AI LAB / DESIGN NOTES
Explore the architecture
behind the intelligence.
Working questions, practical design notes and selected primary references for turning AI ideas into useful systems.
A PRACTICAL FRAMEWORK
From business problem
to measured outcome.
The route for deciding whether — and how — intelligence belongs in a system. Each step is a gate, not a ceremony.
01 →
Business problem
Name the task or decision that should improve, and for whom.
02 →
Data readiness
Establish what the system may know: sources, owners, quality, access.
03 →
Automation opportunity
Separate what should be deterministic from what needs judgment.
04 →
AI / agent decision
Choose models, retrieval and autonomy only where they earn their place.
05 →
Architecture
Design boundaries, integrations and the path to reliable operation.
06 →
Security & governance
Set permissions, evaluation and accountability before launch.
07 →
Implementation
Embed the capability in a real workflow with a real owner.
08
Measurement
Evaluate outcomes against the baseline — and decide what scales.
01 / FROM USE CASE TO CAPABILITY
An approach to
AI transformation.
Where this sequence is applied to a real organisation, it is the basis of AI strategy advisory; the readiness gates are covered in The AI Readiness Assessment: 20 Questions Before You Build.
A working sequence for connecting strategy with execution. It is a way to structure the conversation, not a claim of a proven or proprietary methodology.
01
Identify
State the task or decision that should improve. Describe the people affected and the result they need.
02
Prioritise
Compare expected value with readiness, effort and the consequences of failure.
03
Architect
Connect knowledge, models, workflows, access and human oversight in one design.
04
Integrate
Place the capability inside an actual workflow with clear ownership and support.
05
Measure
Evaluate quality and business usefulness against a baseline, including the cost of operating the system.
06
Scale
Expand only when the evidence supports it, carrying the operating lessons into the next use case.
02 / KNOWLEDGE SYSTEMS
Treat retrieval as
a design decision.
Retrieval-augmented generation combines a language model with access to external information. The original RAG research is a useful foundation for understanding that combination.
A practical design brief should identify the source material, who owns it, how it changes and who may access it. Then define what a useful answer looks like and how a reviewer can check its evidence.
Questions to take into a design review: Which sources are authoritative? What happens when documents disagree? How should the system respond when evidence is insufficient?
Related thinking: Private LLM for Enterprise
Service: AI Architecture Review
Primary reference: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks ↗
Related thinking: The Enterprise AI Stack Explained ↗
03 / AGENT DESIGN
Use autonomy
with a purpose.
Anthropic distinguishes predefined workflows from agents that choose their own process and tools. Its design guidance favours starting simply and adding complexity when it improves results.
For a proposed agent, write a short operating contract: the task, allowed actions, evidence of completion and the conditions for escalation. Include a way to stop and recover when the system cannot proceed reliably.
A useful design review asks where flexibility creates value and where a predictable workflow is enough.
Related thinking: Designing for Agents
Service: AI Architecture Review
Primary reference: Building effective agents ↗
Related thinking: Agentic AI and software development ↗
04 / GOVERNANCE
Make responsibility
part of the design.
NIST’s AI Risk Management Framework offers a voluntary foundation for considering trustworthiness across the design, development, use and evaluation of AI systems.
For an individual initiative, translate governance into everyday decisions: name the owner, document the intended use, define evaluation criteria and agree how changes or incidents will be reviewed.
The aim is an operating practice people can follow, with evidence that helps them decide whether the capability is ready to expand.
Related thinking: An AI Governance Framework CTOs Can Actually Run
Service: AI Readiness Assessment
Primary reference: NIST AI Risk Management Framework ↗
START A CONVERSATION
Have an architecture
question worth exploring?
Bring the use case, the constraint or the design choice you are working through.