Straight answers, including the ones that are no.

Software you install on your own servers that turns incoming documents — scans, PDFs, images, office files, e-invoices — into checked, structured data. It imports from fourteen kinds of source, reads pages with OCR and optionally AI models, classifies and splits documents, extracts fields through versioned templates, validates them, sends the doubtful ones to a person, and exports the result to your systems.

Three ways. Your documents are processed on your infrastructure, not sent to a provider. You pay a fixed annual licence instead of a price per page. And you get the whole process — review screens, workflow, validation, export — not only a recognition call.

No page-based licence and no licence server. A browser interface for everything, including scanning. An API for everything the interface does. Language models as an option at several stages. And configuration by blueprints and a visual designer rather than a long consulting project per document type.

We publish no accuracy percentage, because one number across unknown documents means nothing. Clean printed text reads well with Tesseract; poor scans, unusual typefaces and handwriting need a vision model, a trained language pack or a reviewer. The honest way to answer is to run your documents — which is what we offer.

Yes. Text models in GGUF format run inside the application, and vision models run in a local model server managed by the platform on the same host. You can also point a step at your own Ollama, vLLM, LM Studio or other OpenAI-compatible server. In all of these cases pages stay on your network and there is no per-token charge — you pay in hardware instead.

Remote providers remain available per organisation and per step when you decide a hosted model is worth it, with a monthly spending ceiling.

On clean 300 DPI print, Tesseract 5 is competitive. On degraded scans, dense forms and non-Latin scripts, the best commercial engines still read measurably better — we will not pretend otherwise. The answer for those documents is the hybrid path: Tesseract first, and low-confidence words or whole pages re-read by a vision model, local or remote, at your choice.

A fifteen-minute scanner and volume call first. Then someone reads your documents, authors templates, runs the Tesseract, hybrid and model paths over them, and writes up field-level accuracy with the cost per thousand pages of each mode. It runs one to two days — two is the ceiling. You keep the report whatever it concludes.

Yes. The live demo is the application's own interface running in your browser on synthetic documents — about forty screens, no sign-up.

KognitCapture is self-hosted: there is no licence server that can switch your installation off. Source access and escrow are agreed in the contract conversation. Your documents are files on your disks and your data is in your own PostgreSQL database throughout.

The interface is in English. Recognition is a different matter: around 129 OCR language packs can be installed, several per document.

KognitCapture is a product of Automatize BV in Heist-op-den-Berg, Belgium, developed together with A4IT and A4ME. It is designed, built and maintained in Europe.

Yes. Table extraction uses column zones, row grouping and header detection, and follows a table across pages. Line items are also extracted by a model into the template's own row structure when you enable one. Complex line layouts are where a test on your suppliers' invoices tells you most.

One template with alternatives. Each alternative is a layout variant; all of them produce the same field set. For the long tail of suppliers you see twice a year, label search and model-based extraction read invoices without a dedicated variant.

On the receiving side. It reads UBL and CII documents with an embedded PDF, and ZUGFeRD, Factur-X and XRechnung files, and maps them onto the same template as your paper invoices. It does not send e-invoices and is not a Peppol access point.

It can check what rules can express: totals add up, VAT is consistent, the supplier exists in your master data, the purchase order number is found in a dataset, database or API. Whether the goods were received is a question for your ERP.

It delivers the data in the format and through the channel your package accepts: a file, a database insert, a REST call. There are no named connectors for specific accounting products.

Yes. Each field keeps the recognised value, every amendment, who verified it and when, next to the page image it was read from.

No. The Scan Hub runs in the browser and the server talks to network scanners that support eSCL (AirScan / Mopria).

Not directly; those drivers are not supported by design. Set the scanner or its software to scan to a folder and let a hot-folder source import the files.

Nothing is lost. The scan session is held on the server; reopen it and continue.

No. Opening a document checks it out. Locks can cover a whole batch, expire by themselves, and can be released by an administrator.

Not to a named person today. Access is by hub grant per project, and the queue is ordered by service level, priority or arrival. Assignment to individual reviewers is not available.

Set service-level targets per project. The hubs can order work by deadline, documents close to breach are marked at risk, and alerts and digests are mailed.

