Four phases, in this order.
Most AI projects do not fail at the model. They fail because nobody agreed what a correct answer was, because the hardest assumption was tested last, or because the system was never wired into anything real. The order below is arranged around that.
Understand the problem
We map how the work is done today, who does it, what information already exists, where it breaks, and what a measurably better result would be. No model is selected at this stage.
Output A written problem definition and the criteria we will later measure the system against.
Prove what is uncertain
Every project has one assumption that carries most of the risk. We test that one first, on your real data and on cases you consider representative. Sometimes it is document extraction. Sometimes retrieval quality. Sometimes visual matching, or whether an agent can be trusted to complete an action on its own. We do not decide in advance which one it is.
The phase ends in one of four ways: go, iterate on the data or the scope, pause until a blocker is cleared, or stop. Stop is a real outcome, and it costs less here than anywhere later.
Output A measured answer on feasibility, and an architecture chosen from results rather than assumptions.
Build and integrate
Only now is the architecture chosen: models, retrieval, extraction, vision, agentic components - and plain deterministic software wherever that is the better answer. The system is wired into the data sources, authentication, applications and processes the organization already runs on.
Output A system running inside your environment, not an isolated demo.
Operate and measure
Launch is not the end of evaluation. Quality, cost, latency and the ways the system gets things wrong are tracked on real usage. New cases found in production become part of the evaluation set, and the system improves in measured increments.
Output A system whose quality can be seen, argued about, and improved.
Phase 04 feeds phase 02. That is the loop, and it does not end at launch.
AI has to earn its place.
Sometimes the best system is a simpler workflow, a better index, or deterministic automation.
We do not put a generative model where conventional software solves the problem more reliably and more cheaply. If your problem is not one for AI, we will say so in the first conversation. That is a cheaper answer for both sides than a pilot that was never going to work.
Five principles behind that order.
Why the phases are shaped the way they are.
Problem before model
We do not choose technology before we understand the process, the data, the user and the definition of a correct result. The model is one component of an architecture, and rarely the one that decides whether the project works.
Evaluation is a deliverable
Representative cases and measurement criteria are agreed before the system is built, and the tooling that runs them is handed over with it. Quality is something you can check, not something you have to take on trust.
Architecture follows the problem
Retrieval, fine-tuning, vision, agents and conventional software are not starting positions. We use the combination that produces the best result for the specific case, and we are prepared to explain why that one.
The expert stays in the system
In specialist domains, the people who already do the work define what is correct, what is acceptable, and where the system must not decide alone. AI does not replace the expert - it changes where the expert spends their time.
Production shapes the design
Security, audit, monitoring, cost, latency and integration are not added at the end. They constrain the architecture from the first week, because a system that cannot be operated is not finished.
A demo has to work once. Production has to hold.
Six things we hold to whatever the system turns out to be.
Evaluation
Quality measured on representative cases, before launch and continuously after it.
Traceability
Where the domain requires it, an answer can be traced back to its source or to the decision the system made.
Human control
We define which decisions the system may take on its own and which need a person to confirm.
Observability
Errors, drift, latency, cost and usage patterns are visible to you, not only to us.
Access control
The system reaches the data and the tools it needs for its task, and nothing beyond that.
Deployment
Cloud, private cloud, on-prem or hybrid, decided by where the data is allowed to live.
What you keep when the engagement ends.
Ownership is settled at the start, not negotiated at the end.
Start with the problem, not the AI solution.
Show us the process you want to improve, the data you have, and the result you need. We will tell you where AI actually helps, what has to be proven first, and what the production system would have to look like.
Helpful to bring: how the work is done today · a sample of the real data · who decides what “correct” means