Use of artificial intelligence | Docs

Introduction

Artificial intelligence has been part of Kabeen from the start, but always in service of a single goal: automating the documentation of your information system. Mapping an IT landscape is long, repetitive work that is rarely up to date. AI lets us produce that documentation from real usage — no questionnaires, no inventory workshops, no spreadsheet to maintain.

This article describes precisely what AI does in the platform, which data is sent to it and how you stay in control.

Our philosophy

AI documents, it does not decide

Kabeen AI produces proposals: an application candidate for the inventory, a business feature likely in use, an analysis summary. Each proposal comes with a confidence score and remains subject to your validation. The decision — accept, reject, correct — always belongs to a human.

The minimum data, never more

We apply a simple rule: an AI process receives only the metadata strictly required to produce its result, never the content of your information system at large. In practice that means URL paths and page titles rather than application content, usage aggregates rather than nominative logs, and bounded excerpts rather than full exports.

Whenever a process can be performed without calling a language model, it is. Application auto-discovery and team auto-discovery, for example, are entirely algorithmic: no data leaves the platform for either of them.

Reversibility and transparency

Every Kabeen AI function can be enabled and disabled independently from the workspace settings. None of them is a prerequisite for the platform to work: turning AI off reduces the level of automation, never the availability of your data.

No training on your data

Data sent to AI providers is sent as part of an on-demand processing request. No data from your workspace is used to train models, neither by Kabeen nor by its providers.

Choosing the AI engine

Three providers, one at a time

Kabeen relies on three language-model providers:

Provider	Location	Status
Mistral AI	France / European Union	Default provider
OpenAI	United States	On selection
Anthropic	United States	On selection

The provider is chosen at workspace level, in the Kabeen AI settings, by an administrator holding the workspace administration right. By default, a workspace is configured on Mistral AI, our French provider.

The choice is watertight

The selected provider applies to all AI processing in the workspace. If your workspace is set to Mistral, only Mistral models are called: there is no fallback, no switchover and no cross-provider call to another vendor's models under any circumstances — including an outage. A failing process fails honestly rather than routing your data elsewhere.

The same isolation applies to requests: only workspace administrators may pick a provider. A request crafted by an unauthorised user is ignored and falls back to the provider configured for the workspace.

A mix of models within the chosen provider

Within a given provider, Kabeen does not use a single model: we select the model suited to each process, which keeps both consumption and latency down.

Conversational and reasoning workloads (Prism) use the most capable model in the provider's range.
Background enrichment workloads (feature auto-discovery, deep analysis, conversation title generation, history compaction) use light, fast models.

These models evolve along with our providers' ranges, but always within the provider selected for your workspace.

AI processing today

Overview

Process	Calls an AI model	Data sent
Application auto-discovery	No — algorithmic	None
Team auto-discovery	No — statistical	None
Feature auto-discovery	Yes — light model	URL metadata, page titles, usage aggregates, capability catalogue
Prism assistant	Yes — reasoning model	Question, conversation history, page context, tool results
Deep analysis (Prism)	Yes — light models	Application records within the requested scope
Web search (Prism)	No — search service	The search query only

Application auto-discovery — no AI model

The application inventory is built by algorithmic matching: every URL collected by the agents (domain and path only) is compared against the Kabeen catalogue of more than 80,000 professional applications. Unmatched URLs are then screened by shape heuristics (app., cloud., portal., admin. subdomains, /app, /dashboard, /portal paths…) to tell a professional application from a consumer website.

No call to an AI provider is made for this process. The data stays entirely inside the Kabeen infrastructure.

See Application auto-discovery for the functional details.

Team auto-discovery — no AI model

Linking an application to the teams that use it is derived statistically from observed usage and from your workspace's organisational tree. Here again, no data is sent to an AI provider.

Feature auto-discovery — the only enrichment that calls a model

This is the process that turns an application inventory into a map of the business capabilities actually exercised. It is the platform's only automatic enrichment that calls a language model.

What is sent to the model, for one application:

The record metadata: name, description, category, vendor, access URL — each truncated to 200 characters.
A sample of observed URL paths, capped at 100 entries, each path truncated to 500 characters.
The associated page titles, truncated to 100 characters.
Usage aggregates over the last 90 days: number of views, number of distinct users, number of distinct paths. These are counters, not identities.
Your organisation's functional capability catalogue, which is the closed list the model must choose from.

