| inputs | Email body text, PDF, images (drawings, photos), Excel up to hundreds of lines; agent requests over MCP, REST and a generated CLI. |
| classification | Each email is sorted into one of five kinds: a real RFQ, a vague inquiry, an internal forward, a reply in an existing conversation, or not an RFQ. Thresholds and your internal domains are set per customer. |
| extraction | The order details and each line, with a confidence score per field and a pointer back to where it came from in the document. A request whose lines do not add up is held for a person. |
| matching | Six tiers: customer part number, tenant part number, product code, fuzzy name, catalog search, model. Corrections become cross-references. |
| quote targets | Salesforce CPQ (SBQQ quote and quote lines) in production. Revenue Cloud Advanced transaction objects by the end of 2026. Draft opportunity and quote, your calculator, your price books, contracted prices, documents attached as Files. |
| release rule | Set per customer: which buyer types may receive a quote automatically, the minimum confidence, a value cap, what to do with a buyer's first quote or a revision, and the response-time promise. One decision, one check before anything goes out, every disclosure recorded. |
| modes | Shadow (everything runs, nothing goes out), live, paused. An unusual burst of traffic or a drop in quality pauses a customer automatically. |
| tenancy | Every record is tagged with the customer it belongs to. Requests per minute are limited per customer. Tested live for leaks between customers and between buyers. |
| quality | A fixed set of test emails must pass before any software release. Nightly runs are recorded. A sample of delivered quotes is audited. Each customer's quality is watched for drift. |
| salesforce footprint | One login with only the access it needs, and a small installed package: an entry point, review-status fields, a queue, a notification flow, and a review panel. |