Building AI-native analytical systems

Context

I am interested in what happens when AI is treated not as a chatbot added to an existing application, but as part of the analytical architecture itself. Much of my experimentation comes from quantitative and investment-risk workflows: combining structured data, deterministic calculations, reusable analytical tools, and language models into systems that can investigate a question rather than simply display information.

How I approach it

I keep a clear boundary between what software should calculate and what AI should reason about. Financial metrics, transformations, and controls remain deterministic and reproducible. Models are more useful for deciding what to inspect, connecting information across analytical domains, explaining results, identifying unusual patterns, and synthesising large amounts of evidence.

• Analytical agents using specialised tools rather than one large prompt
• Parallel analysis followed by model-driven synthesis
• AI grounded directly in structured data and documentation
• Provenance and traceability from conclusions back to evidence
• Interfaces where conversation, charts, documents, and traditional dashboards coexist

Why it matters

The biggest change from AI may not be faster report writing, but a different analytical workflow: instead of analysts navigating applications to find the right data, software increasingly assembles the relevant analysis around the question being asked. The challenge is designing that system without losing precision, reproducibility, or human judgement.