What is never sent: page content, query-string parameters, user identities, session identifiers, data entered in the application.

The protections applied:

Input sanitisation: control characters removed, prompt-injection sequences neutralised, systematic truncation. An application whose name or URLs contained disguised instructions cannot hijack the analysis.
Output validation: only capabilities matching an entry in your catalogue exactly are kept; off-catalogue proposals are discarded. Regular expressions produced by the model are checked before being stored.
Silent degradation: on failure, the process returns an empty result. It is never blocking, and does not retry indefinitely.
Light model: this process uses the selected provider's small model, not the conversational one.

See Feature auto-discovery for the functional details.

Prism assistant

Prism is the platform's conversational agent. It is the richest AI process, and therefore the one whose data scope deserves the most precision.

What is sent to the model on each conversation turn:

Your message.
The conversation history, capped at 50 messages. Beyond that, the oldest turns are replaced by a summary produced by a light model, so the full exchange is not resent indefinitely.
The context of the page you opened Prism from, reduced to a closed list of items: entity type, the entity's technical identifier and active filters from an allowed list (status, category, criticality, organisational scope, period…). Nothing else from the page is sent.
Your assistant memory: a preferences file you populate explicitly ("I always reason by subsidiary", "show costs including tax"), capped at 4,000 characters and attached to your account alone.
The definition of the tools available for the turn, restricted to the relevant domains rather than the full catalogue.
The results of the tools called — that is, the data from your workspace needed to answer. Each result is capped at roughly 8,000 characters: over-long lists are truncated, and the model is told how many items were omitted. Prism therefore never receives a full export of your repository.

Data isolation:

Every tool runs with your workspace identifier and your permissions. Data your account cannot see in the interface is equally invisible to Prism: it bypasses no access right.
The workspace's identity-tracking mode applies. In anonymous mode, Prism handles counts, not names.
Every write action (creation, update, linking) is recorded in the audit trail with author, timestamp and context. Destructive actions require explicit confirmation.
Prism only speaks from your workspace's data. When information does not exist, it says so rather than inventing it.

Two derived processes:

Deep analysis answers portfolio-wide questions ("which applications are candidates for rationalisation?"). It launches one bounded sub-agent per application — 25 applications at most, light model, read-only — and returns only a condensed summary. This avoids sending the whole portfolio into a single context.
Web search lets Prism qualify a vendor or a technology from public sources. Only the search query is sent to the search service: none of your workspace data is attached to it.

Retention: your conversations are stored in your workspace so you can find them again. You can delete a conversation, or resume the thread from a given message, at any time.

Technical traceability: each exchange generates a technical trace (question, answer, tools called, token consumption) in our observability platform hosted in the European Union. It is used solely for diagnostics and quality of service.

See Prism assistant for the functional details.

What is never sent to an AI engine

Whatever the process and whatever the selected provider, the following never leaves the platform:

The content of the web pages visited (text, images, forms)
Credentials, passwords and authentication tokens
Keystrokes and screenshots
Files opened or modified on workstations
URL query-string parameters
Conversations and messages from your business tools

This list follows from collection itself: those data are not collected by the Kabeen agents, so they cannot be sent. See Data Privacy.

Staying in control

From your workspace's Kabeen AI settings, an administrator can:

Enable or disable independently the Prism assistant, application auto-discovery, team auto-discovery and feature auto-discovery.
Choose the AI provider applied to all processing in the workspace.

Beyond those settings, data visibility remains driven by roles and permissions and by the workspace's identity-tracking mode, which apply to AI processing as they do to the rest of the platform.

See Roles and permissions and Data Privacy.

Upcoming processing

Kabeen's AI roadmap follows the same line: automate more of the IT documentation, without widening the scope of shared data. Work in progress covers automatic qualification of application flows, detection of functional redundancy across the portfolio, and automatic enrichment of records from public sources.

Each of these will follow the same principles: metadata only, a model suited to the need, a single provider chosen by the workspace, proposals subject to human validation, reversible activation. This article will be updated as those processes go live.