KognitCapture is a multi-tenant, self-hosted capture platform that replaces ABBYY FlexiCapture, ABBYY Vantage, Kofax / Tungsten, ChronoScan, Rossum and Ephesoft — with no per-page meter, no station licences and no module unlocks. Your documents stay on your hardware. Your bill stops moving when your volume grows.
Paper, PDF and structured e-invoices come in through one intake. Line items are extracted per supplier layout, rules check totals and master data, and reviewers only see the invoices that need a decision. The fee does not depend on how many invoices you process.
A scan desk in the browser that flags bad pages while the paper is still in hand. Review queues ordered by deadline. Keyboard-first verification with the page beside the fields, and locking so no two people open the same document. Nothing to install on any desk.
A single Java application on Linux with PostgreSQL. OCR and models run in-process — no external recognition service, no licence server, no message broker. Scale by adding nodes and telling each what it does. Air-gapped operation is a documented deployment.
An OpenAPI 3 specification generated from the running code. One call to capture a document. Template versions you can pin. Scripts in JavaScript, Groovy and Java, a plugin SDK with five extension points, embeddable review screens. And an MCP server with more than a hundred tools.
Recognition runs inside your installation. Remote AI is opt-in per organisation and per step. Access is role-based with per-project grants; every field keeps its history; retention is set per project. And we list plainly what the product does not do.
Licensed per installation by your turnover band, from €900 a year: unlimited users, projects and pages, increases capped at 25%. No runtime licence gate. Blueprints create a working project in one step instead of a quarter of consulting. Your data stays in your database.
Tenants, organisations and projects keep clients apart, down to the directory tree. Name, logo, colours, domain and mail provider are set per client. The review and scan screens embed in your own product. Your connectors ship as plugins.
Runs on Linux and PostgreSQL. Browser-only operator hubs — nothing installed on a desk. Tesseract, local GGUF models or fifteen named LLM providers, switchable per project and per step.
The established platforms in this category tend to share two properties. They meter you per page, so the more value you extract, the more you pay. And configuring a single document type is a specialist consulting project. The cloud alternatives remove the installation — and require your invoices, claim forms and identity documents to leave your building, which for a growing number of organisations ends the conversation.
Language models read documents that classic OCR cannot. But the usual way to get one is somebody else's API: each page becomes a request to a third party, each request is billed by the token, and the model behind the endpoint can be changed or retired on the provider's schedule, not yours. For invoices, claim files and identity documents, that is a meter and a data transfer in one.
A language model reads the invoices and forms that classic OCR struggles with. Bought as a cloud service, it brings back the two things you were trying to leave: a bill that grows with every page, and documents that leave the company to be read. A budget you cannot predict, for a process you cannot fully audit.
A cloud language model is a data transfer. Every page a step sends becomes a request to a third party, under that party's retention and sub-processor terms, in a region you may not choose. KognitCapture does not redact personal data before a model call — so the only complete answer for sensitive material is a model that runs where the documents already are.
KognitCapture treats the model as a component you choose, not a service you are tied to. It can run inside your installation.
Local first, remote by choice, and the same workflow either way: switching a step from a local model to a provider — or back — is a setting, not a rebuild.
Honest limits: a local model is slower than a large hosted one on the same page unless you give it a GPU, and smaller models read difficult documents less well. AI is optional at every stage — Tesseract, native PDF text, rules and in-process machine learning need no language model at all.
| Question | KognitCapture | Classic capture suite | Cloud capture API |
|---|---|---|---|
| How is it priced? | Annual fee per installation, by turnover | Usually per page or per volume band | Per page or per call |
| Where are documents processed? | On your infrastructure | On your infrastructure | In the provider's cloud |
| Runs without internet? | Yes, documented air-gapped install | Often, with a licence server | No |
| Human review screens | Included, in the browser, embeddable | Included, often a desktop client | You build them |
| Workflow and routing | Included, visual designer | Included | You build it |
| Scanning | Browser scan desk, driverless network scanners | Desktop scan station, TWAIN / ISIS | Not part of the service |
| AI models | Local or any of 15 providers, your choice per step | Vendor's own engines | Provider's own models |
| API coverage | Everything the interface does | Varies; often partial | Recognition only |
| New document type | Blueprint or visual template designer | Often a consulting engagement | Prompt or pre-built model |
| Access for AI agents | MCP server, OAuth 2.1, audited | Rare | Rare |
| Vendor security attestation | None — runs under your controls | Often available | Usually available |
| Named ERP connectors | None — generic REST, database, file, bus | Often available | Rare |
| TWAIN / ISIS scanner drivers | Not supported | Supported | Not applicable |
Each page is a straight read: where that product genuinely beats us, where we beat it, and what the move actually costs. We publish both columns, because a comparison that only lists strengths is worth nothing to you in an evaluation.
