All posts

How we meter AI usage in a multi-tenant SaaS

AI API costs are per-token and unpredictable. Here is how we built usage metering that gives tenants visibility, enforces quotas fairly, and prevents one tenant from running up everyone else's costs.

Engineering5 min read·
I
Ibrahim Chehab

The problem

AI API costs are a function of token usage — which varies dramatically by use case, tenant, and conversation pattern. A tenant running complex multi-step AI workflows uses 10x the tokens of a tenant doing simple FAQ lookups. In a flat-pricing model, power users subsidise light users (fine) — but if a single tenant spikes unexpectedly, the platform owner absorbs the cost.

The solution: metered billing with per-tenant quotas, real-time usage visibility, and graceful enforcement.

What we meter

We meter three things: - AI tokens consumed (input + output), per model tier - Messages processed (incoming + AI-generated), per channel - Knowledge base queries (vector search operations per month)

Each is tracked in a usage_events table with tenant ID, event type, quantity, and timestamp. This table is append-only and feeds both billing and the tenant dashboard.

Quota enforcement

Each plan defines a monthly quota per metric. At 80% of quota, we send an email alert. At 100%, we have two options:

  1. 1.Hard stop — block further AI processing, serve a fallback "our AI is temporarily unavailable" response to customers.
  2. 2.Overage billing — continue processing, charge the overage at a per-unit rate.

Enterprise tenants get overage by default; SMB plans get hard stop with an option to upgrade. This is a business decision, not just a technical one — and it is important to communicate clearly in the plan terms.

Preventing abuse

The most common abuse pattern: a tenant enables public-facing AI with no rate limiting on the widget, a bot hits the widget endpoint continuously, usage spikes to 100x normal. We prevent this with:

  • Per-session rate limiting on the widget endpoint (10 messages per minute per session)
  • Per-tenant daily soft cap (alert at 5x daily average, hard stop at 10x)
  • Anomaly detection that compares rolling 24h usage to the tenant's historical baseline

Visibility

Tenants can see their usage in real time in the dashboard: tokens today, this month, vs. quota. Graphs show usage by channel and by AI model tier. This visibility reduces support tickets ("am I about to hit my limit?") and helps tenants make informed plan upgrade decisions.

The investment in metering infrastructure paid back in the first month: one tenant's runaway bot would have cost us $4,000 in API fees. The anomaly detection caught it in 20 minutes.