AI Architecture
Design the steps AI needs to produce an answer.
Decide what information AI receives, which tools it can use and how to check the result. BLUEURBAN implements these steps in software for each product’s work.
BLUEURBAN
Deliver an inspectable result
A shared design principle
Define what AI does and what the system checks.
Models interpret language and compose outputs. Surrounding systems select information, run tools and check access and structure. Each technology combines these responsibilities for its task.
Separate responsibilities from question to answer.
- 1
Prepare information
Select the data, documents and business definitions needed.
- 2
AI processing
Interpret intent and compose queries, text or tool calls.
- 3
System checks
Check permissions and execution conditions, then apply tools and validation.
- 4
Return the result
Return data, generated work or findings with sources and unresolved items.
What the design needs to account for
- Select the information needed
- Choose the material needed for the current question or task, distinguishing confirmed information from what is still unknown.
- Check permissions before execution
- The system validates the work AI requests before running it. Database queries must satisfy access permissions and read-only requirements.
- Keep the material behind the answer
- Retain query results and references so readers can check what an answer is based on.
- Choose a measure for each technology
- SQL execution accuracy, long-form consistency, source visibility and document-answer F1 measure different things. Evaluate each technology using a measure suited to its purpose.
Use business tools in your AI client
Publishing a business capability as a tool lets a connected AI client discover and invoke it.
Explore MCP tool publishingDesign AI connections around your business requirements.
Review responsibilities from preparing information to checking permissions and returning results.
Discuss the architecture