Setting up document extraction used to mean building templates or training a model on hundreds of examples. Now a language model can do it from a written instruction, and vendors are making bold claims about it. The work that decides whether a document process can be trusted has not changed: knowing the documents, measuring against ground truth, designing the review and keeping every value traceable to its source. The latest news makes that work more valuable, not less.
Generative Models Move Into the Core Platforms
On 20 October 2026, ABBYY announced PHOENIX Plus, a generative AI layer inside its Vantage platform. Output schemas are defined with prompts instead of labelled training data, and customers can connect models from OpenAI, Google and Anthropic. The release places it alongside validation rules, confidence thresholds, human review and source traceability.
For the IDP specialist, the range of extraction approaches gets wider: templates, trained models, pre-built models and now prompt-defined schemas. Matching the approach to the document population, and knowing how each approach fails, matters more than ever.
Accuracy Claims Still Need Your Own Ground Truth
The same release reports one logistics customer extracting handwritten and printed text from complex forms with 99% accuracy on the first pass. The figure may be true for that customer, but the release gives no method, sample or baseline. Is it 99% of fields or 99% of documents? Measured on which documents, against what?
Those are the questions an IDP specialist answers with evidence: field accuracy, document accuracy, straight-through processing and escape rate, measured on a ground-truth set that looks like your own production. Our earlier article, How to Measure Accuracy in Intelligent Document Processing, covers how.
Document Processing Is Becoming a Feature Everywhere
Forrester's Boris Evelson wrote that generative and agentic AI is becoming an equalizer among vendors, and that document capabilities now sit in process automation, robotic process automation, content management, records management and business application platforms. Buyers are reconsidering whether to buy or build.
That shifts the skill from knowing one product to judging any of them. The capture case, built from a sample of the real documents, their types, volumes, channels and quality, is what tells an organization which option fits.
Confident Mistakes Still Need Catching
A language model returns a plausible value even when it is wrong, and a plausible value passes a glance. Confidence thresholds set against the cost of each kind of error, validation rules that catch confident mistakes, and review stations that guard against rubber-stamping are what keep those errors out of downstream systems. Source traceability, every value linked to where it came from on the page, makes each one checkable.
The Data Downstream Depends on Capture
Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. Much of that data starts as a document. Data extracted once and only once, with the source document kept as a record linked to it, is data the next AI project can rely on.
Regulation: Transparency Now, High-Risk Later
The EU's Digital Omnibus, given final approval by the Council on 29 June 2026, moves the AI Act's high-risk obligations for uses such as credit scoring, recruitment and education to 2 December 2027, while transparency obligations applied from 2 August 2026. The new dates take effect once the amendment is published in the EU's Official Journal. Document processing that feeds those decisions, such as reading a loan application or a CV, may fall in scope when the obligations apply. Traceable capture decisions built in now will make that easier, wherever the organization operates.
Where to Start
The Certified AI Intelligent Document Processing Specialist (AI-IDP) program covers this ground: the capture case, extraction approaches including language models, validation and confidence, human review design, measurement against ground truth, and the integration, records and governance that make extracted data trustworthy.
Follow IIAAIT on LinkedIn for new articles.