Gross margin by customer

Gross margin by customer

AI products have a cost of goods that moves with usage, so two customers on the same plan can have very different margins. This kit puts revenue, model costs, and the rest of cost of revenue side by side per customer and month, so pricing conversations start from facts.

Inputs

1. Usage in the shared schema (timestamp, customer, task, model, tokens..., cost_usd). From app logs, or from provider exports via the llm-usage-normalizer kit. Blank cost_usd is priced with --prices (check those prices against each provider's current pricing first).
2. Revenue data/revenue.csv: customer,month,revenue. Use recognized monthly revenue (an annual contract is 1/12 per month), net of discounts. Customer names must match the usage file exactly.
3. Other cost of revenue (optional) data/cogs.csv: item,month,amount,allocation. Allocation is usage (by LLM cost share, good for hosting and vector databases), revenue (payment fees), even (per-customer tools), or customer:<name> (dedicated infrastructure). Only cost of revenue belongs here, not sales or R&D.
4. Target margin (default 70%, --target 0.7). Ask the user; 60 to 80% is common for software with AI features.

Steps

1. Check that customer names match between usage and revenue. List names that appear in only one file and fix them with the user (usually a naming mismatch or an unmapped API key).

2. Run:

   node scripts/margin.mjs data/usage.csv --revenue data/revenue.csv --cogs data/cogs.csv --prices data/llm-prices.csv --target 0.7 --out out/margin

3. Read out/margin.md and report per month: overall gross margin, each negative customer, each below-target customer, and the task that drives their cost.

4. Explain the levers for each problem customer, in this order:

   • Cost: cheaper model for that task's simpler steps, better prompt caching, fewer retries (a low success rate burns money).
   • Packaging: usage caps or overage on the heavy task (see consumption-guardrails).
   • Price: price_for_target is what they would need to pay to hit the target. Use usage-pricing-simulator to test a new model on everyone before changing one deal.
5. Mention unattributed LLM cost (internal use, evals, trials). It is real cost even though it is not in any customer's margin.

Output template

**October gross margin: 57% (target 70%).**
- cobalt-labs: -116% ($60 revenue, $130 cost). 96% of cost is research_agent. Needs $433/mo for 70%.
- northwind-co: 40%. research_agent again. Needs $477/mo.
- Everyone else is above 80%.
Biggest lever: research_agent cost per run. Next: model choice for its steps, then usage-based pricing for it.

Guardrails

• Margins use the price table you give it. Say "based on the example prices" if the user has not checked them.
• Revenue and margins per customer are sensitive. Keep them in out/ and do not paste the full table into shared channels.
• This kit does not change prices or contracts. It informs a decision people make.

Gross margin by customer

Shows the gross margin of every customer of your AI product: revenue minus model costs minus allocated hosting and other cost of revenue. It flags who is negative or under target, which task drives their cost, and what they would need to pay to hit your target.

Who it is for: founders and finance leads at AI products preparing pricing changes, renewals, or board reviews.

Example

Input: examples/usage.csv (October LLM usage), examples/revenue.csv, examples/cogs.csv (hosting, vector database, support tooling, dedicated storage), examples/prices.csv.

Output (out/margin.md):

What it can access

• Runs node scripts/margin.mjs (Node built-ins only).
• Reads data/. Writes out/margin*.
• No network.

Setup

None. If your usage comes from provider exports, run @agentcosts/llm-usage-normalizer first. Check the example prices against each provider's current pricing before trusting the numbers.