The publisher ABBYY presented on October 20, 2026 PHOENIX Plus, an artificial intelligence component integrated into its document processing platform Vantage. Two years after the launch of PHOENIX, the company is expanding its approach to data extraction from more variable content, as well as their interpretation, while maintaining control mechanisms. Its availability remains limited to certain Vantage clients.
- ABBYY lance PHOENIX Plus, une couche d’IA générative intégrée à Vantage pour traiter les documents.
- La solution extrait des données de documents variables sans entraînement préalable de modèle.
- Elle interprète le contenu et détecte les incohérences, tout en gardant des mécanismes de contrôle.
- Sa disponibilité reste limitée à certains clients Vantage, avec un coût annuel fixe et sans benchmark indépendant.
Companies seek to leverage their documents in automated processes, but invoices, contracts, and forms do not always have a homogeneous structure. ABBYY intends to handle these variations without systematically having to train a specific model. The solution combines classical document processing techniques with generative models, in an environment controlled by the publisher.
An evolution of PHOENIX launched in 2024
PHOENIX was introduced in 2024 during the ABBYY Ascend event. It was originally based on a multimodal approach using small language models (SLM), designed to specialize AI in document tasks. The goal was notably to facilitate data labeling and accelerate the preparation of extraction models. The solution was part of the Purpose-Built AI Center, an environment intended to group the publisher’s AI technologies.
PHOENIX Plus is not simply a new version of this initial model. It fits into a broader architecture, called PHOENIX, which now distinguishes a deterministic document processing core, PHOENIX Core, and a complementary generative layer, PHOENIX Plus. The first groups the historical functions of optical character recognition (OCR), layout analysis, classification, and extraction. The second mobilizes generative models selected and optimized by ABBYY, which intervene when classical rules or pretrained models are insufficient.
Extracting and interpreting without prior training
One of the main announced contributions concerns so-called zero-shot extraction, which should allow retrieving information from unknown or highly variable documents without first creating a dataset to train a dedicated model. According to the press release, in the case of a supplier contract, the sought information may appear in clauses written differently from one document to another. PHOENIX Plus should be able to identify relevant elements based on predefined instructions, then return results according to an expected structure.
The solution can also intervene after an initial extraction to interpret free text, reconcile information, detect inconsistencies, or produce recommendations intended for subsequent steps in a business process. ABBYY thus targets uses that go beyond simple reading and transcription of fields. For the publisher, this hybrid approach makes deterministic techniques coexist, reserved for operations with established rules, and generative techniques, mobilized for content requiring more flexibility. This organization relies on data validation, confidence thresholds, traceability back to original documents, and the possibility of requesting human verification.
ABBYY claims that a client in the logistics sector achieved 99% accuracy during a first pass on complex forms mixing handwritten and printed text. However, this is a result communicated by the provider, which does not allow drawing general conclusions about the product’s performance. The operational interest will also depend on the complexity of the documents, the quality of available data, and the level of verification required.
Hosting, pricing, and availability
PHOENIX Plus is integrated into Vantage and hosted by ABBYY. According to the publisher, data is processed in the same region as the client’s Vantage environment and is not retained by the service after returning the result. Economically, ABBYY plans to add it to the Vantage subscription for a fixed annual cost, to avoid managing a separate contract with a model provider and billing indexed on the number of tokens (units of text processed by the model) consumed. The amount of this option is not specified in the announcement.
Clients can also continue to connect their own models, notably those from OpenAI, Google, Anthropic, or Mistral, thanks to Vantage’s BYOM (Bring Your Own Model) option. In this case, the company chooses its model and must consider the conditions of the chosen provider. PHOENIX Plus, however, is only accessible to certain Vantage clients, as part of a selective deployment, and the publisher invites interested organizations to check their eligibility with its teams. The published information also does not include independent benchmarks allowing comparison with other document extraction tools or generalist models.
