A Madrid document AI platform that parses, classifies, splits, extracts and validates in one workflow call, and attaches a source citation and a calibrated confidence score to every value it returns.
Overview #
Most extraction APIs in this IDP vendor directory compete on whether a value is right. anyformat competes on whether the platform knows when it is wrong. Each extracted field comes back with the page and bounding box it was read from and a confidence score the vendor says is calibrated, and review is routed per field rather than per document: when one supplier name on an invoice is uncertain, a reviewer checks that field and the other forty pass through. For a buyer, that design decision matters more than any single accuracy percentage, because it determines how much human time survives automation.
The platform is two years old. ANYFORMAT SL was incorporated in Madrid on 27 June 2024 by CEO Juan Huguet García and CTO Diego Pérez Sastre, who both came from the ESG data company Clarity.ai. In November 2025 it closed a €3.3 million seed round led by Kibo Ventures, with 4Founders, Abac Nest Ventures and Decelera Ventures, on top of a €520,000 pre-seed. That release forecast more than €1 million in annual recurring revenue for 2026 and listed ISO 27001 as a goal; the homepage now lists the certification as held.
The company positions itself as enterprise document automation for European companies, and the buyer it is built for is a finance or operations team that needs EU data residency or its own infrastructure. Its closest comparisons are developer-facing extraction platforms such as Reducto, Extend and LlamaParse, with European rivals Klippa and Konfuzio competing on the data-sovereignty argument.
Each value carries its source and a calibrated confidence #
A single workflow chains up to six operations. Parsing turns PDFs, images and scans into structured Markdown with layout detection, which is the OCR stage. Classification sorts documents into customer-defined categories, and splitting separates bundled files into individual documents, the segmentation problem that breaks many invoice pipelines. Extraction fills a typed schema zero-shot, with no labeling or model training, and validation checks the results against rules written in plain language.
Calibration is the claim that sets anyformat apart. A calibrated system that reports 90% confidence is right about 90% of the time, which is what makes a threshold for human review meaningful. The vendor reports an adaptive expected calibration error of 0.009, which it presents as 99.1% calibration accuracy. The practical output is the L'Oréal invoice deployment: the time saved came less from faster extraction than from no longer checking values that were already correct.
The company does not publish which models run inside the pipeline. Its own benchmark scores Gemini, Claude and GPT models as competitors, which says nothing either way about whether those models sit behind the API.
The benchmark is the vendor's own, and the parsing lead is thin #
anyformat publishes its accuracy case in a July 2026 benchmark post written by its CEO. The parsing test covers more than 1,000 real documents drawn from an MIT-licensed open-source PDF sample, with ground truth the team rebuilt by hand in Markdown. The results against eleven systems:
| System | Parse Score | Cost per 1,000 pages (as reported) |
|---|---|---|
| anyformat | 78.1% | $25 |
| Gemini 3.5 Flash | 77.9% | $31.33 |
| Claude Opus 4.8 | 77.3% | $83.50 |
| GPT-5.6 | 74.2% | $82.15 |
| GPT-4.1 | 72.3% | $14.43 |
| Mistral OCR 4 | 70.5% | $4 |
| [Azure Document Intelligence](../../vendors/microsoft/) | 69.9% | $10 |
| [AWS Textract](../../vendors/textract/) | 65.4% | $15 |
| Google Document AI | 55.0% | $1.50 | | Docling | 52.3% | Self-hosted | | PaddleOCR | 43.2% | Self-hosted |
On parsing, anyformat trails nobody but leads Gemini 3.5 Flash by 0.2 points, a margin a different document sample could erase. The case is stronger on long and complex documents. On nearly 100 mixed-format invoices, anyformat captured every field and line item on 80% of documents of 16 pages or more, against 53% for GPT-5.6 and 27% for both Gemini 3.5 Flash and Claude Opus 4.8. On a single table grown to 50 pages and about 2,400 rows, it reports roughly 99% row recovery while the general-purpose models degraded well before the end.
Two caveats apply. The invoice set is small, and every number comes from the vendor running its own test. At publication the improved dataset and scoring methodology had not been released; the post promises both, and until they appear the results cannot be reproduced. The more useful number for a buyer is the one the platform's datasets and evals feature produces on their own documents.
Developers get workflows, typed SDKs and a Claude skill #
The API is organised around workflows rather than single endpoints. A workflow is created once at POST /v3/workflows/, documents are sent to POST /v3/workflows/{id}/upload/run/, and results are polled at GET /v3/runs/{id}/. Each result carries the extracted value, its confidence, and the page number and bounding box it came from. The same workflows can be built in the web Studio without code, and the Business plan adds evaluations and a monitoring dashboard on top.
