AI Token Management

The problem

As enterprises shipped LLM features, AI spend became their fastest-growing and least-understood cost line. Token costs were scattered across provider dashboards with no way to answer basic questions: which application drove this spike? What does a customer actually cost to serve? Is a team's usage anomalous or just growing?

What we built

• Unified visibility — one view of token spend and usage across all four major providers.
• Near-real-time anomaly detection — token-spend spikes flagged as they happen, not at end-of-month invoice time.
• Unit economics — cost-per-token, cost-per-request, cost-to-serve per user or customer, cost-per-feature.
• Cost attribution — spend mapped to the applications, users and teams driving it.

Why it isn't cloud cost 2.0

The obvious move was to treat tokens as one more cloud line item. They don't behave like one. Token spend moves on a different cadence, varies model by model, and spikes for product reasons — a retry loop, a prompt change, a feature call pattern — not infrastructure ones. Anomaly detection tuned for cloud bills misses exactly the events that matter here. And attribution has to follow how engineering teams actually structure AI applications: by feature and team, not by account and region.

Try the idea

A simplified, simulated slice of the core loop — spend visibility plus anomaly detection:

Scope, honestly

This is anomaly detection with human-in-the-loop review — the system flags in near-real time, a person decides what to do. It does not auto-remediate, and I wouldn't claim otherwise. Likewise, every spend figure in the demos on this site is illustrative: the product handles real customer data, which is confidential; the demos show what the feature does, not what any customer spent. The “10+ enterprises” adoption figure is real.

My role

I owned the product end to end, from spec to delivery — defining what “unit economics for AI” should mean for an enterprise buyer, making the detection and attribution calls described above, and carrying the capability through to the 10+ enterprises now running on it.