Margin Erosion

Margin erosion is the gradual decline of gross margin as a business grows — revenue rises, but the cost of delivering that revenue rises faster. It is dangerous precisely because the headline numbers look healthy: more customers, more usage, more revenue, and a slowly worsening ratio that nobody notices until it is structural. Software has historically enjoyed very high gross margins because the marginal cost of an additional user was close to nothing. AI features change that arithmetic. Inference is a real variable cost incurred every time a customer uses the feature, and unlike hosting it does not amortise across users. A product priced on a flat seat model with an unbounded AI feature has, in effect, sold an all-you-can-eat plan for an ingredient it buys by the gram. The mechanisms are worth separating because they call for different responses. Adoption erosion happens when a feature succeeds and usage per customer climbs; it is good news with a cost problem attached. Efficiency erosion happens when cost per unit of work rises — a longer system prompt, more retrieved context, a router sending more traffic to an expensive model, retries that nobody counted. Mix erosion happens when growth concentrates in heavy-usage customers or in a cheaper plan. Discount erosion happens when average realised price falls through negotiation while costs hold. Diagnosing which one you have requires unit economics rather than totals. Track revenue per account against cost to serve per account, and decompose the cost into volume and cost per unit. Without that decomposition, every explanation sounds plausible and none can be tested, which is how organisations end up cutting the wrong thing — reducing quality to save cost when the actual problem was pricing, or repricing when the actual problem was an inefficient prompt. The responses divide accordingly. Engineering fixes efficiency: prompt and retrieval trimming, caching, deduplication, cascading to smaller models, sensible output limits. Pricing fixes structure: usage allowances, credits, tiering by intensity, or charging separately for the expensive capability. Product fixes behaviour: making the cheap path the default, removing invitations to unbounded use, and designing features that resolve in one interaction rather than ten. The discipline that prevents surprise is watching gross margin per unit from the first release rather than at the first board meeting where it matters. A feature whose cost per interaction is rising while the price is fixed does not fail suddenly; it becomes gradually less profitable until the growth everyone celebrated is the reason the numbers stopped working.