Document automation AI built and hosted in France
Recital is an AI-powered document processing platform for enterprises handling high volumes of business documents.
AI Panel Score
6 AI reviews
Reviewed
Facts checked against recital.ai
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.Recital's platform is used by enterprises to automate document-heavy workflows: incoming emails, PDFs, and scanned files are processed to extract structured data, classify document types, and route information into downstream business systems. Customers cited on the company's site include Safran, ADA, Enership, and ENGIE Solutions, who use the platform for tasks like processing garage repair quotes for rental vehicles, handling quotes and orders received by email, analyzing energy invoices, and automating technical intervention documents (DTA).
The platform is built on proprietary research from Recital's R&D team, which has published work on topics including document layout attention mechanisms (Skim-Attention), synthetic data augmentation for cross-lingual question answering, and generative cooperative networks for text generation. These publications indicate the underlying technology stack relies on custom-trained language and vision models rather than solely wrapping third-party AI APIs. The product integrates with existing enterprise systems, including at least one deployment alongside Kofax TotalAgility for a Safran use case.
Recital targets mid-size to large enterprises across sectors such as insurance, finance, logistics, legal and industry that process large volumes of incoming documents. Pricing is not published on the website and is handled through direct contact with the sales team; the documentation adds that LLM-based tasks carry a per-token surcharge on top of the platform licence.
Recital is available in two deployment modes: as a SaaS product or installed on-premise (On Prem), which allows customers with data residency or security requirements to run the platform within their own infrastructure.
Retrieval-augmented assistant that answers questions, summarizes and analyzes content across a document base.
Uses proprietary AI research (e.g. Skim-Attention) to analyze document layout and extract structured information from documents.
Workflows ingest documents from email (IMAP), the API or scheduled sources, split multi-document bundles, and return results by webhook.
Maintains an in-house R&D lab publishing scientific research on document layout analysis and NLP techniques underlying the platform.
Automates extraction, classification, and processing of invoices, as demonstrated in the Enership energy invoice automation story.
Automates the handling of quotes and orders received by email, as implemented for Safran with Kofax TotalAgility integration.
Automates the processing of garage repair quotes for vehicle fleets, as shown in the ADA automation story.
Secures and automates the processing of technical intervention reports (DTA), as used by ENGIE Solutions.
Offers an on-premise deployment option for enterprises requiring in-house infrastructure.
Offers the document automation platform as a cloud-based SaaS solution.
Sector use cases for insurance (subscriptions, claims), finance (KYC/AML, financial documents), logistics (delivery notes, customs documents), legal (contracts) and industry (technical sheets, production reports).
Integrates with third-party platforms such as Kofax TotalAgility to enhance document workflow automation.
French-hosted document AI backed by real research, priced behind a sales call.
“Recital brings proprietary layout and NLP research to invoice, quote, and report automation for enterprises. Data residency and on-prem options matter more than the average IDP pitch here.”
Founded 2017, Paris. That's not a resume, that's a fact worth noting when a vendor claims deep R&D rather than an API wrapper. Skim-Attention and their cross-lingual QA work back that up — this looks like custom-trained models, not a GPT wrapper with a logo.
Customer stories are concrete: Safran running quotes through Kofax TotalAgility integration, ADA processing garage repair quotes, ENGIE Solutions on technical intervention reports. That's real workflow depth, not a feature list.
Pricing is contact-only, so I can't judge whether the value lands fast or drags through a long enterprise sales cycle. On-prem plus OVH-hosted cloud in France is a genuine differentiator for regulated buyers who can't touch US-hosted alternatives. The tradeoff: that same sovereignty focus narrows the addressable market if you don't care where the servers sit.
Proprietary layout-analysis research and France-based hosting are architectural differentiators in a crowded document-processing market.
On-prem deployment and OVH France hosting address data residency concerns directly, a meaningful signal for regulated sectors like energy and transportation.
No published pricing or trial means procurement friction before any payback starts; sales-cycle length is the bottleneck, not the tech.
Automates document-heavy workflows like invoice and quote processing rather than just cutting cost on existing tools — real advancement for ops-heavy enterprises.
Published research (Skim-Attention, generative cooperative networks), a documented API and a public changelog with a release roughly every two months point to active, shipping technical work.
Mid-size to large enterprises in regulated sectors that need document automation with data residency guarantees.
Skip it if you need self-serve pricing or a free trial before looping in procurement.
Studio IA hands you the model choice per document class — a capability you then have to staff.
“Recital's Studio IA aims at the messy end of document work: mixed bundles that need splitting, plus checkboxes, signatures and QR codes. It lets you pick the model type per use case and train on your own hardware, which is real capability and real obligation at once.”
A bundle lands as one file with several document types inside it. Studio IA identifies the parts, cuts them apart and renames each one, then reads checkboxes, signatures and QR codes alongside the text. Sorting that pile is what stalls a capture programme — reading a clean invoice never was.
