Where should you start?

Three questions. No contact details required.

1. Which of these best describes your situation?

A — We have data and a potential use case, but have not committed to a project yet. B — We have a clear use case but data sovereignty or compliance is the main constraint. C — We have an AI system running but cannot tell whether it is performing well. D — We are building AI-native products and working through architecture decisions.

2. What is the most pressing question for your organisation right now?

A — What would this cost, and is our data actually good enough to start? B — How do we use AI without data leaving our own infrastructure? C — How do we know our model is working, and how do we improve it systematically? D — How should retrieval, memory, and agent pipelines fit together?

3. Where does most of the uncertainty sit?

A — Whether the conditions for success exist -- data quality, team capability, realistic timeline. B — Whether we can meet compliance requirements and still use modern AI tools. C — In production -- we shipped something but cannot measure its real-world behaviour. D — In system design -- how the components should fit together at scale.

The letter that comes up most often points to your profile below. Mostly A: Data Readiness. Mostly B: Data Sovereignty. Mostly C: Measurement and Evaluation. Mostly D: Architecture and Agentic Systems.

---

The profiles

Profile A -- Data Readiness

You have data. The question is whether the conditions for success exist.

Most organisations at this stage have more data than they think and less structure than they need. The right first step is a clear-eyed assessment of what you have, what it would take to use it, and whether the expected return justifies the investment. That assessment prevents the more expensive mistake of committing budget before the conditions for success are understood.

• The Cost of ML
• What Data Do I Need?
• What Is Data Maturity?

Profile B -- Data Sovereignty

You have a use case. The constraint is keeping your data inside your own infrastructure.

This is a solvable architecture problem, not a reason to abandon the use case. Open-weight models run inside your own infrastructure -- the same network boundary that already governs your databases and application servers. The data stays where it is. The compliance objection dissolves before it reaches legal review.

• Private Inference: Running AI Inside Your Own Infrastructure
• Why Private Inference
• Marigold: Privately Hosted AI Inference on AWS

Profile C -- Measurement and Evaluation

You have a system in production. The problem is knowing whether it works.

A model without a measurement framework is a liability. The eval dataset -- not the model -- is the primary artefact of any production AI system. It persists across model versions, provides the deployment gate, and makes the difference between a system that improves systematically and one that degrades silently.

• Build Measurements, Not Models
• A Measurement Framework for Production AI
• Labels Are the Task

Profile D -- Architecture and Agentic Systems

You are building AI-native systems and working through how the components fit together.

Agentic systems introduce design problems that single model calls do not: memory across sessions, retrieval strategy, pipeline composition, failure modes that compound across steps. The architecture decisions made early are expensive to reverse. Getting the structure right before scale is where the leverage is.

• Agent Memory
• RAG Strategy
• Static vs Dynamic Planning

(If you would like to talk through what these suggest for your situation, get in touch.)