anyformat vs Klippa, now Doxis AI.dp (2026)

Klippa, rebranded Doxis AI.dp in March 2026, matches anyformat on EU hosting, ISO 27001, no-code workflows and per-field confidence. The difference is deployment and scale: Klippa is Azure SaaS with no documented on-premise option, and documents over 50 pages time out, while anyformat runs on-premise and held ~99% row recovery at 50 pages.

Is Klippa the same as Doxis AI.dp?

Yes, they are the same platform under a new name. SER Group of Bonn, Germany acquired Klippa on 18 March 2025, and the rebrand to Doxis completed on 30 March 2026; the legal entity is now Doxis AI Solutions B.V., still based in Groningen, Netherlands (SER Group, 2025; Klippa newsroom, 2026). The product names moved with it: DocHorizon, the core intelligent document processing platform and its Flow Builder, is now Doxis AI.dp; Klippa SpendControl, the accounts-payable and expense SaaS, is now Doxis SpendControl; the identity-verification line runs on the same extraction engine. Klippa was founded in Groningen in 2015 and lists customers including Eurofins, SNCF and Siemens. If you are reading a quote, a contract or a documentation page issued before March 2026, it describes this platform under the previous name, and the DocHorizon domain still serves the technical documentation.

Where Klippa and anyformat match

Five things buyers often treat as differentiators are not differentiators in this comparison, and it is worth stating that plainly before the rest of the page.

European sovereignty. Klippa is a Dutch company under German ownership, hosting on Azure regions in Germany, where the full service catalogue runs, and in Amsterdam, where a limited set of services is available (Doxis AI.dp documentation, 2026). That is a real EU posture and it survived the acquisition, because SER Group is itself German. anyformat is EU-native, but sovereignty alone will not separate these two vendors and we do not present it as if it did.

ISO 27001. Klippa has held ISO 27001 certification since around July 2023, with the certificate scope now issued to Doxis AI Solutions B.V. (Klippa newsroom, 2026). anyformat is ISO 27001 certified across the document pipeline. Both vendors are certified; the question is scope, not existence.

No-code workflow building. Flow Builder is a genuine visual drag-and-drop builder, not a label on a config file: ingestion from email, cloud storage, ERP systems and API, human-in-the-loop review steps, and 50+ integrations (Doxis AI.dp documentation, 2026). Explicit if/else conditional branching could not be confirmed from the public documentation, which is a question for a demo rather than a gap we can assert.

Zero-shot custom fields. Prompt Builder creates custom extraction models without labelled training sets, and Klippa reports 88% less time to build a custom model (Klippa Prompt Builder page, 2026). The same page carries the vendor's own caveat that the outcome can be unpredictable and occasionally incorrect, which is an unusually honest thing to publish and worth crediting.

Per-field confidence with coordinates. The Candidates component returns confidence values per field together with bounding-box coordinates, and is enabled automatically when human review is switched on (Doxis AI.dp documentation, 2026). Klippa does not publish a calibration methodology. anyformat measures calibration accuracy at 99.1% with an Adaptive ECE of 0.009 (anyformat benchmark, 2026) and attaches visual citations linking each extracted value to its exact position on the page, which is the part that makes a straight-through threshold defensible rather than approximate.

Deployment: Azure regions versus on-premise

Klippa runs as multi-tenant SaaS on Microsoft Azure, and no on-premise, self-hosted, Docker or air-gapped deployment option is documented anywhere in its public material. Region selection happens at project level between Germany and Amsterdam, and the region cannot be changed after a project is created, so the choice is effectively permanent and a migration means rebuilding the project (Doxis AI.dp documentation, 2026). anyformat deploys as SaaS, in a customer-controlled private cloud or VPC, and in air-gapped on-premise environments on the Enterprise plan (anyformat pricing, 2026). The distinction is narrow but decisive for a specific set of buyers: defence suppliers, parts of healthcare, public-sector bodies and banks operating under internal data-egress rules, where an EU region on a hyperscaler is a hosting location rather than a perimeter the organization controls. If your documents can go to a European cloud, this section does not apply to you. If they cannot leave your own infrastructure, it is the whole decision.

