Guides/Pricing & unit economics

AI pricing & unit economics

AI features fail commercially when teams surprise finance after launch. Ground conversations in token skew (output ≫ input), tool-line items, and how caching changes marginal cost — then pair with packaging experiments instead of hoping gross margin holds.

Cost drivers

Most bills enumerate input vs output tokens separately — agents and verbose completions concentrate spend on output. Hosted retrieval, web search, or bespoke tools often bill per call — treat them as first-class COGS lines, not “misc infra.”

Prompt caching lowers marginal input cost when stable prefixes repeat (system prompts, large RAG dumps). Model its adoption honestly — not every workload repeats enough to benefit.

Illustrative COGS stack

Percentages shift by product shape — use your vendor invoices to replace fiction with quarterly refresh charts for exec reviews.

Where spend often concentrates (one generic copilot-style turn)

Output-heavy agents skew output further; cached prefixes shrink input.

Validate with finance — use this to argue for caching, shorter answers, or cheaper routing models on simple turns.

Margin sensitivity (schematic)

Small UX choices that lengthen outputs or invoke tools move margin faster than headline subscription price debates — scenario-plan before committing roadmap.

Margin pressure vs length & tools (schematic index)

Output priced per token — small UX changes that lengthen answers hit margin nonlinearly.

Pair with pricing experiments — shorter defaults and smarter routing recover margin faster than arguing over base subscription alone.

PM checklist

Publish an internal “cost per successful task” definition aligned with analytics — not raw tokens alone.
Tie roadmap bets (longer answers, bigger context, more tools) to margin bands before engineering commits.
Instrument downgrade paths — cheaper models or shorter defaults when sessions exceed COGS targets.