Skip to main content
This is a prototype. AI Token Lens currently runs on a deterministic seeded dataset with no backend and no stored credentials. Two real connectors exist (Anthropic and a ChatGPT Enterprise CSV import); everything else is representative mock data. Treat the numbers in a demo as illustrative, not as your actual bill.
Finance teams can usually tell you what they spend on cloud. Almost nobody can tell you what they spend on AI, because the money leaves through three unrelated doors at once.

​
The three spend layers

Cloud AI

AWS Bedrock, Azure AI Foundry, Google Vertex — metered inference on your cloud bill.

SaaS AI seats

ChatGPT Enterprise, Microsoft 365 Copilot, GitHub Copilot — per-seat licences bought by different departments.

Bespoke apps and agents

Direct OpenAI and Anthropic API keys, gateways, and agent runtimes billed to whoever created the key.
Each layer is visible in a different console, owned by a different team, and denominated differently — per-token, per-seat, per-request. AI Token Lens puts all three in one ledger with per-person and per-agent attribution, budgets, anomaly alerts, and model-choice savings recommendations.

​
The problem it is built around

The demo narrative is “the surprise bill you caught in time.” An agent enters a loop overnight and reprocesses the same batch hundreds of times. Nobody notices, because the spend is spread across an API key nobody owns, on a bill that arrives three weeks later. By then the money is gone and the conversation is an autopsy rather than a decision. Catching it needs three things at once, which is why the views are shaped the way they are:
  1. A single trend line covering all three layers, so a spike is visible at all
  2. Attribution down to a person or an agent, so the spike has an owner
  3. A threshold that fires on day three, not at invoice time

​
What it is not

  • Not a billing system. It reads and reconciles; it does not invoice or pay.
  • Not a chargeback engine. It attributes spend, but does not move money between cost centres.
  • Not real-time. Provider data arrives on provider schedules — daily granularity, not per-request.

​
Where to go next

The five views

Overview, Agents, Alerts & Budgets, People, and Connectors — what each answers.

Connectors

The live Anthropic connector, the ChatGPT Enterprise CSV import, and the honest limits of each.

Running it

Install, run, test, and how the seeded dataset is generated.

Budgets and alerts

Alerts are computed from editable rules, not seeded strings.