Altilia vs Dataiku (2026)

F>P, and this is the closest call on the record, so the reasoning is recorded rather than left implicit. ONE FACT MOVED OFF THIS CELL AND IT WAS DOING MOST OF THE WORK. The July basis credited INGESTS ANY DOCUMENT OR DATA SOURCE here, listing PDFs, emails, handwritten notes, scanned images and legacy databases. That is ingestion for grounding and belongs on Know, where the Knowledge Graph carries Full. Reading a source is not integrating with a system the agent then acts in, and this record's real strength is the reading half. WHAT REMAINS IS A CATEGORY WITHOUT MEMBERS. API-first architecture and PRE-BUILT CONNECTORS are stated repeatedly, with agents EMBEDDED INTO YOUR ENTIRE TECH STACK, FROM LEGACY SYSTEMS TO MODERN CLOUD APPLICATIONS. Across two passes covering the home, platform, why-altilia, ai-agents and modules pages, not one connector is named. No CRM, ERP, ticketing or storage system appears, no catalogue exists, and no count is given. THE AXIS MEASURES BREADTH ACROSS CLASSES, and an unenumerated claim of breadth cannot evidence it. This is the same position ai-library was held at earlier today for the same reason, and consistency requires the same answer. WHAT DOES SUPPORT PARTIAL is real: the API-first architecture is a genuine integration route, agents perform actions rather than only answering, and the platform is distributed through the Microsoft commercial marketplace, which implies at least an Azure-side path. I RECORD THIS AS THE CELL MOST LIKELY TO BE UNDERSTATED ON THIS RECORD. A vendor with thirty-plus large enterprise clients in banking, insurance and public administration necessarily connects to core systems; the connector list simply is not published on the pages reached. A partner or integrations page would settle it and is the first check at lane close.

Stands at F, and the orchestration primitive is named rather than described as drag-and-drop. VISUAL SCRIPTING IS THE MECHANISM: USE VISUAL SCRIPTING TO DESIGN AI AGENT WORKFLOWS, AUTOMATING AND SEQUENCING ACTIONS, CONNECTING DATA SOURCES WITH FINE-TUNED MODELS, CHAINS AND PROCESSING FUNCTIONS. Sequencing actions and composing chains against specific fine-tuned models is genuine multi-step construction, and binding a step to a particular model is a level of control most visual builders in this lane do not offer. SKILLS ARE THE VENDOR'S NAME FOR PROCESS ORCHESTRATION TASKS, built in the low-code interface and combined with agents; independent analyst coverage describes the agents-plus-skills combination as providing robotic process automation capabilities. The vendor's own framing of ONTOLOGY-DRIVEN WORKFLOWS ties the orchestration layer to the knowledge graph, so a process step can be defined against business entities rather than against raw data. THE MULTI-AGENT LAYER IS DOCUMENTED as governable agentic orchestration coordinating multiple agents across departments, and the secondary Multi-agent platform categorisation on this record is consistent with that. THE ARCHITECTURAL POINT WORTH CARRYING is the split the vendor makes between assistants and robots inside one orchestration model: a workflow can hand a step to an autonomous agent or route it through a person, which is why the oversight cell and this one describe the same visual surface from different angles. What is not documented is control-flow vocabulary. No branching, looping, conditional or parallel construct is named on any page reached, and no failure or retry behaviour is described. Sequencing and chaining are documented; expressiveness beyond that rests on the visual scripting claim.

Stands at P. The July basis called channel coverage limited and that reading holds. THE EVENT CLASS IS DOCUMENTED THROUGH THE WORK ITSELF. Agents are triggered by document and data ingestion, with the platform ingesting from any source or format including legacy databases, emails, scans and forms. For a document automation platform that is the primary and correct entry route: work arrives because a document did. A SECOND ROUTE IS THE API. An API-first architecture with agents embedded into the customer's applications means an external system can invoke an agent, which is the inbound programmatic class and is credited on Ext as a surface but counts here as a route by which work reaches an agent. WHAT IS ABSENT IS THE REST OF THE SPREAD. No scheduled, cron or recurrence capability is named. No inbound webhook or external event subscription appears as a distinct capability. And no channel layer exists at all: no chat, email, messaging or embedded conversational surface through which a person reaches an agent is documented, beyond Altilia Insights as an in-platform assistant. THAT IS COHERENT FOR THE PRODUCT rather than a deficiency, and the grade should be read that way. This is back-office document and knowledge automation for regulated enterprises, not a conversational front end. Work arrives as documents and results go into systems; there is no customer sitting on WhatsApp waiting for a reply. A vendor with five channels and no knowledge graph would score better on this axis and be worse at the job Altilia does. Confidence medium: no triggers, scheduling or automation page was reached across two passes, and a platform doing overnight batch document processing for thirty enterprise clients almost certainly schedules, so this is likelier undocumented than absent.

