AI is in the product and in the burn rate. Frontier-model overkill on routine features is the fastest way to watch gross margin evaporate as you scale.
For a SaaS company, AI COGS is the new AWS bill, except it scales with feature usage instead of headcount, and it is set by someone else's price list. Every AI feature shipped on a frontier model by default is a margin decision made implicitly, usually by whoever wrote the first prototype.
The pattern we see in audits is consistent: a large share of production tokens go to tasks a specialized small model handles indistinguishably: classification, extraction, formatting, first-draft generation. Users cannot tell the difference; the gross-margin line can.
Agent features raise the stakes: a loop that retries, reflects, and fans out can spend tokens super-linearly with no user-visible signal. Telemetry per feature and per tenant, with hard ceilings, is the difference between a usage spike and a board-meeting surprise.
SLM-first routing for in-product AI features, with frontier models reserved for the genuinely hard reasoning users actually pay for.
Per-tenant boundaries, hard budget ceilings, and governed agent loops that can't run the meter unattended.
A token-spend optimization audit and a routing-plus-telemetry layer that protects margin as usage, and the bill, keep climbing.
Every AI call classified and routed: SLM by default, frontier by explicit escalation with a price tag.
Cost per feature, per tenant, per model, in the same dashboards as your infra metrics.
Hard caps per workflow and tenant; agent loops degrade gracefully instead of running the meter.
Per-tenant data isolation through the AI layer, matching the promises in your DPA.
Blind bake-offs before any downgrade ships, so margin work never degrades the product silently.
Not if it is gated: we run blind bake-offs on your real traffic and only route task classes where the small model matches or beats the incumbent at a pass/fail threshold. The escalation path stays for the hard cases.
It depends on your task mix, which is the point of the audit. Routine-task tokens dominate most SaaS workloads, and those are exactly the tokens a specialized model serves at a fraction of the cost. The TCO calculator gives a directional answer in minutes.
The routing decision is cheapest before the traffic arrives. A day of architecture now (model-agnostic call sites, telemetry from the first feature) prevents the margin rebuild later. That is exactly what the implementation engagement installs.