The client libraries are young but moving fast. The Python SDK shipped its first release on 27 March 2026 and reached a 1.0.0 release candidate on 20 July, twelve versions later. The TypeScript SDK sits at 0.2.3, and there is a Python CLI. A separate npm package, @anyformat/skill, installs a Claude skill that builds and runs extraction workflows against the API, and the documentation covers MCP access for coding agents including Claude, Codex and Gemini. That puts anyformat among the vendors, alongside Unstract and LandingAI, treating coding agents as a primary integration channel rather than an afterthought.
Use cases #
Accounts payable across European entities
The flagship reference is L'Oréal. According to the vendor's invoice-processing page, the deployment covers more than 1,500 invoices a month from hundreds of suppliers across several European entities, at 99% extraction accuracy and 60% less processing time than the 38 hours a week of manual entry it replaced. New supplier formats are handled without configuration, and only low-confidence fields reach the finance team. The figures are vendor-published; L'Oréal has not released its own account. Invoice-heavy teams will also want to compare Rossum and ABBYY, both of which anyformat names as alternatives.
Spanish tax documents
anyformat's home market shows in the details. A July 2026 post by the CTO covers extracting IVA (Spanish VAT), IRPF withholding and NIF and CIF tax identifiers from Spanish invoices, processed with EU data residency under GDPR, where global parsers tend to return the text but miss what the fields mean.
Logistics, KYC, healthcare, real estate and legal
The homepage lists all five as target industries beyond accounts payable. Named logos include Iberia, whose parent IAG Group appears in the funding release, the Singapore Prison Service, the Spanish logistics firm Adea, auto-parts distributor Recalvi and Grupo Fire. None of these has a published case study with numbers yet.
Pricing is metered in credits the vendor does not define #
The pricing page has three tiers. Pay as you go is free to start with 50,000 signup credits and no card, then €0.0015 per credit, capped at one user, five development workflows, two production workflows and 15 MB files. Business costs €499 a month, or 20% less billed annually, and includes 500,000 credits a month with overage at €0.001, unlimited users, evaluations, monitoring, a private Slack support channel and 25 MB files. Enterprise is custom and is the only tier with self-hosted deployment, custom MSA, DPA and SLA terms, advanced role-based access control, custom models and audit logs.
The pricing page does not say what a credit buys. The benchmark prices anyformat at $25 per 1,000 pages, so a buyer can't convert a monthly credit allowance into pages from public information. Ask for the credit cost of parsing and extracting a page, which almost certainly differs by operation, before comparing the plan with per-page competitors.
EU residency and self-hosting, with a certificate to request #
anyformat's security positioning has three parts: ISO 27001 certification, EU data residency, and deployment in its cloud, a dedicated VPC, or fully on the customer's own infrastructure. It also states a zero data retention policy. The certification appears on the homepage and pricing page; the November 2025 funding release still described it as a goal, so it was obtained within the last ten months. The public Trust Center did not expose the certificate scope, issuing body or a subprocessor list when checked in September 2026. Those three documents, plus which model providers process customer data in the hosted service, are the right requests for a procurement review.
Technical specifications #
| Feature | Specification |
|---|---|
| Operations | Parse, extract, classify, split, validate, orchestrated in one workflow call |
| Parsing output | Structured Markdown with layout detection |
| Extraction | Zero-shot typed schemas, no labeling or training |
| Evidence | Page number and bounding box citation per value |
| Confidence | Calibrated per field, Adaptive ECE 0.009 (vendor-measured) |
| Human-in-the-loop | Field-level routing of low-confidence values |
| Validation | Rules written in plain language |
| Input | PDFs, images, scans; 15 MB (free), 25 MB (Business), custom (Enterprise) |
| API | REST v3, workflow-based ( /v3/workflows/ ,/v3/runs/ ) |
| SDKs and tools | Python ( anyformat , 1.0.0rc1), TypeScript (@anyformat/sdk , 0.2.3), Python CLI |
| Agent access | MCP for coding agents; @anyformat/skill Claude skill |
| Studio | No-code workflow builder, datasets and evals, monitoring dashboard |
| Deployment | anyformat cloud, dedicated VPC, self-hosted (Enterprise) |
| Compliance | ISO 27001, GDPR, EU data residency, zero retention |
| Concurrency | 2 jobs (free), 10 (Business), custom (Enterprise) |
| Model providers | Not disclosed |
| Pricing | Free with 50,000 credits; €499 per month Business; Enterprise custom |
| Benchmark | Vendor-run, published July 2026; dataset release pending at publication |
Resources #
Company information #
ANYFORMAT SL
Calle María de Molina 41
28006 Madrid, Spain
CIF B19820828, incorporated 27 June 2024
Founders: Juan Huguet García (CEO) and Diego Pérez Sastre (CTO) Investors: Kibo Ventures, 4Founders Capital, Abac Nest Ventures, Decelera Ventures
info@anyformat.ai
anyformat.ai
Sources disagree on the date of the €520,000 pre-seed: the company's seed release places it in June 2025, while a Spanish startup database dates it to October 2024.