AWS shows how AI can prepare supply-chain decisions for human approval

In the AWS walkthrough, a supplier emails prices in response to a quotation request. The agent checks whether the supplier is approved and prices meet purchasing guidelines, then prepares a review card. A person compares responses and selects a supplier and issues the purchase order, or rejects the response.

The sample's agent instructions explain the division of work. Operational records belong in the connected data space; connectors supply business rules, such as approval thresholds. The agent is instructed to retrieve facts first, pass them to the relevant connector, then explain the decision with its numbers and rule. Before a write or commitment, it must summarize and request confirmation. Missing information must be identified rather than guessed. These are prescribed behaviors in the configuration.

Access also requires explicit setup. Under the registry documentation, after an administrator links the registry, users choose a prefilled connector card and authenticate with their own credentials. They then select enabled tools, create the connector and share it with relevant groups. The registry does not hold those credentials. Removing its link later leaves already-created connectors in place; discovery and the connector's continued existence are separate states.

The sample demonstration documents invoice checking with synthetic data. Its setup requires a deployed Space populated with datasets and a document knowledge base, completed dataset ingestion, and an agent connected to that Space. Invoice and governance connectors must also be attached, with access shared to the relevant users. In that configured environment, the documented invoice example matches the purchase order and calls the rules, returning approval or director review with the variance percentage. This describes an expected demonstration result, not our own test or evidence of real payments.