Opens in a new tab
Logistics · AI Agents

Freight-quote agents that turn inbox chaos into booked shipments

Quote requests arrived as unstructured emails in every format imaginable. We built a multi-agent workflow that reads, extracts, prices against carrier rules, and drafts responses — with the ops team approving every send.

The challenge

Every quote started as an email: some structured, most not, in multiple languages, with attachments ranging from spreadsheets to photos of documents. Ops staff re-typed details into the pricing system, cross-checked carrier rules by memory, and drafted replies by hand — slow during business hours, stopped entirely outside them.

The approach

The spec mapped the whole flow and drew the human line explicitly: agents prepare, humans approve. We built a pipeline of specialized agents — one extracts and normalizes request details, one prices against codified carrier rules, one drafts the response — orchestrated so exceptions route to a person instead of guessing. The evaluation suite was built from months of real historical emails, and the system had to clear it before touching a live inbox.

What shipped

A production agent workflow connected to the ops inbox: requests are parsed and priced within the rules, draft replies appear in a review queue with the agent's reasoning attached, and one click sends. Anything ambiguous is flagged for a human — by design, not by failure.

Placeholder narrative for design and structure review — details, client identity, and specifics to be replaced with a client-approved write-up before launch.

Under the hood

What it took.

Multi-agent orchestrationEmail & document parsingMCP integrationsRules-engine pricingHuman-in-the-loop reviewEvalsGoogle Cloud
The SoftmaxAI Method

This build ran the standard four stages — spec approved before code.

01 / SPECSpec

A written, client-approved specification before any code — architecture, edge cases, acceptance criteria.

02 / BUILDBuild, with agents

Engineers orchestrate AI agents executing the spec in parallel. Weeks, not quarters.

03 / VERIFYVerify

Evals, tests, and security checks alongside the build. Humans review every decision that matters.

04 / SHIPShip & improve

Deployed to your GCP/AWS with monitoring, production evals, and retraining loops.

Start here

Have an AI idea? Let’s pressure-test it — free.

A 30-minute call with our engineers, not a salesperson. You’ll get an honest feasibility read, a rough architecture, and a clear next step.