Kortex

Kortex · Governed AI on your data

Your AI assistants, answering with your company's data. Without handing them the database.

Kortex connects claude.ai, ChatGPT, Copilot and the AI tools your team already uses to your data warehouse. Everyone asks with their corporate identity, sees only what their role allows, and every query is audited. Your data never leaves your infrastructure.

Book a demo How it works

Works withclaude.aiChatGPTCopilot StudioClaude CodeCursor

Currently being implemented at a Chilean pension fund (AFP).

Conversación en claude.ai donde Kortex responde la facturación por país, filtrada a la región EMEA

A real query in claude.ai: Kortex automatically applies the EMEA region filter from the user's role.

The problem

AI is already in your company. Your data isn't yet.

Spreadsheets pasted into chats

Without a secure connection, people copy sensitive data into AI tools, beyond any control.

Shared accounts

An integration with a single service account sees everything, and nobody knows who asked what.

Freewheeling SQL

Letting AI write free-form SQL produces made-up joins, inconsistent KPIs and answers nobody can defend.

How it works

A governance layer between AI and your data.

AI assistants

claude.ai, ChatGPT, Copilot Studio, Claude Desktop, Claude Code, Cursor

Your infrastructure

Kortex

Identity, permissions, semantic layer and auditing

Your data warehouse

Read-only user, aggregate queries

Question

The person asks in plain language, in the assistant they already use.

Identity

Kortex verifies who they are through corporate SSO: Entra ID or Google.

Permissions

It applies their roles: which metrics and which rows they can see, server-side, on every query.

Governed query

It translates the question into an aggregate query over the metrics defined by your data team.

Audited answer

It returns the result to the assistant and logs who queried what, when and from where.

Real permissions

Same question. The answer each person is entitled to.

"How are we doing this month?"

EMEA Sales

Sales and orders, for their region only.

Finance

Revenue, margins and costs across the whole company.

Regional HR

Headcount and absenteeism for their region, no sales data.

No role assigned

Nothing. Having a directory account isn't enough.

Semantic layer and cubes

Your business, defined once. In cubes AI understands.

Kortex doesn't let AI guess at tables. Your data team defines cubes with the business's metrics, dimensions and relationships, and every assistant answers using those same definitions.

Cubes: business metrics, not raw tables

Each cube describes part of the business in its own language: what counts as revenue, how margin is calculated, what an active customer is. Joins and calculations are defined once, and the agent only chooses which metric to request.

Measures, dimensions and relationships with names and descriptions
Views that combine cubes for each business area
A single definition of every KPI across all assistants

Pre-aggregations: from minutes to seconds

Cubes can declare pre-aggregations that Kortex materializes and keeps up to date inside your infrastructure. Common questions are answered from those aggregates, without scanning millions of warehouse rows: the agent gets in seconds what would otherwise take minutes.

Refreshed automatically at the frequency you set
Less load on your data warehouse
Permissions still apply to the aggregates

model/ventas.yml

cubes:
  - name: ventas
    sql_table: dwh.ventas
    description: Ventas netas, sin impuestos
    measures:
      - name: facturacion
        type: sum
        sql: monto_neto
    dimensions:
      - name: region
        type: string
    pre_aggregations:
      - name: ventas_por_region_mes
        measures: [facturacion]
        dimensions: [region]
        granularity: month
        refresh_key: {every: 1 hour}

Defined in files, in your own Git

The semantic layer lives in YAML configuration files, in a repository your company owns. Your team edits it like any code: with pull requests, review and automated validation. Every approved change reaches production in seconds, with no restarts.

Open, readable format, no proprietary languages
Full history of who changed each metric and why
Separate branches for development and production

Excel and BI tool connectivityComing soon

The same governed metrics and the same permissions, from Excel and your BI tools too.

Why Kortex

Better than wiring AI straight to the database. More open than a platform semantic layer.

	AI wired directly to the database	Platform semantic layer (Microsoft Fabric or others)	Kortex
Where your data lives	In your database, but the AI can read everything the account can see	Usually has to be moved or replicated to the vendor's cloud	In your warehouse. Kortex runs on your infrastructure and copies no data
Who sees what	One shared service account, no individual identity	Per-person permissions, within the vendor's ecosystem	Your corporate identity, with metric- and row-level permissions on every query
Consistent answers	Freewheeling SQL, made-up joins, different KPIs every time	Metrics defined in the semantic model	Curated metrics in cubes, identical across all assistants
Speed	Heavy queries on raw tables: minutes	Depends on purchased capacity	Pre-aggregations on your infrastructure: seconds
Supported assistants	Whichever you integrate, one at a time	Built mainly for the vendor's own assistant	claude.ai, ChatGPT, Copilot, Claude Code, Cursor and more, via open standards
Semantic model	None: the AI interprets the schema	In the vendor's format and tools	Open YAML files, in your own Git
Adoption	Immediate, but uncontrolled	Often requires migrating data and adopting the whole platform	On your current warehouse, in weeks

Where your data lives

AI wired directly to the database: In your database, but the AI can read everything the account can see
Platform semantic layer (Microsoft Fabric or others): Usually has to be moved or replicated to the vendor's cloud
Kortex: In your warehouse. Kortex runs on your infrastructure and copies no data

Who sees what

AI wired directly to the database: One shared service account, no individual identity
Platform semantic layer (Microsoft Fabric or others): Per-person permissions, within the vendor's ecosystem
Kortex: Your corporate identity, with metric- and row-level permissions on every query