Only through a vision model; there is no separate handwriting engine. Expect handwritten fields to need review.

A Linux x86-64 server, Java 25, and PostgreSQL. As a floor: 4 GB of heap for a small installation, 8 GB for real throughput. Local language models need additional memory outside the heap — budget roughly 4 to 5 GB for each 7-billion-parameter model at 4-bit quantisation.

Linux x86-64 is the supported platform, because that is where the bundled OCR libraries are self-contained.

We do not publish images or manifests today. The deployment unit is a JAR run as a system service.

Add nodes. They share one PostgreSQL database and one filesystem; there is no message broker. Each node is given duties and capabilities, so you can dedicate machines to OCR or to export.

Not for OCR, classification or the workflow engine. Local language models run on CPU and benefit from a GPU; whether you need one depends on the model and the volume.

Replace the JAR and restart. Database migrations run on start-up, and the built-in self-verification scenarios check the installation afterwards.

A GraalVM native-image build profile exists but is not validated, so we do not offer it as a supported deployment.

Back up the PostgreSQL database and the file tree with your own tooling. The product has no built-in backup.

The web interface is itself a client of the REST API. The OpenAPI 3 specification is generated from the running application, and the built-in console shows the operations your account may call.

With a bearer token for the general API. API keys are used for specific lanes: embedded hubs, mobile devices, dataset loading and the MCP server.

Not if you pin a version. Publishing takes a snapshot; a workflow step or a capture call can name the version it wants.

Outbound: a webhook export goal with basic, bearer or API-key authentication, and batch completion callbacks. Inbound: a webhook import source verified by HMAC. There is no WebSocket or server-sent event stream.

No. A document follows one path at a time, with decisions. Parallel branches and joins are not supported.

Fifteen named providers, any OpenAI-compatible endpoint, local GGUF models, and anything else through an LLM-provider plugin — there is a sample project for exactly that.

Software is not compliant; an organisation's processing is. What KognitCapture gives you is control: processing on your own infrastructure, access control, audit logs, retention policies and deletion. The documentation maps these controls to GDPR obligations to support your own assessment.

No. There is no vendor attestation for KognitCapture. Because you host it, the software sits inside your own certified scope.

Not yet. Both are on the roadmap. Today users sign in with a username and password, protected by lockout and rate limiting.

Secrets — credentials and provider keys — are encrypted by the application with AES-256-GCM. Document files and the database are not; encrypt the volumes and the database with your platform's facilities.

Nothing by default. If you configure a remote provider and use it in a step, the pages scoped for that step are sent, without automatic redaction. A local model keeps everything inside the installation.

Until you say otherwise. Retention policies are defined per project and are off until you create one.

One annual licence per installation, based on your turnover: from €900 a year, €2,500 at €5M turnover, €13,000 at €100M. Unlimited users and pages. The pricing page has the full table and a calculator.

A server and database to run it on, the effort to configure templates and integrations, and — only if you choose a remote AI provider — that provider's usage charges.

Your documents are files on your disks and your data is in your PostgreSQL database, in a documented schema. There is no runtime licence gate. Configuration can be exported, and a full project report documents every setting.

Installation is one JAR and a database. A blueprint gives you a running project in one step. The time that matters is tuning templates and rules on your documents and connecting the export — which depends on the documents.

We prefer to prove it on your own documents. Ask us about references in the conversation.

There is a page for each of nine platforms, with both columns: where that product beats us and where we win, plus reported cost. Start at Alternatives, or put your page volume into the cost comparison.

At low volume. Below roughly 100,000 to 150,000 pages a year a pay-per-page service is often cheaper than running your own infrastructure, and we will say so. Above that, self-hosting pulls ahead and the gap widens with every additional page.

Technically yes: tenants and organisations separate clients, with their own storage, branding, domains and mail providers. Commercially this is agreed separately from the standard licence.

Yes, for resellers, software vendors embedding capture, and service providers running many clients. See Partners.

Yes. Product name, logo, colours, typeface, sign-in page and footer links can be set at platform, tenant and organisation level.

The Verification Hub and Scan Hub embed in an iframe, opened by a server-side session. Everything else is available through the API.

Yes, as plugins: import connectors, export connectors, workflow steps, email providers and model providers.