Bulk CSV: turn LLM token exports into kg CO₂e (finance & CSRD prep)

Once you have model names and token counts in a spreadsheet, the last step for CSRD-style activity data is to turn those rows into CO₂e with the same coefficients you use in production. The bulk CSV tool and POST /api/v1/batch-estimate do exactly that: no prompt bodies, only metadata.

1. What to import

The parser accepts common export shapes. You need a model column (aliases include deployment_name for Azure OpenAI) plus token counts:

• Classic: prompt_tokens + completion_tokens
• OpenAI-style: input_tokens + output_tokens
• Total + one side: total_tokens with either prompt or completion to infer the other
• Total only: total_tokens with model — the tool applies a 50/50 split for coefficient lookup; the sum of tokens is unchanged, but you should state that assumption in your methodology note

Delimiters: comma, semicolon (typical for EU Excel), or tab. A UTF-8 BOM on the first cell is fine.

2. Where the numbers usually come from

Hosted APIs (OpenAI, Anthropic, Azure OpenAI) return usage on the response. Persist those fields next to your cost centre or tenant id, then export to CSV for a month or quarter. That path matches our field-by-field logging note.

Microsoft 365 Copilot embedded in Word, Outlook, or Teams is a different product boundary: there is no universal “monthly tokens.csv” for every tenant comparable to an API log. For Copilot, you may only have adoption metrics unless you control a server-side integration. Use bulk CSV when you already have token rows from a system you instrument — not as a substitute for data Microsoft does not expose.

3. From spreadsheet to audit trail

Run the batch estimator in the browser (paste or upload) or call POST /api/v1/batch-estimate with the same payload for automation. Download the results CSV to attach alongside your coefficient snapshot version and scope (dynamic vs lifecycle).