Skip to main content

Boardroom Answers · Technology & Platform · Technical Architecture & Scalability

Your platform is built around today's AI patterns — prompts, RAG, vector search. When the AI paradigm shifts again in eighteen months, does your data architecture survive?

The question a Chief Solutions Architect (Enterprise B2B SaaS) asks.

The short answer

Models and prompts are swappable by design; the compounding asset is the governed, provenance-linked decision corpus in standard Postgres — the next AI paradigm plugs into the same contract layer without touching the system of record.

The full executive answer

The architecture separates what ages fast from what compounds, and the durable asset is the data layer. The system of record is Postgres with a governed, tenant-isolated schema; vector search runs on pgvector — an open-source extension inside that same database — so embeddings live next to the governed data rather than in a separate vector-database vendor that may not exist in three years. Embeddings are, by nature, regenerable: if better embedding models or entirely new retrieval paradigms arrive, we re-index from the source records; the investment that compounds is the structured, provenance-linked decision and evidence corpus, and that is model-agnostic by construction.

The AI-facing layers are deliberately thin and swappable: prompts are versioned assets, providers sit behind one gateway, and — most importantly for paradigm shifts — the platform's contract with the AI is defined at the output side: typed schemas describing what a valid analysis is, enforced by validation and the CI evaluation harness. Whether the thing producing that output is today's tool-calling model, tomorrow's agentic system, or something after that, the governed contract layer is the stable interface. We already run scheduled background AI workers and ship an MCP surface — the emerging standard for AI agents connecting to tools — so the agentic direction is a lane we are in, not a disruption we await.

What I will not claim is prophecy. The defence against paradigm shift is not betting right — it is keeping the switching cost of the AI layer low and the value concentrated in what does not deprecate: governed data, provenance, audit history, and the customer's accumulated decision memory. In evolutionary-architecture terms, the fitness function is 'can we adopt the next paradigm without touching the system of record' — and every layer boundary was drawn to keep that answer yes.

Grounded in: Evolutionary architecture (fitness functions); pgvector/open-standards data strategy

Want this answered live, on your data?