Use the least complex AI that solves the problem
One schema for every kind of "AI" solution. Walk the gates top to bottom, answer each one, and stop at the first yes. The more powerful the tier, the more unpredictable and expensive its failures, so we take the simplest approach that's genuinely powerful enough for your case
This answers the HOW, not the WHY. The business case for doing the thing should already be settled: this only decides what to build it with.
Decision framework
Walk the gates from the top, answer each, and stop at the first yes.
The five tiers, simplest first, are listed under “The whole map” below. The more powerful the tier, the more unpredictable and expensive its failures, so take the simplest approach that is genuinely powerful enough.
Five tiers, simplest first
Same ladder, laid flat. Colour runs from low complexity and low failure cost to high
Plain coded logic. Fully auditable, instant, free to run and trivial to debug: if conditions are knowable, nothing beats it.
Trained on tabular data with labelled examples. Fast, explainable and cheap at high volume: the workhorse for prediction on structured signals.
One stateless prompt in, one answer out. No tools, no steps, no branching, just pure language transformation.
A loop that calls tools, reads results and branches, bounded to one coherent goal. More power, more surface area to test.
Specialised roles working in parallel or critiquing each other. The most capable, and the hardest to predict, observe and debug. Use only when nothing simpler fits.
Hard constraints push you simpler
The framework is meant to be indicative and simple: by design it can't catch every nuance of a real project. So we don't apply it blindly: latency, compliance, cost-per-call, failure cost and volume can all pull a decision back toward a more deterministic tier
Let's place it on the map together
Tell us the problem and the constraints: we'll scope the least complex architecture that solves it, and the path to production behind it