Data residency, GDPR and retention

Klippa's GDPR position is that of a processor acting under a data processing agreement, which is how its own compliance page frames it (Klippa compliance page, 2026). That is standard and legitimate; it is also contractual rather than architectural, which is why the table above records GDPR as partial rather than yes. Two details belong in a procurement checklist. First, one compliance page lists optional server locations in the United States and Switzerland alongside the European ones, which sits awkwardly next to a purely EU-native reading of the architecture, so confirm in writing which locations apply to your tenant. Second, the certification wording describes servers as ISO 27001, SOC 1 and SOC 2 certified, phrasing that reads as though part of it may be inherited from the Azure platform underneath; no SOC 2 report, auditor or Trust Services Criteria scope is published, so treat first-party SOC 2 as unconfirmed until you have seen the report. On retention, Klippa documents PII masking that deletes the unmasked value after redaction, but no account-level zero-retention switch. anyformat offers zero data retention as a contracted Enterprise option, and its ISO 27001 scope covers the document pipeline itself rather than only the infrastructure beneath it.

Long documents, tables and figures

Klippa's own file-limitations documentation states that documents longer than 50 pages time out and fail, and that documents with many line items can time out below that threshold; images are capped at 20 MB, and the identity model at 16.78 megapixels with no side longer than 4096 pixels (Doxis AI.dp file limitations, 2026). That is a ceiling rather than a degradation curve. A 120-page contract pack, a full-year bank statement, a consolidated annual report or a long customs manifest does not process, and the workaround is a splitting step you build and maintain yourself. anyformat was benchmarked on exactly this failure mode: on line-item tables scaling from 1 to 50 pages and roughly 2,400 rows it held ~99% row recovery, and it extracted 94% of mixed-format invoices with every field and every line item correct, reaching 80% on documents of 16 pages or more where the next best system stopped at 53% (anyformat benchmark, 2026). Separately, no figure, chart or diagram detection is documented for Klippa; anyformat detects visual elements and returns structured descriptions of them, which matters when part of a report's meaning sits in a chart rather than in a table.

Evaluation and monitoring

We found no evaluation infrastructure in Klippa's public material: no accuracy dashboard, no ground-truth dataset management, no regression suite for workflow changes, and no published precision, recall or F1 for any model. Human-in-the-loop review exists and is documented properly, but review catches errors one document at a time; it does not tell you whether a change to a prompt, a schema, a threshold or an underlying model made the system better or worse across your document mix. This is stated as not found rather than absent, since something may exist behind an enterprise agreement. The relevant fact for a buyer is that it cannot be verified from the outside, and that the vendor's own caveat about unpredictable Prompt Builder outcomes makes a measurement loop more important, not less.

Which LLM does the Prompt Builder use?

Klippa confirms that Prompt Builder uses LLM technology, but its public documentation does not disclose which model provider is involved or where inference runs (Klippa Prompt Builder page, 2026). We are not asserting an answer, because there is no sourced one to assert. We are pointing out that the question is open, and that for a vendor whose primary security argument is European data residency it is a material one: which provider, in which region, under which sub-processor agreement, with what retention on the model provider's side, and whether that changes when Prompt Builder is used rather than a pretrained model. Ask it in procurement and get the answer in the DPA. anyformat's architecture is model-agnostic, and the answer is contractual rather than implied: zero data retention is available on Enterprise, and an air-gapped on-premise deployment keeps inference inside your own perimeter (anyformat pricing, 2026).

