Building an AI agent that chases supplier quotes: what we learned about RFQ emails
Описание и применение
Автоматизированный агент Dov отправляет структурированные запросы цен, отслеживает ответы, парсит различные форматы котировок, нормализует стоимость до уровня доставки и предоставляет сравнение цен для утверждения человека.
Задача
Automating the procurement process of sending RFQs, chasing supplier responses, and comparing quotes while handling unstructured data and ensuring compliance.
Результат
The AI agent reduced manual follow‑ups, ensured consistent RFQ structure, parsed diverse quote formats into comparable tables, and enabled human approval of landed‑cost based decisions.
Рабочий процесс
- 1. Draft a structured RFQ with part description, material, tolerances, finish, quantity tiers, Incoterms, lead time, and numbered questions.
- 2. Send RFQ via email and log the message.
- 3. After 2 business days, send first follow‑up; after 4 more days, send second; finally send a closing nudge.
- 4. Receive quotes in various formats (PDF, Excel, text, photo).
- 5. Extract text and tables from each format.
- 6. Map extracted data onto a fixed schema (unit price, currency, MOQ, tooling cost, lead time, payment terms, validity date).
- 7. Flag missing fields and send targeted follow‑ups to suppliers.
- 8. Normalize each quote to landed cost per unit, including freight, duties, tooling amortization, payment terms, and lead‑time cost.
- 9. Compile a comparison table with assumptions for buyer review.
- 10. Human approves shortlist and final decision before payment.
Доказательства
Our AI agent, Dov, sends a lot of RFQs. Somewhere around the thousandth email we realized that "request for quotation" is a generous name for what actually happens: you send a clear question and receive back a PDF named final_v3_REAL.pdf, a one-line reply saying "price attached", or silence.
Numbered questions are the single best trick we found. Suppliers tend to answer in the same order, which makes parsing far easier.
Most quotes arrive after a nudge, not after the first email. Our cadence is roughly: a first follow-up after 2 business days, a second after 4 more, and a short final "should I close this out?" message.
Quotes arrive as PDFs, Excel files, inline text, and the occasional photographed price list. The pipeline we settled on: 1) Extract text and tables from whatever format arrived.
The rule: never let the model silently fill a gap. A blank cell is information. An invented price is a lawsuit.
A quote of $1.10 per unit EXW from far away is not cheaper than $1.35 delivered nearby. Real comparison needs freight, duties, tooling amortization, payment terms, and the cost of a long lead time.
Agents are great at finding suppliers, sending clean RFQs, chasing, and building the comparison table. They should not unilaterally commit money.
- Дата записи: 2026-10-21T00:56:02.421Z