ROI model for an AI feature — with the assumptions visible

Build an ROI model for {{AI feature}}.

Inputs:
- The feature: {{what it does, who uses it}}
- Expected user volume + frequency: {{paste — usage assumptions}}
- LLM + infra cost per call: {{paste — model used, average input/output tokens, vector DB costs if RAG}}
- Pricing context: {{free / paid / what tier}}
- Quality bar (eval pass rate): {{paste — what counts as working}}
- Cost of getting it wrong (refund risk, churn risk, brand risk): {{paste}}

Output:

## Unit economics
| Component | Per-call cost | Per-user-month cost | Per-user-month revenue (if any) | Net |
|---|---|---|---|---|
| LLM inference | | | | |
| Vector DB / retrieval | | | | |
| Infra (compute, storage) | | | | |
| Support escalations expected | | | | |
| Total | | | | |

## Sensitivity table
| Variable | Pessimistic | Expected | Optimistic | Impact on margin |
|---|---|---|---|---|
| Usage frequency per user | | | | |
| Token cost (model price change) | | | | |
| Quality requirement (eval pass rate) | | | | |
| Support cost (% of calls needing human follow-up) | | | | |

## The 3 assumptions that break the model if wrong
The variables this is most sensitive to. For each: what evidence supports the current assumption, and what the early signal would be that it's wrong.

## Comparable features (real-world)
1–2 features at other companies that have public economics signals (earnings call mentions, public pricing changes, blog posts about cost). What we can learn from their experience.

## When this feature pays back
Time to positive contribution margin under expected assumptions. Time to positive in pessimistic case.

## When this feature should NOT ship
The conditions under which this model says "don't build." Be honest — the answer is sometimes "this is a feature that loses money even at scale."

## Monitoring plan
The 3 numbers to watch monthly post-launch that would tell us if the model is holding or drifting.

Hard rules:
- Don't model on "things will get cheaper" — assume current prices for year 1
- Quality is not a free variable. "Premium model" → real cost. "Smaller model + retries" → also real cost.
- Don't assume usage from people who haven't pre-committed. Top-of-funnel curiosity ≠ usage.