Pricing

Klippa publishes no pricing for its core IDP API; the published numbers belong to the accounts-payable product. Doxis SpendControl lists Effective at €95 per month for up to 4,000 invoices per year and up to 10 users, Premium at €275 per month for up to 12,000 invoices per year and up to 30 users, and a Custom tier that adds SSO and API access; add-ons are priced separately, with workflows at €50 per month, accounting integrations from €50 per month, ERP integrations from €100 per month and extra invoices at €0.28 each (Doxis SpendControl pricing, 2026). That is legible for an AP team with predictable volumes, and it is more than most vendors in this category publish. It does not extend to general document processing, where you are back to a sales conversation.

anyformat bills in credits, charged per page and per operator, so the cost tracks the operations a workflow actually runs: 10 credits cost €0.01, so Parse is 25 credits per page (€0.025, published as $25 per 1,000 pages), Extract 35 credits per page, Classify 10, Split 25, and Validate 25 credits per rule (anyformat pricing, 2026). Pay As You Go is free to start with a one-time grant of 50,000 credits, Business is €499 per month (€399 billed annually) for 500,000 credits, 5 users and 10 workflows, and Enterprise is custom-priced with unlimited users and workflows, VPC or on-premise deployment and zero data retention.

What does anyformat ship that Klippa does not?

An evaluation suite, and a set of guarantees around the workflow layer rather than the workflow layer itself.

Evaluations, available today. Every anyformat Extract and Classify workflow has a Health tab. You build a dataset with verified ground truth, either by promoting a document the workflow already processed and confirming its values in the review interface, or by uploading labelled documents with their expected JSON. Files are taggable into sub-datasets, so accuracy reads per supplier, per document type or per difficulty slice instead of as one average. An evaluation re-extracts every in-scope document against a chosen workflow version and scores it field by field, with a result-versus-expected view on each failure; runs are numbered and immutable, so the effect of a change is measured rather than assumed. Health Overview then places the dataset benchmark next to live production accuracy, confidence and through rate, and reports the gap between them (Evals). No equivalent surface is documented for Klippa.

Optimizer, announced and not yet available. anyformat has announced Optimizer, a workflow that tunes itself against its own dataset using your evaluations as the target. It is on the roadmap, not in the product today (Evals).

The pipeline you do not have to build. This is the one place where the usual comparison does not transfer cleanly, because Klippa ships an orchestration layer too and it is a real one. The difference is what surrounds it. anyformat pairs Studio workflows with calibrated per-field confidence at 99.1% calibration accuracy, visual citations that link every value back to its position on the page for audit, the evaluation loop above, and a parse then extract architecture built for documents that keep going past 50 pages (anyformat benchmark, 2026). Setting up a production workflow with the Annie assistant takes around 5 minutes against roughly 4 hours of manual configuration (anyformat internal benchmark, 2026).

When to choose Klippa (Doxis AI.dp)

Choose Klippa when identity verification is a first-class requirement rather than an occasional document type: the same engine covers 10,000+ ID document types across 150+ countries with liveness detection and NFC passport reading, which anyformat does not offer. Choose it when you want published self-serve pricing for accounts payable and your volumes fit the SpendControl brackets, because a price list you can read without a sales call is worth something. Choose it if you are a Benelux or DACH buyer who values a vendor with a decade of production deployments and a German parent with an established enterprise content management business behind it, and if EU residency in Germany or the Netherlands under a DPA-based GDPR position meets your requirement. And choose it when your documents are short and standard, such as invoices, receipts, payslips, bank statements and IDs, where the pretrained models apply directly and the 50-page ceiling never becomes relevant.

When to choose anyformat

Choose anyformat when your documents cannot leave your own infrastructure, since air-gapped on-premise and VPC deployment are available and no self-hosted option is documented on the Klippa side. Choose it when documents run long, past 50 pages, with tables of thousands of rows or embedded figures, where the measured result is ~99% row recovery rather than a timeout. Choose it when you have to prove accuracy and keep it: datasets with verified ground truth, sub-dataset slices, immutable evaluation runs and a live comparison of benchmark against production. And choose it when calibrated confidence is the basis for straight-through processing, because a threshold is only as good as the calibration behind it.

These two products overlap more than most pairs in this category, and the honest summary is that European hosting and ISO 27001 will not decide it. Deployment model, document length and whether you need to measure accuracy over time rather than review it document by document usually will.