My AI features cost real money every time someone uses them. How do I stop gross margin from collapsing?
Accept the new baseline first: AI native products are running somewhere near 50 to 65 percent gross margin, not the 80 to 90 percent that defined the last decade of SaaS, and inference cost tends to rise as a share of spend rather than fall as usage grows. Then get instrumented. You cannot manage a cost you cannot see per customer, so meter inference per account and watch the tail, because a small minority of users typically drive the bulk of consumption. After that the levers are ordinary: route cheap work to cheap models, cache aggressively, cap or credit the heaviest actions, and put a consumption component in your price so the customers costing you the most also pay the most. If you earn in rupees and pay for compute in dollars, add currency to that list, because a weak rupee quietly eats a margin point at a time.
Go deeper
6 resources, 2 India-specific, 6 link-checked.
🎧 Podcast
✓ Link checkedIndiaFreeIntermediate
Kiran Darisi (Atomicwork) and Sreedhar Peddineni (GTMBuddy), both out of the Freshworks lineage, on why per seat pricing stops making sense when the product is doing the work, and what you sell instead. The practitioner view behind the margin numbers.
The pricing advisor with the widest view of AI deals on how to charge when every use has a real cost: hybrid models, floors, and pricing the outcome without giving away margin. Directly addresses the gross margin trap.
The margin defence most companies are actually reaching for. Poyar's number is the one to remember: roughly 70 to 80 percent of token consumption comes from about 10 percent of users, which is why a flat seat price on an AI product bleeds quietly.
Gives you the current benchmarks to set expectations with your board (roughly 45 to 53 percent depending on how much of the stack you own) and four concrete moves: pick a margin posture deliberately, meter cost per customer, route to cheaper models by default, price with usage.
One metric you can start tracking this month: AI product revenue divided by inference cost. It moves before your gross margin does, which makes it the early warning you want rather than the postmortem you get from the P&L.
Inference efficiency ratio is AI product revenue divided by inference cost for the same period.
For AI infused SaaS, 8:1 is the floor and 10:1 or better is healthy, below 5:1 is a warning.
For AI native products the healthy zone is 5:1 or better, since inference is structurally about 20 percent of revenue.
Unlike gross margin, which lags, IER is a leading signal: model routing, prompt caching and tiered pricing moved one example from 4.4:1 to 8.0:1 at the same revenue.
Names the version of this problem that is specific to Indian companies: revenue in rupees, inference billed in dollars, and a currency move you did not budget for eating margin. If you sell in INR and run on foreign models, this is your risk in one page.
