Start with one question
Before connecting an AI agent to Amazon data, write down the question you want it to answer. For example: which campaigns account for the most ad spend in a chosen period? Treat this as an analysis task, with a named account and a date range you can check. [cs]
Commerce Spine provides normalized Amazon data through a read-only API for agents, scripts and automations. You bring the client that asks for the data. The API response gives you something concrete to inspect before you rely on an agent's explanation. [cs]
Choose how to connect
The API accepts requests at a single /graphql route. Protected requests use an agent bearer token in the Authorization header. Keep that token out of shared prompts, article drafts and public repositories. Use your client's secret handling instead. [cs]
Check access with amazonAccounts before asking for performance data. It returns the accounts allowed by the organization and user permissions. Choose an account from that response. A familiar store name in a prompt is not proof that the connection can read it. [cs]
Check what the data covers
Ask for dataFreshness to see the latest available date before choosing the period to analyze. Use dataCatalog or dataAsset to inspect available fields. This makes it easier to distinguish a field that exists from one an agent has merely suggested. [cs]
Amazon's Reports API covers report retrieval and management, including inventory and orders. Amazon also warns that report fields and formats can change. If you maintain a direct report integration, build those changes into your maintenance checks. Commerce Spine's reference is the place to check the contract exposed by its own API. [amazon] [cs]
Check the answer against the response
For a first analysis, request a small set of fields and keep the returned rows alongside the agent's explanation. Check that the account and period match your question. If you rank campaigns by spend, make sure the ranking uses spend rather than a similarly named revenue field. [cs]
Results include pagination information. Follow the returned cursor when there is another page; do not treat the first page as the complete result. The response also includes query metadata that can help you inspect what came back. [cs]
A useful instruction for your agent is: "Show the account, period and fields you used. State whether more pages remain. Separate observations in the returned data from suggestions that need investigation." This is a suggested review habit, not a claim that every client applies it automatically. [cs]
Know where reading ends
This guide covers the read-only data API. A response showing campaign spend does not establish that a bid or budget changed. Keep any proposed action in a separate review step, and confirm which tool would perform it before proceeding. [cs]
Start with one question you can verify. Once you can trace the answer back to the response, expand the workflow to more accounts or recurring analysis. No performance improvement is promised here; the first objective is a connection whose answers you can check. [cs]