What is the AI pyramid?
The AI pyramid shows which layers an organisation needs to let AI agents work safely. Data at the bottom, above it a security layer with permissions per agent, then the agent layer and at the top access for people.
- Access for peopleChat, web, dashboard, app
People work with agents where they already are.
- Agent layerAgents working together through an orchestrator
Specialised agents, connected horizontally and vertically.
- Security layerPermissions per agent: create, read, update, delete
Start with read, add write per use case, deleting stays human work.
- DataStructured and unstructured
ERP and CRM, plus contracts, manuals and email.
Explanation
Data: structured data from ERP, CRM and other systems, and unstructured data such as contracts, manuals and email.
Security layer: this is where you set, per agent and per system, which permissions it has: create, read, update or delete. We always start with read-only access and add write permissions per use case. Deleting stays human work. Every write action is logged under the agent's name and its passport, so an audit becomes a search query instead of a project.
Agent layer: specialised agents work together, horizontally through an orchestrator and vertically with the people who direct them.
Access for people: people work with agents where they already are, in a chat in Teams, a web app, a dashboard or a mobile app.
Agents are as good as the layers beneath them. A smart agent on messy data or without clear permissions does not deliver reliable work.
What it looks like in practice
At a care organisation, a planning agent reads rosters and client data and may only create draft rosters. A planner approves them in the dashboard she already uses.