Stands at F and is the strongest cell on the record by a clear margin. It is also where several facts removed from other cells correctly land. THE KNOWLEDGE GRAPH IS THE PLATFORM'S CORE, not a retrieval add-on. The vendor states it AGGREGATES AND ORGANIZES ALL INFORMATION WITHIN A COMPANY, UNIFYING DATA FROM EVERY POSSIBLE SOURCE, STRUCTURED, UNSTRUCTURED, AND EVEN PHYSICAL DOCUMENTS, INTO A SINGLE COHERENT DATABASE. A graph is the strongest form of maintained retrieval structure there is, because it encodes relationships between entities rather than similarity between text chunks, and it is queryable by traversal rather than only by nearest neighbour. THE ARCHITECTURE IS HYBRID AND THAT IS THE DIFFERENTIATOR. Large and small language models are combined with the graph in a neuro-symbolic design, described as GRAPHRAG WITH TRAVERSAL REASONING. The symbolic half is what makes an answer traceable to a structure rather than to a probability. INGESTION BREADTH IS EXCEPTIONAL AND NOW SITS HERE rather than on Int: any source or format, including long complex documents, standardised forms, SCANNED IMAGES OR HANDWRITTEN NOTES, emails and legacy databases. Handwriting and scans matter for the regulated European buyers this targets, where the source of truth is frequently paper. CITATION IS DOCUMENTED, which few records in this lane manage: users interrogate the knowledge base in natural language and receive PRECISE, FACT-BASED ANSWERS WITH SOURCE ATTRIBUTION. Attribution is what makes a grounded answer checkable, and it is the property that carried over from Obs where it did not belong. RAG techniques retrieve and fine-tune LLMs and SLMs against enterprise knowledge, and the vendor reports above ninety-five percent accuracy on domain tasks, which is vendor-reported and recorded as such.

Stands at P under the state limb of the 31 August ruling, and the interesting part is a distinction the July basis blurred. CONTINUOUS LEARNING IS NOT MEMORY, and this record is the clearest case for separating them. The vendor documents that THROUGH CONTINUOUS LEARNING AND HUMAN FEEDBACK, AGENTS BECOME INCREASINGLY ACCURATE, and that reviewed corrections feed improvement over time. Retaining corrections across runs is the accumulation limb of the ruling, and on the wording alone this would reach Full. BUT THE MECHANISM IS FINE-TUNING, NOT STATE. The platform page describes creating TRAINING DATASETS from real documents and using RAG techniques to RETRIEVE AND FINE-TUNE LLMS AND SLMS. Corrections are absorbed by updating model weights offline, not written to a store the agent reads at runtime. Those are genuinely different capabilities: a fine-tuned model is better at the task in general; a memory lets an agent recall what happened with this customer last month. Crediting weight updates as memory would let every platform with a feedback loop claim the axis. WHAT IS LEFT IS THIN. Altilia Insights is a conversational assistant, so session context exists within a conversation, which is the ruled definition of Partial. The Knowledge Graph persists and is addressable, but it is the grounding structure and is credited on Know; treating a corpus as agent memory would be the same double-count. NO CROSS-SESSION AGENT STATE IS DOCUMENTED: no memory store, no retention policy, no per-user context, no primitive by which an agent writes state in one run and reads it in another appears on any page reached across two passes. WORTH CARRYING FOR THE LANE: this is the second record this session where a learning loop presented as memory. The test that separates them is whether the artefact is read at inference time or absorbed into weights.

