The AI usage section tracks your team's spending on large language model APIs — OpenAI, Anthropic, Google, and others. It answers questions like "which workflow is costing the most?" and "which user is driving 60% of our AI spend?"
What to upload
A CSV where each row is one LLM API call or a daily/monthly rollup of calls. Columns:
| Column | Required | Notes |
|---|---|---|
provider | Yes | e.g. openai, anthropic, google |
model | Yes | e.g. gpt-4o, claude-3-5-sonnet |
workflow_type | Yes | e.g. email-draft, data-extract |
source_type | No | e.g. api, sdk, zapier |
user_email | No | Who triggered the call |
occurred_at | Yes | ISO timestamp or date |
input_tokens | No | Input token count |
output_tokens | No | Output token count |
cost_usd | Yes | Cost for this row in USD |
Download the exact template from AI Usage → Import → Download template.
Where to get this data
OpenAI — go to platform.openai.com → Usage → Export.
Anthropic — go to console.anthropic.com → Usage → Export CSV.
Google (Gemini/Vertex) — export from your Google Cloud billing dashboard, filtered to AI/ML services.
Your own logs — if your team routes all AI calls through an internal gateway or proxy, export from there and map your column names to the template.
How to import
- Go to AI Usage → Import in the left sidebar.
- Click Choose file and select your CSV.
- Click Import.
The AI usage importer is stricter than the spend importer — column names must match exactly.
What you see after import
- Summary cards — total cost by provider, model, and workflow
- Event table — filterable by provider, model, workflow type, and source
- Cost trend — month-over-month spend per provider
- Monthly delta — how much this month compares to last month
Spike detection
If Nagata sees month-over-month AI spend growth of 50% or more and the dollar delta is at least $200, it creates an alert automatically. This requires at least two months of data to fire.
Common problems
Column names must match exactly — unlike the spend importer, AI usage won't guess column mappings.
cost_usd must be a plain number — 0.0032 works; $0.0032 does not.
Only one month of data — the spike alert and MoM comparisons require at least two months.