Member Tools — AI Cost Forecast
Portfolio-wide per-member COGS projection across 15 shipped tools
Prepared by A Player Labs
22 April 2026 · Draft v1
Bottom line
4 of 15 tools drive variable cost. The other 11 are calculators at $0/run. Blended AI COGS stays under $3/member/month even at 1,000 members.
that's the advantage.
Blended / mo
$2.85
per member
Power user p95
$18.00
top 5% of users
1,000 members
$2,850
total monthly AI COGS
Per-tool unit economics — only AI tools incur variable cost
| Tool (UI name) | Folder | Type | Model | Calls / run | In / Out tokens | Cost / run |
|---|---|---|---|---|---|---|
| Agentic Commerce Agent | aca | AI | Claude Sonnet 4.5 | 1 | 5K / 3K | $0.06 |
| Agentic Commerce Rewriter | acr | AI | Claude Sonnet 4 | 50 per catalog | 1K / 0.5K ea | $0.40 |
| Review Scoring & FAQ Engine | RSF | AI | Claude Sonnet 4.5 | 6 | 4K / 2K avg | $0.25 |
| Winning Creative Remixer | wcr | AI | Gemini 2.5 Flash + Sonnet 4.5 | 4 | 6K / 2K avg | $0.15 |
| Google Campaign Toolkit | gct | Calc | — | — | — | $0.00 |
| The Scale Into Profit Calculator | TSC-V2 | Calc | — | — | — | $0.00 |
| Meta Campaign Spend Calculator | MFC | Calc | — | — | — | $0.00 |
| Media Optimization Ad Calculator | MOAC | Calc | — | — | — | $0.00 |
| The Meta Optimisation Calculator™ | MOC | Calc | — | — | — | $0.00 |
| The Growth Rate Calculator | GRC | Calc | — | — | — | $0.00 |
| Meta Media Buyer's Orderbook | metatron | Calc | — | — | — | $0.00 |
| Size & Stock Snapshot | sss | Calc | — | — | — | $0.00 |
| Size & Stock Snapshot (Vendor) | sssvend | Calc | — | — | — | $0.00 |
| Stock on Hand Snapshot | TWP | Calc | — | — | — | $0.00 |
| The Claude Project Setup Guide | cps | Config | — | — | — | $0.00 |
Forecast by member segment
| Segment | % base | AI runs / mo | $ / member / mo |
|---|---|---|---|
| Power user | 5% | 14 | $18.00 |
| Active user | 35% | 3.5 | $2.46 |
| Dabbler | 60% | 0.3 | $0.06 |
| Blended average | 100% | — | $2.85 |
Total COGS at scale
| Member base | Monthly AI COGS | Top 5% burn | Bottom 60% burn |
|---|---|---|---|
| 250 | $710 | $225 | $9 |
| 500 | $1,425 | $450 | $18 |
| 1,000 | $2,850 | $900 | $36 |
| 2,500 | $7,125 | $2,250 | $90 |
Assumptions & recommendations
Pricing
Sonnet 4/4.5 at $3/$15 per MTok. Gemini 2.5 Flash at $0.075/$0.30. ~50% prompt-cache hit. No Opus in the mix — adding it lifts the affected tool ~5×.
No rate limits applied
Power-user figures assume unthrottled use. A 20-runs/mo cap flattens the top-5% to ~$6 and cuts blended COGS roughly in half.
ACR scales linearly
Baseline assumes 50-product catalog. A 500-SKU bulk rewrite costs ~$4.00. Cap per-account to prevent catalog-dump abuse.
Meter first, price later
These are directional. Add a run counter to each AI tool and calibrate against real usage before setting member pricing tiers.
Cap the two risks
WCR and ACR are the only tools with batch-input exposure. A monthly run limit is the cheapest insurance.
Haiku fallback at 1,000+
Route ACA / RSF "cheap path" summarisation to Haiku. ~70% cost reduction with acceptable quality on summarisation tasks.
Scale forecast — what happens as the AI tool count grows
Linear growth — blended COGS per member / mo
| Tool count (AI only) | Dabbler | Active | Power | Blended |
|---|---|---|---|---|
| 4 today | $0.06 | $2.46 | $18.00 | $2.85 |
| 8 tools (2×) | $0.12 | $4.92 | $36.00 | $5.70 |
| 12 tools (3×) | $0.18 | $7.38 | $54.00 | $8.55 |
| 16 tools (4×) | $0.24 | $9.84 | $72.00 | $11.40 |
Max-usage stress test — worst case ceiling
| Scenario | $ / member / mo | 1,000 members |
|---|---|---|
| Today · power-user uncapped | $18 | $18,000 |
| Today · every member = power user | $18 | $18,000 |
| 12 tools · every member = power user | $54 | $54,000 |
| 12 tools · batch-abuse (ACR × 10) | $90 | $90,000 |
Read this as a ceiling, not a forecast
These figures assume zero rate-limiting and every member hitting max use. Reality lands at 5–15% of ceiling in practice. Budget = blended × 2 for safety.
Cost-minimisation levers — codify first, call the LLM second
1 · Codify the repeatable
Most "AI" output is actually template + variables. Product descriptions in the same category, email sequences for the same funnel stage, SOPs for the same task — generate the master once, substitute variables client-side. Replaces 100s of LLM calls with 1.
2 · Rule-based pre-classification
Most classification work is deterministic. MOAC already sorts ads via CPP/ROAS thresholds — zero AI. Apply the same pattern to review scoring, creative tiering, product grading. Only escalate the ambiguous 10% to the LLM.
3 · Response caching with TTL
Identical or near-identical inputs should hit a cache (Redis, 24-hour TTL). A popular review theme or common FAQ gets generated once per day, served 1,000× from cache. 90%+ hit rate on mature tools.
4 · Retrieval over generation
Instead of asking Claude to write a new brief, retrieve the closest previous brief by embedding similarity and adapt it client-side. ~100× cheaper than regeneration. Works well for WCR's brief library once 50+ briefs exist.
5 · Haiku triage layer
Use Haiku ($1/$5 per MTok) to route the request. Simple cases resolved in-layer; only the hard 20% escalates to Sonnet. Net cost reduction 60–70% on the full AI portfolio.
6 · Batch before send
ACR currently runs 50 prompts for 50 products. Batch 10 products per prompt → 5 calls instead of 50. ~40% token savings with no quality loss. Same lever applies to RSF and ACA.
7 · Client-side pre-filter
Before sending data to the LLM, filter it client-side by rules (rating thresholds, keyword matches, length). RSF doesn't need to process all 500 reviews — the top 50 cover 80% of the signal. Linear input-cost reduction.
8 · Prompt-cache discipline
Stable system prompts + reference CSVs cached for 5 min = 50% input-cost cut already baked in. Discipline in how requests are batched and ordered within a session pushes that to 80%. Free money if engineered tightly.
9 · Hard caps per account
Every AI tool ships with a monthly run ceiling by tier. Power users hit it and either wait or upgrade. Non-negotiable floor below which the worst-case ceiling cannot breach.
Stacked impact
Levers 1, 2, 3 and 6 applied together reduce portfolio LLM spend by an estimated 60–75% at steady state. This matters more as the tool count doubles — at 12 tools, the blended $8.55 becomes ~$2.50 with discipline. Grow the portfolio, flatten the cost curve.
Excluded — HWE (internal ops), MOAC-Sales (fork), TSC v1 (superseded), cps-sales (internal), ee-app / ee-platform / ee-quiz (infra)
● Directional estimates pending real usage telemetry