Price and Package the AI Feature

Workflow area: AI Product Economics · Tier: Advanced · Hands-on time: ~40 minutes

Pricing an AI feature is where two truths collide: your costs scale with usage (pb-46 just showed you the curve), but users hate metered anxiety and love flat predictability. Every packaging model is a different allocation of that tension — who absorbs variance, you or the customer. Teams that skip this thinking default to "include it in the plan," discover their power users are unprofitable, then bolt on limits that feel like a bait-and-switch. The one-pager forces the trade-off to be chosen, not stumbled into.

When to use — and when not to

• Use it before GA of any feature with material per-use costpricing bolted on after launch always reads as a take-away.
• Use it when the unit-economics model shows an unprofitable tailpb-46's breakeven percentile is precisely the input that decides between flat and usage-based.
• Use it at repackaging momentsmoving AI from "beta perk" to paid tier is a pricing launch with churn risk; it deserves the full pass.
• Skip heavy machinery when cost rounds to zeroa genuinely cheap feature priced as a differentiator inside existing plans needs a paragraph of justification, not this playbook.
• Skip outcome-based pricing without outcome measurementcharging per resolved ticket requires pb-41's containment definition to be bulletproof and auditable; without it you'll litigate every invoice.

The recipe

1. 1.

   Bring the three inputs to the table. From pb-46: blended cost per user, the breakeven usage percentile, and the stress-scenario margin. From your metrics ladder: what a "successful outcome" is and whether you can measure it per-account. From the market: what the two nearest competitors charge and how they meter. Missing any of these, stop and get it — pricing debates without them are taste contests.

2. 2.

   Score the four models against your situation. Run the structured comparison:

   Copy

   Help me choose pricing packaging for an AI feature. Score each
   model honestly against MY numbers, not in general.
   Feature + value: [WHAT IT DOES, TIME/MONEY IT SAVES A USER]
   Unit economics: [BLENDED COST/USER/MO, BREAKEVEN PERCENTILE,
   WORST STRESS-SCENARIO MARGIN — FROM MY MODEL]
   Customer base: [SMB/ENTERPRISE MIX, CURRENT PLANS + PRICES]
   Usage distribution: [p50 / p90 / p99 INTERACTIONS PER USER]
   Outcome measurability: [CAN WE COUNT SUCCESSES PER ACCOUNT? HOW
   DISPUTABLE IS THE COUNT?]
   
   For each of: (1) seat add-on flat fee, (2) usage credits,
   (3) hybrid (allowance + overage), (4) outcome-based —
   score 1-5 on: margin protection at my p99, buyer predictability,
   sales friction, gaming/dispute risk, ops complexity. Then state
   which usage-distribution shapes favor each model, which my
   distribution actually is, and recommend one with the runner-up
   and the condition under which the runner-up wins.
   

   Sample output fragment:

   Copy

   Your p99 is 14× your p50 — a long-tail distribution. Flat seat
   add-on scores 2/5 on margin protection: your top decile breaks
   even at $X. Recommended: hybrid — allowance set at your ~85th
   percentile (covers "practically everyone" psychologically) with
   metered overage. Runner-up: pure credits, which wins if your
   buyers are already credit-literate (dev tools, API customers).
   

3. 3.

   Set the allowance line with psychology, not just math. The allowance should cover enough of the distribution that the median buyer never thinks about it (80th-90th percentile is the usual zone), while the overage price sits above marginal cost with room for the stress scenario. Anchor the tier price to value delivered (hours saved, tickets resolved) — cost sets your floor, value sets the number.

4. 4.

   Design bill-shock prevention as product work. This is where AI pricing launches die. Spec four affordances: a live usage meter in-product (not buried in billing), proactive alerts at 50/80/100% of allowance, a hard-cap option the admin controls (degrade gracefully — queue or downgrade quality, don't silently bill), and no-surprise defaults: overage off until explicitly enabled for SMB, invoice-with-review for enterprise. Metering visibility is a trust feature; invisible meters read as traps.

5. 5.

   Write the pricing one-pager. One page: chosen model and the sentence of rationale · price points and allowance line with the percentile it covers · margin at p50/p90/p99 under the chosen scheme (from the pb-46 sheet, re-run) · bill-shock affordances committed · grandfathering/migration plan for existing users · the review trigger (model prices, usage distribution shift, competitor move). This page is the artifact leadership approves — everything else is appendix.

Make it yours

• Enterprise variant: replace per-seat metering with an account-level annual commitment sized on projected usage, true-up at renewal — enterprises trade a higher price for zero monthly variance, which also caps your invoice-dispute surface.
• Credits-as-platform variant: if you'll ship multiple AI features, one credit currency across features beats per-feature pricing — decide the credit's exchange rate per feature from each feature's cost, and keep one feature's price change from cascading.
• Outcome-based pilot: rather than betting the pricing on outcomes, run one design-partner cohort on per-outcome pricing with a cap, for a quarter — you learn dispute rate and measurement robustness before committing the price book to it.

Quality check

• Margin holds at p99 under the chosen modelre-run the pb-46 sheet with the pricing applied; if the power decile is still underwater, the packaging failed its one job.
• The allowance covers a named percentile"generous" is not a spec; "covers 87% of current users at median month" is.
• Every metered surface has a visible meterusage users can't see is usage they'll dispute.
• Overage requires an explicit opt-in or capthe default path can produce a larger-than-expected bill for no customer, ever.
• Migration is written downexisting users' path (grandfathered, migrated with notice, beta sunset) is in the one-pager, because repricing surprises churn harder than high prices.