The studio also puts the model decision in your hands per use case: extractive, generative or fine-tuned, deep or not, trained on reciTAL's GPUs or on your own hardware. Extraction is quoted at under a second per page. That's an ownership position, not a settings panel.
But choice at that level is work transferred to you. Someone has to own the document taxonomy and decide when a class earns a retrain, a standing job a template-driven capture engine never asks you to staff. Budget the seat, not just the licence.
It competes at the difficult end of document processing — splitting mixed bundles, reading marks — rather than on clean invoice capture.
Intake starts at the mailbox and the scanner rather than an upload form, which is where enterprise document work actually originates.
Product documentation is public at docs.recital.ai and the models are stated to support all languages, which helps a multi-country rollout.
Training on your own GPUs keeps the models under your control, but the tuning calendar becomes a permanent internal responsibility.
Bundle splitting, non-text marks like checkboxes and signatures, and a choice between extractive, generative and fine-tuned models put the craft ceiling above template capture.
Enterprise teams that process mixed document bundles rather than clean digital files.
Teams with nobody to own model tuning after go-live.
No price on the site. Every budget line waits on a sales call.
“No rate card, no seat count, no usage tier — quoting is a black box until sales gets involved. Seven published customer stories, from Safran to Allianz Trade, suggest real deployments, but none of them tell you what you'd pay.”
Contact sales. That's the whole pricing plan. No per-document rate, no per-seat number, no volume tier to anchor a budget against. For a platform pitched at high-volume invoice and quote processing, the absence of a usage-based unit price is the thing procurement will flag first — you can't model TCO on a phone call.
On-prem vs SaaS is a real cost fork, not a footnote. SaaS runs on OVH in France — fine for data residency, but ask who eats the migration cost if you later need dedicated cloud. On-prem buyers should price their own infra and ops overhead into year one, separate from whatever Recital quotes. The docs add one more line item: LLM-based tasks carry a per-token surcharge on top of the platform licence.
Seven published customer stories, from Safran and ENGIE Solutions to Allianz Trade, make a decent proof set. But named logos don't substitute for a number. Until sales gives you a quote, budget owners are pricing blind.
OVH-hosted France infrastructure and on-prem option help EU data-residency procurement, and the docs say LLM tasks are billed per token on top of the platform licence, though no rates are published.
Enterprise sales-led deals usually mean negotiation room, but term length and renewal terms aren't published anywhere I found.
No rate card, no per-seat or per-document price anywhere I could find — contact sales is the only path.
Named use cases like Enership's invoice automation and Safran's Kofax integration give a concrete task to measure time-saved against.
On-prem vs SaaS choice creates two very different cost structures, but neither has a published starting figure to model against.
Enterprises with document volume high enough to justify a custom quote and procurement cycle.
You need a number to put in a budget spreadsheet before your first sales call.
Solid extraction engine for invoice and quote queues, but I want the sandbox before the sales call.
“Recital reads like a document processing shop with real vision-model research behind it, not a GPT wrapper with a nicer UI. The daily fit will depend heavily on your document mix and how well the workflow engine handles exceptions.”
What catches my eye first: the platform is described as fully API-enabled, with a workflow engine, connectors, and business rule integration. That matters for the person who inherits this after IT hands it off — you're not stuck clicking through a queue tool, you can route exceptions programmatically. The named use cases (garage repair quotes for ADA, energy invoices for Enership, DTA reports for ENGIE Solutions) tell me it's tuned for messy, semi-structured business paperwork rather than clean digital-native PDFs.
The RAG layer for querying document bases is the feature I'd actually spend day three on — summarizing and answering questions across a document set is a different job than field extraction, and bundling both means fewer tools open at once.
What I can't assess from the outside is exception handling on the documents that fall below the automation threshold and land in the review queue. That's where every IDP tool either saves you hours or eats your afternoon. Kofax TotalAgility integration on the Safran deployment suggests this slots into existing BPM stacks rather than replacing them, which is the right instinct for enterprise buyers already invested elsewhere.
Named workloads (invoices, garage quotes, DTA reports) suggest real template variety, but exception-handling behavior on messy scans isn't something I could verify.
Docs and an API reference exist and the research publications (Skim-Attention, cross-lingual QA) read like engineer-written material, not marketing copy.
The RAG layer covering summarization and Q&A on top of extraction could reduce tool-switching, but I found no self-serve way to test accuracy on my own document set first.
Business rule integration and a modular RAG solution suggest headroom past basic field extraction, though discoverability of advanced configuration wasn't clear to me.
Kofax TotalAgility integration on the Safran deployment and a described workflow engine with connectors point to fitting inside existing BPM stacks rather than forcing a rip-and-replace.
Enterprise teams processing high volumes of invoices, quotes, or technical reports who need an extraction and RAG layer that plugs into an existing BPM system like Kofax.
