Freight brokerage runs on speed. The broker who quotes first usually wins the load, and the one who quotes an hour later is quoting into a decision that's already made. Our client, a mid-market 3PL brokerage moving roughly 1,200 loads a week, was losing on speed. Quote requests piled up in shared inboxes while brokers hunted through rate history, load boards, and gut feel to price each lane. Average turnaround was 34 minutes. By the time many quotes went out, the shipper had already booked with someone faster. They didn't need more brokers. They needed the quoting to stop being the bottleneck.

The Problem

The brokerage fielded around 900 quote requests a day across email, a customer portal, and EDI. Each request needed a price, and pricing a lane well is not simple. A broker had to consider the origin and destination, equipment type, current market rate, fuel, seasonality, and what that specific customer had paid before.

So brokers did it by hand. They opened each request, checked recent rates on the lane, glanced at a load board to read the market, factored in the customer relationship, and typed back a number. On a busy lane with clean history it took a few minutes. On a thin lane it took much longer, and the quote often sat in a queue while more urgent work jumped ahead.

The numbers were rough:

• Average quote turnaround: 34 minutes
• Daily quote volume: ~900 requests
• Quote-to-book win rate: 22%
• Brokers spending on quoting: roughly half of every broker's day
• Peak load: requests sitting 2+ hours during morning surges

The real cost was not just broker time. It was the loads lost to slower response. In a market where shippers blast the same lane to five brokers, being third to answer means quoting into a closed deal.

Our Approach

We proposed a system that reads inbound quote requests, prices the lane instantly using the broker's own history and live market signals, and only pulls in a human when the lane is genuinely uncertain. Three phases, one principle: automate the routine lanes, escalate the ambiguous ones.

Phase 1: The Data Problem (Weeks 1-2)

Before pricing anything, we pulled three years of the client's quote and booking history. Roughly 400,000 quotes with outcomes attached, meaning we knew not just what they quoted but whether the load booked and at what margin.

The finding that shaped the build: just under 70% of quote volume ran on lanes the brokerage priced regularly, each with enough recent history to model confidently. The long tail of rare lanes was noisy but small. That meant we could deliver most of the value fast by nailing the high-frequency lanes first, then widening coverage over time.

Phase 2: The Pricing Engine (Weeks 3-5)

We built a pricing model that blended three signals: the client's own historical rates on the lane, live market benchmarks pulled from rate data feeds, and the specific customer's booking behavior. An LLM layer handled the messy intake, reading a quote request in whatever format it arrived and pulling out origin, destination, equipment, weight, and dates even when the email was a sloppy paragraph.

The key design choice was confidence scoring on every quote. The engine didn't just produce a rate, it scored how sure it was based on how much recent, relevant history backed the lane. High-confidence quotes on well-traveled lanes could go straight to a customer-ready number. Thin lanes got flagged for a broker with a suggested range attached.

That distinction mattered. A system that confidently guesses on a lane it's never seen is worse than a slow human, because now you've quoted a bad number fast. By making uncertainty explicit, we turned the AI into an instant first pass rather than an unaccountable black box.

Phase 3: Validation and TMS Integration (Weeks 6-7)

A price alone was not enough. Every quote still had to respect business rules before it went out:

• Does the margin clear the customer's contracted floor?
• Is capacity actually available on this lane right now?
• Does the equipment type match what we can source?
• Is this a customer with special pricing or volume commitments?

We wired these checks into the pipeline so validated quotes wrote directly into the client's TMS through its API. Clean, high-confidence quotes posted in seconds and could be sent with one click. Quotes that failed a rule, or carried a low-confidence lane, landed in a broker's queue with the exact issue and a suggested range called out.

The broker's job changed from pricing every request to reviewing the hard ones. Same team, radically different day.

The Technical Stack

For teams interested in the implementation details:

• Ingestion: Unified intake across email, portal, and EDI, normalized into a single quote queue
• Extraction: An LLM layer parsing free-text and structured requests into clean lane, equipment, and date fields
• Pricing: A model blending historical rates, live market benchmarks, and customer-level booking behavior, with confidence scoring per lane
• Validation: A rules engine checking margin floor, capacity, and equipment before any quote is surfaced
• Integration: REST API writing customer-ready quotes into the TMS, with a review queue for low-confidence lanes

Total build time: 7 weeks from kickoff to production. The system covered nearly 70% of volume in the first two weeks live and climbed as we added lanes.

Results

We tracked performance over the first 90 days post-launch.

Quote turnaround: 34 minutes down to under 4 minutes. An 89% reduction in average response time. For high-confidence lanes that flowed straight through, a customer-ready quote was ready almost instantly.

Win rate climbed from 22% to 26%. A 19% relative lift in quote-to-book conversion, driven mostly by speed. When you're first to answer on a competitive lane, you win a lot more of them.

Straight-through quoting hit 64%. Roughly two of every three requests produced a customer-ready quote with no broker touch. The rest flowed through the review queue, where brokers resolved flagged lanes far faster than pricing from scratch.

The morning backlog vanished. Requests that used to sit 2+ hours during surges now returned a quote in minutes, so the brokerage stopped bleeding easy loads to faster competitors before lunch.

The team was redeployed, not cut. Brokers reclaimed roughly half their day from manual pricing and shifted it to carrier relationships and account growth, the work that actually compounds. Capacity went up without new headcount.

Key Takeaways

This project reinforced a few things we believe about AI in operations:

1. The best automation targets a boring, expensive bottleneck. There was nothing glamorous about quoting lanes. That's exactly why it was underinvested and ripe for a large, measurable win.

2. Confidence scoring is what makes AI trustworthy. The difference between a helpful pricing system and a dangerous one is whether it knows when it doesn't know. Explicit uncertainty plus broker review is what protected margin instead of quietly eroding it.

3. Follow the volume, not the edge cases. Nailing the high-frequency lanes delivered most of the value in weeks. Chasing every rare lane first would have delayed the payoff for months.

4. Speed is a strategy, not a nice-to-have. In brokerage, response time is the product. Cutting it from half an hour to minutes changed who won the load.

If your brokerage is losing loads to slow quotes, the pattern is probably already sitting in your shared inbox. The lanes are structured enough to price, and the lost loads are expensive enough to justify it. We can help you build the system, or talk through your use case first.