AI Thesis: AI as the foundation, not the feature.

An AI-improved product gets a faster autocomplete, a better summary, a smarter search ranker. It would still exist without the model — just slightly worse. An AI-impossible product cannot exist at all if you remove the model: the core experience requires a capability that did not have a price-per-call eighteen months ago. Investors fund the second category. The first is feature-list dressing.

The test is simple. Delete the AI from your description. Can you still write a one-sentence product? If yes, you are AI-improved. If the sentence collapses into nonsense, you are AI-impossible — and your job is to identify which capability curve crossed which threshold in which quarter. That date is your "why now".

The classical loop is: product creates usage, usage emits a signal, the signal improves the model, the improved model differentiates the product. The honest version names what kind of signal: clicks for ranking, corrections for translation, completions for code, outcomes for diagnosis. Without naming the signal you have a slogan, not a flywheel.

The second honest test: does the signal accumulate where you sit, or where the foundation-model vendor sits? If OpenAI sees the corrections before you do, your flywheel powers their factory. If the corrections live in your stack as glossaries, prompt overrides, routing rules, eval sets — then the asset is yours, and a competitor with the same APIs starts from zero.

The cheap-to-smart pattern works because most requests are easy and a small percentage need the heavyweight. Route by default to the cheapest model that meets the latency and quality bar; reroute the segment to a more expensive model when a confidence signal trips a threshold. This pattern is how you keep gross margin alive while still delivering the best result on the long tail.

Two architectural rules. First, every model goes behind an adapter — never call a vendor SDK directly from product code, or you lock yourself into their roadmap and their price list. Second, every stage has a local-only fallback, even if degraded. If your only option is "cloud or nothing", a price hike or an outage is a company-ending event.

Four risk families come up in every conversational-AI product: vendor lock-in (your supply chain), cost spikes (your unit economics), hallucination (your trust budget), and privacy (your regulatory surface). Each has a known structural mitigation. Each must be visible in the architecture diagram, not buried in the FAQ.

The discipline is to convert each risk into a concrete shipping artefact: an adapter interface, a per-user quota, a confidence gate, a retention policy. If a risk has no artefact, it is a wish. Investors and serious customers can tell the difference in under a minute.