Metered platforms charge roughly $0.02–0.10 a page (public list bands). At two million pages that is a page bill on top of a platform fee. Here recognition runs on your own CPU with Tesseract or a local model: the marginal cost of the next page is the electricity to read it. Use a remote LLM if you want to — tokens are logged and costed per tenant, and you decide which projects may call out at all.
Self-hosted on Linux and PostgreSQL, air-gap capable, with local GGUF models for both text and vision. Tenant → organisation → project isolation reaches down to the storage prefix, and every agent, API key and operator action lands in an audit channel. There is no licence server to phone home to, because there is no licence gate.
Scanning, processing and verification hubs are browser pages. Network scanners are driven server-side over eSCL, so a scanning desk needs a browser and nothing else — no thick client, no driver rollout, no per-station licence. The hubs also embed into your own application behind an API key, which is how software vendors ship capture without building it.
FlexiCapture, ReadSoft and ChronoScan expect one deployment per customer. This platform was built tenanted: a service provider runs twenty clients on one cluster, with per-client storage, roles, retention, AI budgets and audit. That is the single strongest number in any business case a service provider puts together.
There is no converter that turns a FlexiCapture document definition into a Vantage skill — practitioners describe the path as inventory, map, rebuild, re-integrate, run in parallel, and such deployments are commonly quoted at eight to twelve weeks.
If you have to rebuild anyway, the honest question is no longer "replace something that works" but "rebuild there, or rebuild here". Here, an importer reads your FlexiCapture 12 project and gives you a head start.
Read the migration brief →Project configuration, the scanning hub with its assembly board, the template and workflow designers, and the verification hub with zone overlays, confidence bars and validation errors — rendered from the application's own stylesheets and markup, on made-up documents.
Nothing exotic to learn before you can run it, and nothing proprietary between you and your data.
One fixed annual fee covers one installation, one tenant and one organisation, with unlimited users, projects and pages. It scales with your own turnover in marginal steps, with a €900 floor and yearly increases capped at 25%.
| Turnover | Per year |
|---|---|
| Up to €500k | €900 |
| €1M | €1,500 |
| €5M | €2,500 |
| €25M | €5,500 |
| €100M | €13,000 |
| €500M | €37,000 |
| €1bn | €52,000 |
Ex VAT. Hosting, remote AI usage and professional services are separate.
We would rather lose the deal in week one than in month six. The full list of what we do not claim → · Where each competitor beats us →
We publish no headline percentage, because one number across unknown documents means nothing. On clean print Tesseract is competitive; on degraded scans the best commercial engines still read better, which is what the hybrid and model paths are for. A bake-off on your own pages settles it — scoped, one to two days, with a field-level report you keep.
Not unless you configure an export goal or a remote AI provider that sends it. OCR, classification and local models run inside the installation.
One annual licence per installation by turnover band: from €900, €2,500 at €5M, €13,000 at €100M. Unlimited users and pages.
A platform licence for the deployment, not a meter on your documents. Above roughly 100,000–150,000 pages a year, self-hosting is decisively cheaper than a metered IDP service and the gap widens every year. Below that, a metered platform is often genuinely cheaper than running infrastructure — we will say so. Run the numbers on the pricing page: the cost model.
No. Recognition can be entirely local: Tesseract for text, a local GGUF model for vision through a managed llama-server. Fifteen named remote providers are supported for teams that want them, per project, with tokens and cost logged per tenant. An air-gapped install is a supported deployment, not a workaround.
From FlexiCapture 12, an importer carries fields, rules, datasets and scripts across as a starting point. From anything else, field sets, validation rules and export mappings are rebuilt here — the same work any migration between capture platforms requires, including ABBYY's own path from FlexiCapture to Vantage. What shortens it: industry blueprints that provision a project, workflow, routing, upload endpoint, export goal and a full field set in one action, and model-assisted template authoring that drafts a template from one sample page. Blueprints ship pre-authored field sets, not pre-trained models — see the migration brief for what that means in weeks.
KognitCapture is self-hosted: there is no licence server that can switch your installation off, and your data stays in your own database. Source access and escrow are agreed in the contract conversation.
Through REST, a database insert, a file, a message bus or mail. There are no product-specific adapters; a script or plugin covers special cases.
It receives and maps them, together with Factur-X, ZUGFeRD and XRechnung. It does not send e-invoices and is not an access point.
Network scanners that support eSCL work from the browser with nothing installed. TWAIN and ISIS devices work through scan-to-folder.
A Linux x86-64 server with Java 25 and PostgreSQL. 4 GB of heap to start, 8 GB for real throughput, more if you run local models.