Avoid if you need to validate extraction accuracy on your own documents before any commercial conversation starts.
Recital's correction screen sweats the small stuff, and that queue is where your afternoons will go.
“The public docs show a Review queue sorted by agent and a correction screen with an undo on every field. How it handles your own scans stays behind a demo booking.”
Anything under the automation threshold lands in Review, and the correction screen is laid out right: extracted fields down the left, the page on the right with detected zones drawn on it. Each field gets a checkmark, an edit box and a Revert to extracted value arrow. Validate sends it and loads the next. Somebody thought about hour four.
The loop back to training is what matters by month three. Copy to dataset pushes a bad document into Studio IA, and Add page to dataset can prefill annotations from the agent's own models. Fixing a recurring format isn't a blank-page job.
But I'm reading all of this off documentation screenshots. The only door is a demo booking and I couldn't find a trial, so how it handles your ugly scans stays unknown until sales lets you in. And 25.8 kicked off a UI redesign, so these screens may move.
Per-field checkmarks, a Revert to extracted value button and validate-then-next show real care for the daily correction loop.
The user guide walks the Review queue and Studio IA annotation step by step, and annotation prefill shortens retraining work.
I couldn't find a mobile app or mobile guidance in the docs, and the review queue reads as desk work.
Access starts with a demo booking rather than a trial, though the public user guide lets you preview the screens first.
OVH-hosted infrastructure and a documented workflow engine with business rule integration point to production-grade design.
Operations teams who spend their days correcting extracted fields on business documents.
Skip it if you need to test accuracy on your own files before booking a demo.
Seven customer stories back the pitch; the real lock-in is the extraction models you train.
“Seven named customer stories and docs that show the actual correction screen leave less to guesswork than the demo-only front door suggests. The catch is exit: extraction models trained in Studio IA likely won't move to another vendor without retraining.”
Better documented than the sales funnel suggests. The site's only door is a demo booking. The docs, though, show the Production review queue, the per-agent correction list and the correction screen itself: fields left, document right. Daily use is visible.
Seven customer stories. Safran's emailed quotes and orders, with Kofax TotalAgility in the loop. Klian's insurance subscriptions. Allianz Trade at 99.5% automated reading of financial documents, per the vendor's own story, so discount it. Still, named deployments beat adjectives. And the Skim-Attention research says this isn't a thin wrapper.
Exit is the yellow flag. On-prem and OVH hosting in France mean you're not hostage to their cloud. But your extraction logic lives in models trained in Studio IA, and I couldn't find a model-export page in the docs. Leaving for another document-processing platform probably means retraining. Ask before signing.
Research base (Skim-Attention, cross-lingual QA) is a real gap versus document tools that only wrap third-party AI APIs.
On-prem deployment helps, but custom-trained models per the R&D publications mean extraction logic isn't easily portable elsewhere.
2017 founding and an API reference page suggest ongoing operation, and the public changelog shows a release roughly every two months through 2025 and 2026.
Claims about proprietary research are backed by named published work, not just adjectives.
Seven named customer stories, from Safran's emailed orders to Allianz Trade's financial documents, back the automation claims with specifics.
Regulated enterprises that need document automation hosted in France.
Teams who need trained extraction models they can take elsewhere.
Common questions answered by our AI research team
Yes, Recital is offered both as SaaS or On-Prem deployment, giving enterprises flexibility over where the platform runs.
Yes, Recital offers a modular RAG (retrieval-augmented generation) solution to exploit document bases for answering questions, summarizing, analyzing, and exploring content across all document types.
Yes, the Studio IA is designed to handle all use cases including long documents, poor quality documents, handwritten documents, and tables.
Recital's cloud infrastructure is hosted on a shared cloud at OVH in France, with options for dedicated cloud or on-premise deployment.
Yes, the reciTAL platform is fully API-enabled, with authentication, workflow and Swagger reference pages in its public docs, built for production with a robust workflow engine, connectors, and business rule integration.
Who Recital competes with, and how each one differs
Targets large enterprises (insurance, banking, government) with a self-training document automation platform, deployed on-prem or private cloud, versus reciTAL's API-first NLP approach for extraction and search.
Offers a low-code platform for building custom document automation workflows across enterprise back-office functions, with a broader partner/app-ecosystem model compared to reciTAL's more focused NLP extraction and semantic search tooling.
Focuses on unstructured document processing for insurance and financial services with a no-code model training interface, contrasting with reciTAL's developer-oriented NLP API for extraction and classification.
Provides pre-built OCR/extraction APIs for specific document types (invoices, receipts, IDs) at a lower price point aimed at developers and SMBs, versus reciTAL's more customizable NLP-driven extraction and search for enterprise document sets.
A smaller, engineering-focused API startup specializing in parsing complex PDFs and tables into structured data for RAG pipelines, competing on technical depth for document ingestion rather than reciTAL's broader classification and search feature set.
