Running a $30M freight brokerage on phone calls and spreadsheets pushed Mukesh Kumar to build AI agents that quietly transform margins, service levels, and his team’s daily work.
In freight brokerage, most revolutions happen behind the scenes. Loads still move on trucks. Carriers still wrestle with docks and detention. And brokers still fight for a few extra margin points in a market that rarely forgives mistakes.
For Mukesh Kumar, president and COO of T3RA Logistics in Northern California, the real revolution started with an uncomfortable truth: his team was spending more time in inboxes than with customers.
T3RA moves roughly $30 million in freight each year, sitting between shippers such as major food manufacturers and defense-related lanes on one side and thousands of carriers on the other. The business was healthy, posting around 8% EBITDA, but Kumar, who also serves as the company’s chief technology and growth officer, kept seeing the same pattern in the operation.
“Every time we mapped a load lifecycle, we saw 20 to 40 manual touches,” he recalls. “Quote to tender, tender to booked, appointments, status updates and invoicing—everything was email threads and portal hops. It worked, but it didn’t scale.”
Kumar didn’t arrive in freight as a pure operator. Before T3RA, he spent a decade in engineering roles at HP and Cigna, then co-founded TruckBook, a trucking startup backed by Moneta Ventures that raised a $3 million round. That experience convinced him that data and systems could move just as much freight as trucks and trailers.
When he left TruckBook and started T3RA, he intentionally leaned into brokerage. “I wanted to own the full P&L, not just the software,” he says. “But once we got to scale, I saw the same thing I’d seen in larger logistics organizations—smart people doing repetitive work that computers could do better.”
The turning point came in early 2025, when the latest wave of large language models and orchestration frameworks started to look more practical than experimental. Kumar began to sketch out what he calls a “digital workforce”: a series of agentic AI systems, each with a narrow mandate, hard guardrails, and an audit trail.
“I wasn’t interested in a generalist AI that could ‘do everything’ in freight,” he explains. “I wanted agents that could do one job extremely well, with very clear red lines.”
The first agents targeted the most painful parts of T3RA’s workflow: tendering, appointments, and tracking. The Tender Agent checks required fields, validates documents, assembles response packets, and routes anything outside pre-approved pricing bands to humans. The Appointments Agent reads facility hours and rules, proposes time windows, and books through email or portals, escalating after a defined number of failed attempts. The Tracking Agent sends updates on agreed intervals, tags exceptions with reason codes, and alerts humans when something genuinely needs judgment.
Each agent operates on a simple traffic-light model: green for fully automatic actions, yellow for actions that require one-click human approval, and red for anything it can’t or shouldn’t do. Every step is logged.
“We designed them to fail safely,” Kumar says. “If a portal times out, if a tender is missing fields, if a facility has weird rules, the agent doesn’t guess. It escalates.”
Over several months, the agents shifted from test lanes into production. In a representative cohort of lanes, T3RA saw a double-digit reduction in touches per load, improved on-time performance, and fewer stale confirmations for after-hours appointments. The team moved roughly two full-time-equivalent hours from managing inboxes toward resolving exceptions and growing customer relationships.
The biggest win came from pricing. Kumar and his team built a Pricing Agent that assembles rates based on historical lanes, contract bands, and market inputs, routing outliers to humans. It doesn’t negotiate or commit to penalties, but it dramatically speeds up cycle times.
“Before, it could take us two hours to go from quote to tender and another four hours to get booked on some lanes,” he says. “Now, responses that used to take hours happen in minutes.”
By T3RA’s internal estimates, the pricing agent alone now saves about $40,000 per month in manual work and supports a margin shift from roughly 11% to 15% on their freight book, thanks to tighter pricing and faster responses.
Those numbers, and the discipline behind them, are part of what make Kumar’s story different from the usual “AI will change everything” narrative. He deliberately avoids automating areas that could create disputes or regulatory exposure. There are no AI-negotiated claim settlements, no autonomous commitments with penalties, and no timestamp changes.
“Automation and adjudication are different things,” he says. “Agents can gather data, propose actions, and execute well-defined steps. But final decisions that affect contracts and relationships should still sit with people.”
That philosophy reflects his academic side as well. Kumar has authored several articles on AI negotiation in freight pricing, LLM-driven logistics workflows, and claims automation, exploring where AI should stop and humans should start. Those ideas now run through T3RA’s daily operation.
For many mid-market brokers, his story offers something practical: a path to automation that doesn’t require massive data science teams or science-fiction promises.
“Any brokerage moving a similar volume can start with one lane, one workflow, and one agent,” Kumar says. “You don’t need a PhD. You need clean examples, clear rules, and the discipline to treat agents like coworkers with guardrails, not magic.”
In an industry where margins are thin and talent is stretched, that kind of incremental, auditable innovation may be more transformative than the flashier, long-horizon bets. For T3RA, it started with a founder who had seen both sides of logistics—software and freight—and decided that the next big leap would live in the workflow no one wanted to look at.