Consistent answers

AI wired directly to the database: Freewheeling SQL, made-up joins, different KPIs every time
Platform semantic layer (Microsoft Fabric or others): Metrics defined in the semantic model
Kortex: Curated metrics in cubes, identical across all assistants

Speed

AI wired directly to the database: Heavy queries on raw tables: minutes
Platform semantic layer (Microsoft Fabric or others): Depends on purchased capacity
Kortex: Pre-aggregations on your infrastructure: seconds

Supported assistants

AI wired directly to the database: Whichever you integrate, one at a time
Platform semantic layer (Microsoft Fabric or others): Built mainly for the vendor's own assistant
Kortex: claude.ai, ChatGPT, Copilot, Claude Code, Cursor and more, via open standards

Semantic model

AI wired directly to the database: None: the AI interprets the schema
Platform semantic layer (Microsoft Fabric or others): In the vendor's format and tools
Kortex: Open YAML files, in your own Git

Adoption

AI wired directly to the database: Immediate, but uncontrolled
Platform semantic layer (Microsoft Fabric or others): Often requires migrating data and adopting the whole platform
Kortex: On your current warehouse, in weeks

General comparison. Each platform's capabilities vary by product and license tier.

The console

Who sees what, always in view.

Your team manages access, roles and filters from a web console, and can answer the question auditors care about most in seconds.

Matriz de accesos de Kortex: usuarios por cubo, con sus roles y filtros de fila

Access matrix: every user, their roles, how much of each cube they see and which row filters apply to them.

Filtros de fila de un rol: región igual a EMEA en cada cubo

Row filters by role: anyone with EMEA Sales only sees rows from their region, across every cube.

Búsqueda inversa: qué roles y usuarios pueden ver la facturación

Reverse lookup: which roles and people can see revenue, and with which filters.

Capabilities

Everything IT needs to say yes.

The assistants they already use

A single URL connects Kortex to claude.ai, ChatGPT, Claude Desktop, Claude Code, Cursor and VS Code. For custom GPTs and Copilot Studio, actions with a ready-to-paste OpenAPI spec.

Individual corporate identity

Everyone signs in with their Entra ID or Google account. No shared accounts or local passwords.

Metric- and row-level permissions

Define per role which measures and dimensions are visible and which rows, e.g. region = EMEA. Enforced server-side, so the agent can't bypass them.

Governed KPIs

Answers come from a semantic layer with curated business definitions, versioned in your Git with review and automated validation.

Safe queries

No free-form SQL: only aggregate queries over the semantic layer, with a read-only database user.

Semantic metric search

The agent finds the right metric by meaning, in multiple languages, and only among those the person is allowed to see.

Full audit trail

Every query logs who, when, from which app, which metrics, which filters and with what outcome. Returned data is never stored.

Access reviews

A who-can-see-what matrix, reverse lookup by metric ("who can see revenue?") and effective permissions per person.

Admin console

Users, roles, connected apps, SSO, data sources and auditing in one web console. Changes apply without restarts.

Connects to your data warehouse

PostgreSQL
Amazon Redshift
Snowflake
BigQuery
SQL Server
Databricks
MySQL
ClickHouse
Athena
Trino
Oracle
DuckDB

For IT and security

Your data, always under your control.

Data never leaves your company

Kortex is installed in your cloud or on your servers. It doesn't copy or index business data: only the model's names and descriptions.

Dataforge has no access to your infrastructure

No remote access, agents or tunnels. Your team decides when to install each version.

Kortex never calls an AI model

Agents call Kortex, not the other way around. No outbound access to AI providers required.

Encrypted credentials

AES-256-GCM, with the master key kept outside the database. A backup reveals no passwords.

Closed by default

No role means no access. The admin console only responds from your network's IP ranges.

No telemetry unless you opt in

If you enable it, only counts and latencies are sent, and you see the exact payload before it goes out.

Deployment options

In your AWS VPC with a CloudFormation template: no SSH, encrypted disk, access via Session Manager
On your servers with Docker Compose
VPN-only, for desktop clients
Air-gapped, with manually loaded images

Implementation

From installation to production in weeks.

Half a day

Installation

A script that only asks for two DNS names and an email address. SSO and the warehouse connection are configured from the console.

1 to 2 weeks

Roles and testing

Roles are defined and tested with real users in a development environment.

Production

Agents connected

The setup is replicated in production, everyone connects their assistant and administrators are trained.

< 10 s

for a semantic model change to reach production, no restart needed

~1 min

to upgrade versions, with a prior backup and rollback

90 days

of audit history by default, with activity charts and per-person summaries

FAQ

Kortex FAQ

Only the metrics their roles allow and only the rows their filters allow. No role, no access. Permissions are checked server-side on every query.

Bring AI to your data without losing control.

We'll show you Kortex running with a real assistant and sample roles, and walk through your IT team's requirements together.

Book a Kortex demo

How we start

We evaluate together whether there is a real opportunity

Diagnostic questionnaire

You complete a short form about your process, systems and volumes to validate fit before meetings.

Business case analysis

We meet to understand the challenge, estimate the solution's potential value, and confirm whether a clear opportunity exists.

Tailored proposal

If there is a clear opportunity, we prepare a proposal with scope, timeline and pricing adapted to your operation.

Tell us about your operation

Fill out the form and we will get back to you by email to schedule a conversation.

We will reply by email to arrange a first conversation.