Export machine-readable attribution for an agent

Three routes, and the right one depends on whether the agent needs to ask questions or just receive answers.

The three

Route one: the export

Run the read, take the output file, put it where the agent can see it. Adequate for the common case where a model writes a monthly summary from a fixed set of numbers. The €99 one-time read on a Google Analytics upload produces it, refundable within 30 days, no questions asked.

Route two: the API pull

A scheduled job pulls the read and writes it into whatever context store your agent uses.

Scope the key to reading reports and nothing else, and record where it lives and who can rotate it.

The field checklist before you hand it over

Run this regardless of route:

Is the window stated inside the payload rather than implied by the filename.

The last one catches a surprising number of problems. A file whose period is only in its name will be summarised as though it were current, forever.

What to put in the prompt, not the data

Two instructions that belong in the system prompt rather than in the payload: treat an interval spanning break-even as unresolved, and never recommend a budget move on a channel labelled unmeasurable. Both are policies rather than facts, so they belong on the agent's side.

What not to build

Autonomous budget changes. There is no such capability here and there is a reason: an observational estimate needs a human weighing context the model cannot see. The wider argument is in can an AI agent do your attribution and the alerting version in why real-time attribution alerts mislead.

Related answers

• What an agent can actually do with attribution data
• What your team should know before agents read the data
• What agents get wrong about attribution data
• Read an attribution methodology doc in 15 minutes
• Measure Shopify channel lift in a single afternoon