Faced with the temptation to plug a generative AI model directly into the entirety of a company's documents, Raphael Godinot, Partner Account Manager (ABBYY) and Olivier Vasseur, Solution Architect (ABBYY) came to the Tech For Industry Show to warn of the risks of this approach and present an alternative: trusted document AI.

The opening message is clear: using a generative AI chatbot to ingest documents and extract information may seem appealing, but it carries several concrete risks. First, cost: token-based consumption billing can quickly become very expensive at enterprise scale. Second, a reliability risk: a generative model remains inherently probabilistic, always carrying some possibility of drift or hallucination — a significant risk when dealing with critical industrial documents. Finally, a traceability risk: it's often hard to explain precisely where a piece of information extracted by a general-purpose LLM actually came from.

ABBYY proposes a different approach: specialized, deterministic AI, meaning algorithms that produce reproducible results — the same document processed twice will yield exactly the same result, which isn't guaranteed with a solution relying solely on generative AI.

This trusted AI rests on several qualities: it's explainable end-to-end (every processing step can be reconstructed and justified), auditable and compliant (an audit trail can be presented to a regulatory body), and frugal, since ABBYY's solutions run on CPU rather than GPU — a point presented as increasingly relevant given growing scrutiny of AI's energy consumption.

The demonstration details how the ABBYY Vantage platform pipeline works: image enhancement, OCR/ICR (recognition of printed and handwritten characters), document classification, data extraction (combining business rules, deep learning, few-shot learning and prompts), then control and traceability of each extraction — with the ability to know the exact location of the information in the source document.

A concrete example illustrates the hybrid approach: on a rail freight invoice, the platform first extracts standard fields (number, date), then identifies a specific text zone (a note about missing wagons). Rather than sending the entire document to an LLM, only that precise zone is passed on for analysis — which limits token consumption, keeps context under control, and preserves the ability to precisely pinpoint the origin of each piece of extracted information.

ABBYY positions its platform as a universal industrial mail room automation tool: whatever the input channel, documents are processed through Vantage, then the results (and source documents converted to PDF/A) are sent to enterprise systems (notably the ERP) via APIs and connectors — always with the option of involving a human operator in case of exceptions, through configurable confidence thresholds.

The session presents a series of examples of documents the platform can already process without custom development: logistics documents (international consignment notes, bills of lading, certificates of origin), energy performance diagnostics (a semi-structured document that varies from one agency to another), purchase orders, technical drawings mixing printed text and handwritten annotations, resumes, payslips — with integration into a document fraud detection platform capable of identifying visual or font manipulations — and energy savings certificates.

Another highlighted feature: the ability to process a single PDF file combining several distinct documents (for example, a supplier file scanned all at once), by automatically detecting the boundaries between documents, removing separator sheets, reordering pages if necessary, then applying the appropriate extraction model to each sub-document — with the option of applying consistency rules across documents.

ABBYY presents its history: inventor of the first OCR technology 40 years ago (technology licensed to numerous scanner manufacturers and software vendors under the FineReader name), more than 10,000 projects delivered, ISO 9001 certification, and leader status in Gartner's Intelligent Document Processing (IDP) ranking. The platform is offered in SaaS mode or on-premise deployment, with a low-code/no-code interface letting business teams configure document automation without needing a data scientist.

The session also mentions a new document format, presented as the result of a collaboration between several major AI players (including Nvidia, Red Hat and IBM) and ABBYY's research lab, specifically designed to be quickly and efficiently digested by AI applications — much like what Word, PDF and HTML once represented for publishing, printing and web publishing. According to the benchmark cited during the session, this format would be around three times more efficient to process than traditional formats, while natively embedding a governance layer (identifying personal data that should not be used, for example).

ABBYY cites clients including groups such as BNP Paribas, CMA CGM, Thales, Gerflor and Saint-Gobain (processing orders and acknowledgments of receipt for the group's distribution business), as well as projects run with France Travail on resume processing. The company positions itself as a complementary player to major generative AI platforms rather than a competitor: some clients, such as CMA CGM, combine a solution like Mistral AI with ABBYY Vantage for their transport and logistics documents.

The central message of this session is a call for methodological caution: rather than entrusting the entirety of document processing to a general-purpose generative AI, ABBYY makes the case for a hybrid approach, where specialized, deterministic AI handles reliable, traceable and auditable information extraction, before an LLM is mobilized in a targeted and controlled way. An approach aimed at reconciling reliability, compliance, cost control and energy efficiency.

Watch the full session, along with the other Tech For Industry Show replays, on our dedicated page.