Value model

A value case is one of the four things that sink AI pilots when it is missing. This page works the model on the demo’s own case: it uses the system’s real self-measurements for the cost-to-run side, and clearly-labeled illustrative assumptions for the business-outcome side, so you can see the method and rerun it on your own numbers.

The shape of the model

Value is the outcome it moves minus the cost to run it:

value = (outcome delta, evidenced) - (cost to run, measured)

The discipline is asymmetric on purpose: the cost to run must be measured, and the outcome delta must be evidenced against your own baseline, never assumed. A value case that assumes its outcome is not a value case.

What the demo measures for real

The cost-to-run side is not hypothetical here - the system already meters it:

Per-turn cost is metered by the as-if-paid ledger at real pay-as-you-go rates; the cost-at-volume model turns that into a monthly figure at your traffic.
Quality is gated: the eval harness scores every change and blocks on the safety dimensions (see the model card and the eval harness).
Safety has a deterministic floor that never depends on the model.

So the left-hand cost term, and the “does it work safely” precondition, are demonstrated, not promised - which is the part a value case usually hand-waves.

The illustrative business layer

The outcome side depends entirely on your panel, so the figures below are illustrative placeholders, not claims. Medication non-adherence is a well-documented cost driver - missed doses drive avoidable admissions and complications - and the model multiplies four inputs you supply from your own data:

avoided cost = panel size x baseline non-adherence x adherence lift x cost per non-adherent patient

Scenario	Adherence lift (illustrative)	Reading
Conservative	a small single-digit lift	run-cost-dominated; value is thin until the panel is large
Base	a modest lift	value crosses run-cost at a moderate panel size
Optimistic	a larger lift	value is outcome-dominated; run cost is a rounding error

Sensitivity: the answer moves most on the adherence lift and the cost per non-adherent patient - both are yours to measure, not ours to assume. Run cost barely moves the result once the panel is non-trivial, which is the point: a cheap, safe, measured system makes the value case turn on the clinical outcome, where it belongs.

How to use this

Replace the illustrative row inputs with your panel size, your baseline adherence, your measured lift (ideally from a small controlled pilot), and your cost per non-adherent patient. Keep the cost-to-run side as the system measures it. If the model only clears the bar under the optimistic row, that is a finding worth surfacing before building, not after.