Manage Gross Margin by Cohort, Not Just in Aggregate
An average gross margin can hide a structurally bad segment. I want to know whether specific plans, customer types, features, or usage profiles are profitable enough to support the service they receive.
- Margin by plan
- Margin by customer
- Margin by feature
- Heavy-tail analysis
- Enterprise exceptions
The situation
If ninety customers are highly profitable and ten are deeply unprofitable, the blended number may still look acceptable. That does not mean the commercial model is stable, especially if the unprofitable behaviour becomes more common as the product grows.
The naive solution
Track one company-wide AI cost percentage against total revenue.
The hidden constraint
Different cohorts can have completely different usage shapes. Agencies may generate many variants, developers may run long repair loops, enterprise users may demand premium models, and free users may consume heavily without revenue.
The decision
Review contribution margin by meaningful cohort and inspect the tails, not just the average.
The implementation
At minimum I would segment by plan, customer, feature, model route, and usage decile. For each segment: revenue, variable AI/tool cost, successful outcomes, support burden where measurable, and contribution margin. Large customers with negotiated contracts should be modelled separately because the economics may include committed spend, BYOK, custom limits, or services revenue.
The tradeoff
Fine-grained profitability data can tempt teams to overreact to short-term variance. I would use rolling periods and cohort trends, not punish a customer because one complex task happened to be expensive.
The evidence I would look for
The analysis should reveal whether poor margin is caused by pricing, model choice, product inefficiency, abuse, support cost, or a specific high-cost capability.
The reusable lesson
Average margin can hide a bad product contract. Inspect who creates the cost and why.
The consultant's questions
- Which customers or plans have the lowest contribution margin?
- Is poor margin caused by high valuable usage or by waste and failures?
- Does the heavy-usage cohort have a commercial path to expand revenue?
- Are enterprise exceptions being measured or simply absorbed?