Gartner cites Altilia in the 2026 Magic Quadrant for IDP. Here is how the platform maps to the market it describes.

On 8 September 2026 Gartner published its Magic Quadrant™ for Intelligent Document Processing. The report evaluates fifteen vendors and cites four further providers as Honorable Mentions: companies it considers noteworthy without having included them in the formal analysis. Altilia is one of the four, and the only Italian company in the report. Rather than frame it as a badge, we take it as the occasion to check the platform against the yardstick Gartner uses: use case by use case, feature by feature.

The yardstick has moved. Gartner now defines IDP as a platform that combines several AI techniques, machine learning, language models and knowledge graphs among them, to capture and enrich data from complex documents and to support workflows across applications. Over the next three years it expects IDP to stop being a capture step in front of a process and to become the intelligence layer that feeds decision-grade data to systems of record and to the AI systems that increasingly consume it. Vendors that only extract, and cannot connect capture to downstream action, are described as exposed to commoditisation. That is a definition of the market we recognise, because it is the one we built the platform for.

Four use cases, one graph

Gartner lists four principal use cases for IDP. Altilia runs all four on the same operating system, and the differences between them are handled by the knowledge graph rather than by four separate products.

Transactional document processing: single, independent documents such as invoices, purchase orders or claims forms. Pre-built agents in Altilia Extract classify, extract and validate at volume; validated data flows to the system of record through Altilia Sheets or APIs; only the exceptions reach the review queue in Notebook.

Interdependent document processing: several related documents that must be processed together, the know-your-customer check being Gartner's example. This is where Altilia has its deepest production record. In corporate credit, agents read a file of 27 document types, populate 250 datapoints, run KYC and ultimate-beneficial-owner checks and reconcile the figures across documents, taking the cycle from five days to two hours. In NPL due diligence, agents processed 130 million pages, extracted 330 datapoints per case and reached 98.3% accuracy on a 10,000-document audit sample. Because every extraction lands in the graph, consistency across documents is a query, not a script.

Contextual intelligence: semantically enriched data that activates knowledge for agents and assistants, technical manuals being the example. This is the use case the knowledge graph was designed for. Documents populate an ontology; retrieval runs over the graph; Notebook answers questions with a citation to the passage the answer comes from; regulatory assistance, internal policies and product documentation become knowledge that other agents can use, not just text that people can search.

Specialised document processing: documents formatted for a technical specialisation, where the layout carries the meaning. Here we do not prompt a general model and hope. Domain ontologies describe what the document contains, and Altilia Model Trainer trains small models on the client's own document families, from court appraisals and cadastral records to the technical annexes of insurance policies.

The four mandatory features

Extraction. Gartner asks for the capture, validation and mapping of entities to structured data. In Altilia, extraction maps to the classes and properties of an ontology, so the output is a typed graph rather than a flat set of fields, and every value keeps a link to the page it came from.

ModelOps. Governing the life cycle of composite AI and decision models. Altilia Model Trainer covers the whole cycle: datasets built with synthetic data generation and human-guided data curation, training, evaluation against the client's own accuracy thresholds, versioning and promotion to production. Models trained on the platform are the customer's property.

Composable architecture. Altilia is an operating system, not a monolith. Robots, Skills and Notebook are its user layers; agents ship as apps (Write, Sheets, Slides, Extract); the knowledge graph, the import layer, the review layer and the model foundry are modules that can be deployed and scaled separately.

Preprocessing. Integration with upstream applications to ingest and prepare content. The import layer collects documents from mailboxes, file shares and document management systems, splits bundles into their constituent documents, classifies them and recovers text and layout from scans.

The sixteen optional features

Gartner's optional features describe what separates a capture tool from a platform. Altilia covers them as follows.

• Postprocessing: validated data is exported to downstream systems through Sheets, APIs and connectors, with the provenance attached.
• Human in the loop: Notebook's review queue contains only what the agents could not resolve, routed by confidence and by the reviewer's expertise.
• Workflow orchestration: document processes are modelled once and executed by agents, deterministically where the rules are known and agentically where they are not.
• Agentic orchestration: agents call skills and other agents to complete a task, and the operating system authorises and logs each step.
• Centralised repository: the knowledge graph is the repository, holding entities, relations and provenance for every document processed.
• Collaboration: reviewers, approvers and subject-matter experts share the same queues in Notebook, with roles that govern who validates and who signs off.
• Data validation: extracted values are cross-checked during extraction against internal data and external registries, as in KYC and beneficial-owner checks.
• Flexible deployment: SaaS, private cloud in the EU or on-premises, with the same platform in each case.
• MLOps: the Trainer pipeline governs machine-learning models as well as language models, and accuracy is monitored in production.
• No/low code: business users describe what a document contains and what an agent should do; the platform takes it from there.
• Other language models: small and domain-specific models trained on the client's data run alongside large models, each bound to the task it fits.
• Prebuilt downstream integrations: connectors to common systems of record and to the apps people already work in.
• Prebuilt upstream integrations: connectors to mailboxes, storage and document management systems for ingestion.
• Rich media capture: native-digital documents, scans and photographed documents, including tables and complex layouts.
• Third-party components: third-party and open-weight models, OCR engines and external services can be plugged in as skills.
• Workflow governance and management: every agent action is permissioned, logged and auditable, and policies are enforced by the operating system, which is how customers meet EU AI Act and DORA obligations.

Two things in the report that most of the market cannot say

The first is the knowledge graph. Gartner includes knowledge graphs in its definition of composite AI, but in the report they appear as one technique among several, not as the foundation of a platform. In Altilia the graph is the foundation. Most IDP tools produce fields; Altilia produces a graph. An ontology says what the enterprise's documents are about, agents populate it, and the graph then does three jobs at once: it makes cross-document reasoning possible, it grounds retrieval so that answers carry citations instead of hallucinations, and it provides the curated knowledge on which large and small models are fine-tuned. The contextual-intelligence use case is not a feature we added; it is the architecture.

The second is sovereignty. The report returns repeatedly to what regulated industries need: sovereign deployment, execution without external connectivity, and the same functionality whether the platform runs in the cloud or in an isolated data centre. Altilia was built for banks, insurers and public administrations in Europe. Models are trained and served inside the customer's perimeter, ontologies, models and agents remain the customer's property, and the platform is certified to ISO/IEC 27001, 27017 and 27018. For a European institution, sovereignty is not a deployment option; it is the reason the platform can be used at all.

Where Gartner says the market is going

Three trends run through the report. Inference layering: routing high-volume tasks to cheaper deterministic or small domain models and reserving frontier models for reasoning. Agentic IDP: agents that decide and act, not only extract. AI in the loop: models that validate, cross-check and pre-correct before a human sees the case. In Altilia each skill is bound to the model that fits its task, agents run whole processes rather than single steps, and validation agents check cross-document consistency before anything reaches the review queue. The direction of the market is the design of the platform.

“Being cited is pleasant. What matters is that Gartner now defines IDP as an intelligence layer built on composite AI, knowledge graphs included, and served under sovereign deployment. That is the platform we have been building since day one.”

Gartner, Magic Quadrant for Intelligent Document Processing, 8 September 2026. Gartner does not endorse any vendor, product or service in its research publications, which express the opinions of its research organisation and are not statements of fact. GARTNER and MAGIC QUADRANT are registered trademarks of Gartner, Inc. and/or its affiliates.