An architecture for AI you can keep

Build your AI platform once. Swap the tools forever.

The model, the chat app, the automation tools — all of these get replaced inside a year or two. What lasts is the work you put in: the instructions you write down, the knowledge you collect, and the way you connect everything to the systems you already run. The platform's job is to keep all of that portable when the tools around it change.

Figure 1 A future-proof AI platform
Where AI meets people surfaces, grouped by app CHAT Claude · ChatGPT app VOICE phone bots · dictation IDE / CLI Cursor · VS Code · Claude Code · terminal EMAIL support@ · dispatch@ WEB / API embedded chat · webhooks — OAuth 2.1 · sign-in — App what users actually open · swap as the market moves Business chat For staff asking questions, drafting, summarising. Claude Desktop · ChatGPT Ent. M365 Copilot · Open WebUI Developer tool For people writing or editing code. Claude Code · Codex · Cursor Gemini CLI · Continue Agent framework For automation, pipelines, scheduled tasks. LangGraph · CrewAI · Mastra Pydantic AI · open-source — Model API · vendor-specific — Model the LLM each app calls Model Vendor app: locked to one model family. Open app: lets you swap models when something better arrives. Claude · GPT · Gemini · Llama · open-weight The durable core your IP · compounds across every replacement Durable core — your IP Skills repository Your rules, prompts, and how-to recipes — the playbook for how your AI should work. In your own repo · versioned · reviewable Knowledge & memory The facts, files, and history your AI draws on — what your company actually knows. Shared across every app · yours to keep Evaluations Tests that prove your AI still works — before and after every tool swap. Versioned with your skills · run automatically Policy & guardrails What your AI is and isn’t allowed to do — approvals, allow-lists, audit trails. Written down · auditable · enforced everywhere — MCP · one protocol for every tool — Integration Model Context Protocol · open glue Documents wikis · drives Records CRM · ERP Data warehouse · BI Identity directory · perms Custom any internal API — HTTP + OAuth · the web — Systems of record what you already run · varies by company Documents SharePoint · Drive Operations Salesforce · SAP Warehouse Snowflake · Databricks Directory Entra · Okta Internal in-house APIs
  1. 01 — Channels

    Where AI meets people

    Surfaces, grouped by app

    • Chat — Claude, ChatGPT app
    • Voice — phone bots, dictation
    • IDE / CLI — Cursor, VS Code, Claude Code, terminal
    • Email — support@, dispatch@
    • Web / API — embedded chat, webhooks
  2. — OAuth 2.1 · sign-in —
  3. 02 — App

    What users actually open

    • Business chat — Claude Desktop, ChatGPT Ent., M365 Copilot, Open WebUI
    • Developer tool — Claude Code, Codex, Cursor, Gemini CLI, Continue
    • Agent framework — LangGraph, CrewAI, Mastra, Pydantic AI
  4. — Model API · vendor-specific —
  5. 03 — Model

    The LLM each app calls

    Vendor app: locked to one model family. Open app: lets you swap models when something better arrives.

    Claude · GPT · Gemini · Llama · open-weight

  6. 04 — your IP

    The durable core

    What persists across every replacement.

    • Skills — your rules, prompts, recipes
    • Knowledge & memory — facts, files, history
    • Evaluations — tests that prove it still works
    • Policy & guardrails — what AI can and can’t do
  7. — MCP · one protocol for every tool —
  8. 05 — Integration

    How the core reaches out

    • Documents · wikis, drives
    • Records · CRM, ERP
    • Data · warehouse, BI
    • Identity · directory, perms
    • Custom · any internal API
  9. — HTTP + OAuth · the web —
  10. 06 — Systems of record

    What you already run

    • Documents · SharePoint, Drive
    • Operations · Salesforce, SAP
    • Warehouse · Snowflake, Databricks
    • Directory · Entra, Okta
    • Internal · in-house APIs

Three principles.

  1. Write down what your AI should do — and keep it in your own repo.

    The rules, the company knowledge, what the AI can and can't do — write them as plain files in a folder you control. When the model or chat app gets replaced, the next one loads the same folder. The files stay yours, and that's where the value is.

  2. Use the same formats and protocols everyone else is using.

    If you connect a system using a one-vendor format, you've married that vendor — and a divorce later means rebuilding from scratch. Stick to the protocols that more than one company supports: MCP for connecting tools, Git for versioning, OAuth for login. Then if the vendor goes sideways, swapping them out is a config change, not a rewrite.

  3. Don't build what big companies are already spending billions on.

    You're not going to make a better model than Anthropic, OpenAI, or Google. You're not going to build a better chat app than the people whose whole job it is. Pay for those. Spend your team's time on what only your team knows: your data, your workflows, how your business actually runs. That's the part nobody else can do for you, and it's the part worth investing in.