# Architecture

> Seven layers, from the AI employees at the top to the infrastructure at the bottom, each one assembled from what is underneath it

Source: https://aipril.dev/docs/overview/architecture

Forward this page to IT. It is the whole stack on one screen: what an AI employee runs on, what a
workspace is assembled from, and what all of it relies on.

## How to read it

Seven bands, joined by the relation each one has to the band beneath: **runs on**, **built on**,
**secured by**, **connects to**, **relies on**.

The shape says one thing: **nothing at the top is a separate product.** An AI employee has no
storage and no permissions of its own. It is L1 executing through the same L4 components a person's
screen executes through, under the same L5 controls, which is why an answer from an agent and a
number on a dashboard cannot disagree.

## L1 · AI employees

The autonomous workforce: hiring and mandates, skills, the tool surface, retrieval over your own
knowledge, model choice, budgets, guardrails and the approval gate, escalation to a person, and a
decision trace on everything.

A runtime with identity, boundaries, a budget and a record, not a chat window. See
[AI harness](https://aipril.dev/docs/harness.md) and [AI organisation](https://aipril.dev/docs/ai-organisation.md).

## L2 · Workspaces

Where work and business state live: applications and portals, spreadsheets, task boards and
inboxes, chat, collaboration threads, dashboards, and the exploration surface.

Every one of them is a view onto the same platform objects. See
[Applications](https://aipril.dev/docs/apps.md), [Sites and portals](https://aipril.dev/docs/cms.md), [Spreadsheets](https://aipril.dev/docs/spreadsheets.md),
[Collaboration](https://aipril.dev/docs/collab.md) and [Analytics](https://aipril.dev/docs/analytics.md).

## L3 · Workbench

How the things above get built and changed: AI-first generation, applications as code in the built-in
git, the process editor, the data-model builder, the marketplaces, the prototype canvas, the page and
chart builders, and the design system.

This is the layer that decides whether you own the system or wait for someone else to change it. See
[Applications](https://aipril.dev/docs/apps/authoring.md) and
[Versions and the built-in git](https://aipril.dev/docs/platform/versions.md).

## L4 · Platform

The core components everything above is assembled from: orchestration and the workflow builder, the
query engine and semantic layer, sandboxed Python pipelines, the data catalogue and master data, the
ledger, the scheduling solver, speech recognition, ETL and writeback, API integration, the MCP
server, timers and triggers, object storage, realtime transport, notifications, roles, the audit trail, and separate development, test and production landscapes.

These are shared services, not per-application copies. That is the mechanical reason two
applications installed side by side behave as one system. See [Platform](https://aipril.dev/docs/platform.md).

## L5 · Security

Built into the core rather than layered on: single sign-on over standard protocols, roles and resource
grants, row-level security, column masking, agent mandates and least privilege, a fail-closed approval
gate in front of consequential actions, organisation isolation by schema, encrypted credentials, the
audit log, and air-gapped operation.

The property worth checking: the platform applies these where it plans the query and where it
dispatches the tool call, not in the interface. See [Access and isolation](https://aipril.dev/docs/platform/access.md).

## L6 · Connectors

What it connects to: business systems, databases, warehouses, object storage, files, directories and
any REST API. 130+ sources, with the typed connector registry behind them.

See [Connectors](https://aipril.dev/docs/platform/integrations/connectors.md) and
[Sources and storage](https://aipril.dev/docs/engine/sources-and-storage.md).

## L7 · Foundation

What it relies on: containerised microservices on Docker and Kubernetes, cloud or on-premises, an
open API-first architecture, an event-driven core, a stateless API layer that scales horizontally, an
in-memory analytical engine, columnar data on object storage, zero-copy reads, and zero telemetry.

See [Platform](https://aipril.dev/docs/platform.md) for the deployment and scaling detail, and
[Privacy and on-prem](https://aipril.dev/docs/harness/privacy.md) for what a self-hosted installation takes.

## What this means in one sentence

**Every property here comes from where it sits in the stack.** Access holds everywhere because L5
sits under L4; the numbers agree because there is one query engine; an AI employee is governed
because it executes through the same components a person does; and you can change any of it because
L3 authors the layers above it as artifacts, not as code somebody else owns.

Next: [All capabilities](https://aipril.dev/docs/overview/capabilities.md).