Stands at F, and the mechanism is documented as a workflow stage rather than as a governance principle. THE REVIEW STEP IS EXPLICIT: REVIEW AND VALIDATE DATA CLASSIFICATIONS, EXTRACTIONS AND AI-GENERATED ANSWERS, with the vendor adding INTEGRATE THE FEEDBACK OF HUMAN EXPERTS TO ENHANCE ACCURACY AND CONTROL AT KEY STAGES. At key stages is the phrase that matters: the checkpoint sits at chosen points in the workflow rather than being a blanket setting, which is the same property that earned agentx its Full earlier today. THE ARCHITECTURAL DISTINCTION IS THE STRONGER EVIDENCE. Altilia ships agents in two declared forms, ASSISTANTS that support a human decision and ROBOTS that automate a process autonomously. A platform that makes the autonomy level a first-class choice at design time has built oversight into its model of what an agent is, rather than bolting an approval toggle onto an autonomous default. FOR THIS PRODUCT THE VALIDATION STATION IS THE RIGHT SHAPE. The work is document classification and extraction feeding downstream processes, so the consequential moment is accepting an extracted value, not calling an external tool. A human confirming a classification before it enters the knowledge graph is the approval gate that matters here, and it is where the vendor put it. THE FEEDBACK LOOP IS DOCUMENTED AS CLOSING: reviewed corrections feed continuous learning, so oversight improves the system rather than only catching individual errors. What is not documented is a policy layer: no confidence threshold routing low-certainty items to review, no action allowlist and no spend or scope limit appears on any page reached. Oversight here is placed by design rather than triggered by rule.

F>P, and the vendor's own security paragraph is what decides it. The July basis read the certifications as held; the sentence says something weaker. THE EXACT WORDING: Altilia ensures robust data security through its Integrated Management System, ALIGNED WITH ISO/IEC 27001, ISO/IEC 27017, AND ISO/IEC 27018 STANDARDS. Aligned with is the second rung of the hedge ladder in section 7 and is named there explicitly as Partial. It says the management system was built to those standards; it does not say a certification body audited and issued against them. THE SELF-CONTRADICTION IS INSIDE ONE PARAGRAPH AND ACROSS TWO PAGES. The same sentence continues THESE CERTIFICATIONS VALIDATE, treating alignment as certification, and a separate page describes an ISO-CERTIFIED PLATFORM. Under section 7 the more precise and technical statement wins over the looser marketing one, and the precise statement is the one that names the management system and the three standard numbers. WHY THIS MATTERS RATHER THAN BEING PEDANTRY: for a vendor selling data sovereignty to European regulated buyers, whether an ISO certificate exists is the first procurement question, and the distinction between aligned and certified is exactly what a procurement team is checking. Recording it as certified when the vendor says aligned would put a claim on the index the vendor itself does not make. THE CONTROL HALF IS GENUINELY MET, which is why this is Partial and not lower: ACCESS CONTROLS, ROLE-BASED PERMISSIONS is named, alongside risk management, incident response, and continuous monitoring and audits. Under the conjunction bar, one half documented and one half asserted is Partial. Two facts moved off this cell as belonging elsewhere: data sovereignty and deployment location are graded on Dep, and customer IP ownership of models is a commercial term rather than a security control. No certificate number, issuing body, scope statement or trust page was reached. That is the single check that would move this to Full.

F>P. The July basis cited AI ops management, monitoring hundreds of agents, debugging and observability, and explainable results. Reading the platform page against those, two of the four are the wrong subject and one belongs to another axis. WHAT THE MONITORING ACTUALLY WATCHES IS MODELS, NOT AGENT RUNS. The documented capability is MONITOR AND TRACK MODELS PERFORMANCE TO SPOT INEFFICIENCIES and EASILY MONITOR AND MAINTAIN UP-TO-DATE AI MODELS THROUGHOUT THEIR LIFECYCLE, with automated resource adjustments based on performance and workload. That is model operations, an accuracy-and-lifecycle discipline, and it is a real capability. It answers whether the model is performing well; it does not answer what a given agent did on a given document and why. EXPLAINABILITY BELONGS TO KNOW AND HAS BEEN MOVED. Verifiable answers with SOURCE ATTRIBUTION come from the Knowledge Graph and are a property of the grounding layer. Citing a source is not a record of execution, and counting it here would let the same architectural fact carry two axes. WHAT IS MISSING IS THE WHOLE OF THE AUDITING HALF. No run history, per-document processing record, agent action log, decision trace, retention period or export path appears on any page reached across two passes. THE GAP IS SHARPER THAN THE GRADE SUGGESTS FOR THIS BUYER. Altilia sells to regulated European enterprises processing licences, tax forms, tenders and public registries, and positions on deterministic auditable AI. Those are exactly the workloads where someone will later ask why a document was classified as it was. Model-level accuracy metrics do not answer that for an individual case. PARTIAL RATHER THAN NONE because model performance tracking is real, is customer-facing, and does tell an operator that something has degraded, which is more than several records in this batch offered.

