Framework · AI Strategy
Brain. Memory. Hands. Control.
LLM, RAG, Agent and MCP map onto brain, memory, hands and control, so instability can be diagnosed layer by layer. If you understand these four, you understand modern AI architecture.
The problem
“We need an LLM.” “Let’s build an AI Agent.” “Add RAG.” “Use MCP.” Pilots begin. Budgets are allocated. Six months later:
The problem is not intelligence. The problem is architectural clarity.
The four layers
It can: Think · Write · Explain · Summarize · Reason
It cannot: Access your private company data (by default) · Press buttons · Update systems · Send emails
It only thinks.
Strategic insightLLM capability alone introduces minimal operational risk but limited enterprise value without context.
It can: Look up internal documents · Reference policies · Retrieve knowledge base articles · Use enterprise data before answering
Without memory, the brain guesses. With memory, the brain checks first. RAG does not make the brain smarter — it makes it grounded.
Strategic insightRAG reduces hallucination risk but introduces data governance responsibility.
It can: Call APIs · Send emails · Update CRM records · Trigger workflows · Approve requests
Now the brain does not just think. It acts. Advice is safe. Action has consequences.
Strategic insightAgent deployment requires operational maturity and governance discipline.
It defines: How tools are described · How context is passed · How capabilities are exposed · How communication is structured
MCP is not intelligence. It is not business rules. It is structured communication that enables governance.
Strategic insightWithout a control layer, agents become fragmented and difficult to audit.
Modern AI systems classify into four layers: reasoning, knowledge, action and control. Confusing these layers leads to instability. Scaling without control increases risk.
Interactive
For each system, tick what it actually does. The classifier names the layers it spans, what each one demands, and where an agent is acting without control or without a defined boundary.
No systems yet. Add every AI system you run or plan — the inventory is phase one of the model. It stays in this browser only.
Implementation · 90 days
Map your AI footprint: list all LLM usage, identify RAG pipelines, audit agent tool access and classify autonomy levels.
Success metricComplete AI capability map
Separate thinking from doing: restrict agent permissions, standardise retrieval pipelines, introduce tool exposure standards and define approval checkpoints.
Success metricNo autonomous agent without defined boundary
Institutionalise AI architecture discipline: an AI architecture review board, tool usage logging, cost per workflow, and alignment with enterprise architecture governance.
Success metricAI governed like any core platform
Go deeper
Pilots multiplying?
Bring the capability map you built above to a free 30-minute diagnostic.