires no external service. The system also runs in isolated mode, without an internet connection; in that case only the features that call external data sources are unavailable, for example the connection to the Hungarian tax authority. In this model the customer is responsible for backups, for restore capability and for the network environment, while K3Net is responsible for the system itself.
If the customer has no operations capacity of its own, K3Net runs the system on a server in Hungary. The data stays within the European Union. Every customer gets a separate system instance, database and search index; documents and extracted data never mix between customers. The AI model server runs on shared infrastructure: it serves the request but keeps no customer data persistently and does not learn from it. In this model K3Net is responsible for backups, for handling the encrypted copies and for running the server.
Access is not decided by folder structure, but by role and document type.
Every DocAI user belongs to one or more groups, and the group determines which document types the user can see and what the user may do with them: read, record, edit, approve. Because permissions are resolved at document type level, your organisational logic maps directly onto the system.
More importantly, permissions do not stop at the list view. Full text search, semantic search and document chat all work exclusively on the documents the given user is allowed to access. If a user cannot open a document, the AI will not answer from it either.
Not just "someone changed something".
The system keeps a field level change log. Every entry shows who performed the action, when, on which document, on which field, and what the value changed from and to. The log is not limited to editing: downloads and detail view opens are recorded too, so it can be established afterwards who downloaded or opened a given document.
In an audit this means a question about one specific document (who saw it, who edited it, when and to what) is not answered by estimation. It can be listed.
Sign-in works with a password, and two factor authentication (TOTP, using an authenticator app) is available and can be switched on per user. A strict password policy can be enabled per customer: at least 12 characters, upper and lower case, digit and special character, 180 day expiry. If the customer's internal policy differs, we agree on it during rollout.
Every uploaded document file is stored with encryption at rest, so filesystem access alone is not enough to read the documents. Data in transit moves over an encrypted channel, TLS, including all traffic between the user interface and the system.
The database is backed up in encrypted form, with a defined retention period, in several copies, daily. Under K3Net operation this is K3Net's task. In an on-premise install, backups fit into the customer's existing routine: at rollout we hand over which directories and which database must be backed up and what to watch for during a restore, but running and verifying the backups belongs to the customer's IT operations.
For each document type you can set how many years the related documents must be kept. From this the system calculates an expiry date per document, so it is always possible to query which documents have reached the end of their retention period. Disposal remains a manual decision: expiry does not trigger automatic deletion. A document also cannot be deleted while an approval is in progress on it. A manually deleted document stays in the recycle bin for one year and can be restored; after that the system removes it permanently.
Documents are not isolated files. The system maintains a chain of document links that can be traversed as a graph:
So when an invoice is disputed, nobody has to hunt down the referenced contract and the acceptance certificate by hand: the connection is visible from the document record itself.
Where a document has to stand on its own authenticity, electronic signing is available: the system applies a PAdES B-LTA signature with a qualified timestamp and logs the signing events.
DocAI's language models run on the system's own infrastructure by default, so document content is not sent to an external provider even for processing. This is not a setting to be found, it is the system's default state.
An external AI service can be brought in if the customer explicitly asks for it, but under three conditions: it has to be switched on deliberately, switching it on is permission bound, and switching it on withdraws the EU AI Act fact sheet, because the processing conditions are different from that point. By default this option is off.
The role of the AI is visible in the interface as well. For every extracted field the system records whether the value came from a user or was read out by the AI. The user supplied value is the source of truth: the AI does not overwrite it. If the AI reads something different, it flags the difference, and the decision stays with the human.
Documents are approved inside the system, not over email. An approval round can have a deadline, a deputy can be set for absences, and every decision in the process is logged: who approved or rejected, when, and with what comment. Approval rounds can be built in several steps, so multiple approvers can be involved.
While an approval is in progress on a document, that document cannot be deleted. This delete protection prevents an ongoing case from losing its trail mid-process.
K3Net signs a data processing agreement with the customer, setting out the purpose and scope of processing, the security measures applied and the question of subprocessors. By default data stays within the European Union; it can reach a third country only if the customer switches on an external AI service, and the data processing record states this explicitly. When a data subject request arrives (access, rectification, erasure), the customer forwards the request to us and we carry it out: today this is not a self service function in the interface, we deliver it as a service.
DocAI is a decision support system operating under human oversight: the business decision stays with a human, its suggestions and automated steps are logged and can be checked against the source document, and the processing steps are retrievable. The classification of any given deployment always depends on the actual purpose of use, so we review this per rollout. The detailed reasoning is on our blog: AI, GDPR and EU AI Act compliance.
Certification is granted to an organisation, not to software, so no system can hand its customer an ISO certificate. What we can hand over is evidence: the controls described above (access management, logging, encryption, backup, retention, approval) can be mapped point by point against the expectations listed in the ISO 27001 annex and among the measures required by NIS2, and the reports the system produces can be submitted as evidence in the customer's own audit.
The same applies to records management: the documented information belonging to a regulated operation (supporting documents, minutes, approved records) can be handled in a traceable way in the system, because every view, download and field level change leaves a trail.
Two of our customers, Trigo Kft. and Duna-Dráva Cement Kft., used DocIT as their records and evidence management system in their successful NIS2 audits. For now they use it without the DocAI extension, but the foundation is shared: everything DocIT can do is also available in DocAI.
Preparing for those audits produced the set of answers that supplier security questionnaires typically ask for. These are the answers we give today:
We write this list down so nothing surprises you during supplier assessment.
There are two operating models. On-premise, DocAI runs on your own server and the documents never leave your network. In the K3Net-operated version they are stored on servers in Hungary, in a separate instance and database per customer. By default, in neither case does data leave the EU; that changes only if the customer switches on an external AI service.
Yes. Every document file is encrypted at rest, transfers use TLS, and database backups are encrypted and kept with a retention period.
Yes. The system logs changes at field level (who, when, what, from which value to which), and records downloads and detail view opens as well. The log is kept in yearly separated records; the application offers no log deletion function.
Yes, with time-based one-time codes (TOTP, e.g. Google Authenticator), enabled per user. A strict password policy can be enabled per customer: at least 12 characters, mixed character classes, 180 day expiry.
No. The language models run locally, on DocAI's own infrastructure. External AI services can only be used after explicit activation, tied to permissions, and switching them on withdraws the EU AI Act fact sheet; they are off by default.
No, and software cannot be: ISO 27001 certificates are issued to organizations. DocAI's controls (permissions, audit log, encryption, backups, on-premise AI) are comparable with the measures expected by ISO 27001 and NIS2, and can be handed over as evidence for your audit.
Yes, as a records and evidence management system. Two of our customers, Trigo Kft. and Duna-Dráva Cement Kft., used DocIT as their records and evidence management system in their successful NIS2 audits. They currently use it without the DocAI extension, but the foundation is shared: everything DocIT can do is available in DocAI.
The retention period is set per document type in years, and the system computes the expiry date. Disposal is a manual decision, and a document under an open approval cannot be deleted. Data subject (GDPR) requests such as erasure or data export are fulfilled on request.
If your security questionnaire contains a point this page did not answer, ask us: we can show you the specific control in operation. In a demo we walk through permissions, the audit trail and the operating model, mapped onto your own processes.
Request a demoFill out the form below and our team will reach out shortly to schedule a demo.
We only use your data to respond to your inquiry. We do not share it with third parties. Legal basis: your consent and pre-contractual steps (GDPR Art. 6(1)(a) and (b)). For details, see our privacy policy.