Stands at F and is among the better-evidenced Dep cells in the lane, because deployment control is the company's positioning rather than an enterprise-tier add-on. THREE MODES ARE DOCUMENTED CONSISTENTLY across the home, why-altilia and platform pages: SaaS, private cloud, and on-premises, described by the vendor as DEPLOYMENT SOVEREIGNTY. Repetition across pages matters here; several records this session rested a deployment claim on one sentence in one place. THE PART THAT MAKES IT SUBSTANTIVE RATHER THAN A HOSTING MENU is that the models travel with the deployment. Domain-tuned SLMs and LLMs are SERVED LOCALLY, so an on-premises customer is not running a thin client that ships documents to a hosted model for inference. Model inference is the leak in most agent platforms' residency story, and this one closes it by design. For a buyer processing tax forms and civil registries, that is the difference between a usable option and a nominal one. DATA SOVEREIGNTY IS STATED AS THE PURPOSE, with data kept within the customer's secure infrastructure and customers retaining ownership of data, models and assets. The company's whole market position is a sovereign European operating system for agentic AI competing against US hyperscalers, which is only coherent if the deployment claim is real. Per the 30 August ruling the sovereignty and residency properties are graded here alone and are deliberately not credited again on Sec, where the attestation now sits at Partial on its own evidence. What is not documented is region selection within the SaaS offering, installation requirements, or which components run customer-side in the on-premises mode.

Stands at F under the 31 August bar, which asks whether the customer receives packaged assets ready to adopt. THE PREBUILT SET IS NAMED BY BUSINESS DOCUMENT RATHER THAN BY CAPABILITY, which is the stronger form. The vendor documents DEPLOY READY-TO-USE AGENTS FOR COMMON USE CASES SUCH AS EXTRACTING DATA FROM INVOICES, ORDERS, AND TRANSPORT DOCUMENTS. An invoice agent is a finished job; a toolkit is a starting point, and the difference is exactly what this axis measures. A SECOND, BROADER SET IS DOCUMENTED BY FUNCTION: specialised agents for document classification by semantic context, data extraction from unstructured sources, semantic search and question answering, summarisation into reports, and document generation for emails and operational content. Those are role-shaped rather than component-shaped. THE TWO-ROUTE STRUCTURE IS EXPLICIT AND IS WHY THIS CLEARS FULL RATHER THAN SITTING WHERE THE GENERATION-ONLY RECORDS LANDED THIS SESSION. The vendor states customers either deploy ready-to-use agents or use the IDE to create bespoke ones, so adoption of a packaged asset is a first-class path beside building, not a fallback. Kalcend dropped to None earlier today because generation from a prompt had replaced the catalogue entirely; here both exist. BLUEPRINTS FOR COMMON USE CASES sit alongside, and the vendor's positioning of a digital workforce implies role-shaped packaging rather than parts. What I did not reach is an enumerated gallery, so the size of the ready-to-use set rests on the named examples plus the vendor's description. Three named document types is thin as a catalogue but specific enough to be checkable, which is more than an unenumerated count would be.

Stands at F, and it is unusually well documented for this axis because model choice is the vendor's strategic position rather than a feature. THE BREADTH IS THE FULL RANGE: any commercial or open-source model, large and small, with the vendor's own framing being YOUR MODELS, YOUR CONTROL. Customers retain IP ownership of models trained on their data. THE SMALL-MODEL AND LOCAL-EXECUTION HALF IS THE DISTINCTIVE PART and separates this from a provider menu. Domain-tuned SLMs and LLMs are SERVED LOCALLY, with model selection driven by COST AND DATA TYPE. Choosing a small local model for a high-volume document class and a large hosted one for hard reasoning is a genuine economic lever, and running the small one inside the customer's own infrastructure is what makes it available to buyers who cannot send documents to a hyperscaler at all. FINE-TUNING COMPLETES IT: the platform retrieves and FINE-TUNES LLMS AND SLMS against enterprise knowledge, so the customer is not only choosing among models but shaping the one they choose. Few records in this lane document that. THE STRATEGIC CONTEXT IS WORTH CARRYING because it explains why this cell is strong rather than incidental. Altilia positions as a sovereign European alternative to US hyperscalers, and model portability is the technical substance behind that claim: a platform that could only call OpenAI could not make it. Per the standing convention, local model execution is credited here as customer model control and the deployment location it implies is graded on Dep, not counted twice. What is not documented is a named provider list, a model picker interface, or automatic routing between models at runtime, none of which this axis requires.

