I have a trading system that runs continuously on a prediction-markets platform. It is at 137 development phases. Each phase is a small, tested, shipped piece of behaviour. Each one went through a discovery note, a plan, a review and a set of tests that had to pass before the phase could be marked complete.
I mention the number 137 because it is the opposite of how most people talk about AI systems.
The standard AI-project narrative is: we built a thing, it does X, look at the demo. The true story, in my experience, is: we built a tiny thing, it did X poorly, we added a second tiny thing that did Y, we spent a week fixing the subtle bug where X and Y interacted badly, we shipped a third tiny thing and so on, for months.
The system I’m describing started with one pattern: screen every live market, rank them, pick the top few, execute a trade. That worked for about a week. Then we added news ingestion, because markets move on news before they move on fundamentals. Then we added a research layer, because news without context is noise. Then we added resolution-confidence gating, because some markets resolve ambiguously and it is better to stay out. Then we added an arbitrage detector, because some markets have internal inconsistencies worth more than the headline signal. Then we added position sizing, because without it a few bad trades wipe out a month of good ones.
Every one of those was a phase. Every one of those shipped and ran before we moved on.
The alternative, building the whole thing in one shot, deploying it, watching it go wrong, is the thing every AI consultancy I have watched has done at least once. It never works. What works is the discipline of incremental shipping: small, tested, live; small, tested, live; small, tested, live. For months.
The tests matter. The phases matter. The willingness to ship something that only does one thing matters. The 1,800-odd tests the system carries are not a flex, they are what lets me change one piece without breaking five others.
If you are looking at an AI vendor who promises to build your whole thing in two weeks, you are looking at an AI vendor who has not run one of these systems in production. Tell them no. Tell them to build piece one. Run it. Then decide.