Stands at F under Mike's 30 August bar, though it is the least verified of the Full cells on this record and the note should say so. WHAT SUPPORTS IT. The vendor states an API-FIRST ARCHITECTURE, and the direction is the Ext one: agents are CONNECTED DIRECTLY INTO YOUR OPERATIONS AND APPLICATIONS and EMBEDDED INTO YOUR ENTIRE TECH STACK, meaning the customer's own systems invoke Altilia agents rather than the reverse. API-first is a claim about how the platform is built rather than a feature bolted on, and every capability being reachable through the API is what the term means. SIMPLE SDKS ARE NAMED AS A FIRST-CLASS BUILD ROUTE, listed alongside natural language and visual scripting as the three ways a customer creates agents. A vendor offering an SDK as one of three peer entry points is describing a developer surface, not an afterthought. CONFIDENCE IS MEDIUM AND THE GAP IS THE SAME ONE I NAMED ON AGENTX TODAY: no API reference, endpoint list, authentication documentation or SDK package was reached across two passes. The API is documented as an architectural property and a deployment route rather than through a developer surface a reader can inspect. I have graded consistently with agentx, where one-click-to-API carried Full on the same basis, rather than applying a stricter test here. No MCP server was found, which under the ruling does not withhold the grade. THE ASYMMETRY WORTH RECORDING: this vendor documents its knowledge architecture in unusual depth and its developer surface barely at all, which is consistent with selling to business buyers through enterprise engagements rather than to developers. A developer portal is the natural check at lane close, alongside the connector catalogue for Int.

Stands at P under the 31 August Eval bar. The July basis was accurate and the refinement here is about which half is present. THE MEASUREMENT SURFACE IS REAL BUT AIMED AT MODELS. The vendor documents MONITOR AND TRACK MODELS PERFORMANCE TO SPOT INEFFICIENCIES, ENABLE CONTINUOUS LEARNING AND IMPROVEMENT OVER TIME, alongside human review and validation of classifications, extractions and generated answers. Reviewed corrections both catch individual errors and feed the training loop, so quality is measured and acted on. WHY IT DOES NOT REACH FULL. The ruled bar asks for a result the customer can read AND COMPARE about the agent's behaviour on their own work. Nothing documents a test set held aside, expected outputs recorded, a scored run, or a comparison between one version of an agent and another. Model performance tracking tells an operator that accuracy has moved; it does not let them establish that a specific change caused it, which is the question a regression harness answers. THE VENDOR-REPORTED ACCURACY FIGURE IS RECORDED AND NOT CREDITED. Above ninety-five percent on domain-specific tasks is a vendor benchmark, and under the standing convention vendor self-benchmarking is Partial-class evidence at best, the same handling applied to codebuff. THE STRAIN WORTH NAMING for a platform of this kind: the review station generates exactly the material a golden dataset needs, since every human validation is an expected output recorded against a real document. Turning that into a scored regression suite is a small step from what exists, and its absence is more likely undocumented than unbuilt. Confidence medium; no evaluation, testing or debugging page was reached across two passes, and the debugging dashboard referenced in the July note was not found on any page this pass.

Stands at N, and confidence rises to high because a genuine near-miss was found and refused rather than nothing being found. THE NEAR-MISS IS THE STRONGEST ON THIS RECORD AND WOULD CATCH A GRADER SKIMMING. Altilia calls its autonomous agents ROBOTS, distinguishing them from assistants, and independent analyst coverage states that the combination of agents and skills PROVIDES ROBOTIC PROCESS AUTOMATION CAPABILITIES. A vendor whose own vocabulary is robots and RPA, in a market where RPA historically means screen automation, reads at first glance like the positive case for this axis. IT IS NOT, AND THE DISTINCTION IS THE ONE THAT MATTERS FOR THIS LANE. Altilia's robots are autonomous agents that classify, extract, summarise and generate against documents and a knowledge graph, reaching systems through an API-first architecture and connectors. RPA capabilities here means the outcome replaces the manual work RPA was bought for, not that the mechanism is screen driving. No browser control, navigation, form filling, screen interaction, visual grounding or desktop automation appears on any page reached across two passes. COMPUTER VISION IS THE SECOND NEAR-MISS AND IS ALSO REFUSED. The platform uses computer vision and document analysis and recognition to read scanned images and handwritten notes. Reading a document image is intake, credited on Know, and is the opposite of operating an interface: the page is being interpreted, not driven. THE ABSENCE IS STRUCTURAL. A platform whose entire premise is turning unstructured sources into a queryable knowledge graph reaches data by ingestion and API, which makes screen automation unnecessary. WORTH CARRYING TO ENTERPRISE OPERATIONS, where IDP and RPA-adjacent vendors will recur: robots in the vendor's vocabulary and RPA in an analyst's are not evidence of computer use, and the test is whether an interface is being driven.