In the not-so-glamorous world of freight brokerage, where success is measured in margin points and minutes saved on quote cycles, Mukesh Kumar is showing that AI doesn’t have to be flashy to make a difference. Kumar is the president and COO of T3RA Logistics, a Northern California company that moves $30 million worth of freight every year between enterprise and defense lanes. He has carefully set up a group of narrow AI agents to do the boring, inbox-heavy work that used to take up his team’s days. The results are clear to brokers: Kumar’s pricing agent alone saves about $40,000 a month by shortening pricing cycles from hours to minutes. This helps the gross margin go up from about 11% to 15% year over year. But the three operational agents he’s made for bidding, appointments, and tracking—agents that any mid-market store could make—show his bigger idea: ship small, keep the scope tight, let people run things, and show the differences.
From manual work to digital coworkers
Kumar did not come to this conclusion quickly. His academic work, which included papers on AI negotiation in freight pricing and LLM-driven logistics workflows that were published in Scholar Publishing and the International Journal of Computer Trends and Technology, gave the theory a base. But theory and reality came together in T3RA’s live operation, where a typical full-truckload move used to need 20 to 40 manual touches for things like setting up appointments, tendering, and status updates. “Every touch is a chance for a delay, rework, or a missed service-level agreement,” says Kumar. “Before agents, it took us an average of 120 minutes to go from quote to tender. It took another 240 minutes from tender to booked. Late-night appointment confirmations were pushed back by off-hour facilities, and our team was buried in inbox work that should have been done in a more organized way.”
When Kumar changed the way he thought about the problem, it got better. He didn’t want to make a general AI that could do everything and might make costly mistakes. Instead, he made what his team calls a “digital workforce”: narrow, auditable agents, each doing a specific job with strict rules and a visible kill switch.
Three agents, three jobs
The Tender Agent never makes up a rate. It checks the required documents, verifies the tender fields, and puts together response packets, only using approved pricing bands or sending the data to humans when it goes outside of the parameters. The Appointments Agent suggests possible times based on the hours of the facility and any other rules that apply. They book appointments by email or through the portal, following the rules for that facility, and if they fail three times, they escalate the issue. The Tracking Agent sends updates on time at agreed-upon intervals, makes exceptions with reason codes, and lets people know when a variance goes over a certain level. However, it can’t change timestamps or close exceptions without human approval.
Kumar says, “Agents aren’t interns; they’re coworkers with audit trails. Each runs on a traffic-light model: green for automatic actions that happen all the time, yellow for actions that need human approval with one click, and red for actions that need to be blocked and escalated.” An unchangeable audit record keeps track of every step, so operations, compliance, and customers can figure out what happened and why.
This design is based on Kumar’s main idea, which he talked about in his research on carrier outreach and detention claim automation: when the dock door is stuck and the shipper wants answers, guardrails are better than being smart. A tendering agent that never sets prices is safer than a generalist who sometimes makes up rates. Less scope means fewer data joins and fewer chances to fail.
The Numbers That Tell the Story
In a representative multi-week cohort on production lanes, T3RA Logistics reports a material double-digit reduction in touches per load, measurable lifts in on-time-in-full delivery, and a decline in exception rates from pre-agent baselines. There were fewer stale confirmations for after-hours appointments, and coverage was better. The team moved about two full-time-equivalent hours from managing the inbox to resolving exceptions and developing customers. Kumar says, “The effect is as much cultural as it is technical. Teams spend less time searching through their inboxes and more time fixing problems and calling customers. That’s where real relationships and money are.”
The pricing agent saved $40,000 a month by getting rid of repetitive rate-building work, but Kumar is quick to point out the compounding benefits: faster response times, better win rates on competitive lanes, and tighter margins on known routes that freed up capital for growth lanes. The increase in gross margin from 11% to 15% was due to both cost savings and better use of human judgment.
Made to Fail Safely
Kumar’s system expects things to go wrong at every level. If a tender is unclear, a “Tender Sanity” pre-check will reject it or flag any missing fields before anyone touches it. When a portal times out, it uses exponential back-off and a human nudge after a set amount of time. Spam updates are limited by confidence levels, and periods of no change go into digests instead of filling up inboxes.
He says, “Data reality in freight is noisy. Facility calendars, portals, and reference numbers are all things that cause problems. Instead of ignoring it, we planned for that friction.”
What Kumar won’t automate is just as telling: no negotiating exceptions, no commitments that come with penalties, and no changes to timestamps. On purpose, those stay with people. This keeps the line between automation and adjudication that he’s looked into in his research on claims handling.
A Guide for the Middle Market
Kumar thinks that any mid-market brokerage can do this in 30 days. In the first week, draw a map of your swimlanes, choose one lane (like tendering), gather 50 clean examples, and write down the red-yellow-green rules. Week two: set up a sandbox, connect read-only integrations, and run shadow mode without writing. In week three, switch to supervised mode with one operations lead, keep track of touches and service levels, and make the guardrails tighter. Week four: add a second lane and send out an Agent Governance memo that talks about KPIs, rules about conflicts of interest, and red lines. Kumar says, “The mandate is practical. You don’t need a PhD or a business round. You need clear boundaries, clean data, and the self-control to treat agents like coworkers, not magic.”
Thought Leadership Based on Operations
Kumar’s academic work, like his recent paper in Scholar Publishing about how AI negotiation is changing freight RFPs and spot markets, comes directly from his work in the field. His ResearchGate preprint on reaching out to carriers and his IJCTT analysis of automating detention claims give procurement teams ideas for how to check out vendors and pilots.
“Freight doesn’t reward theoretical elegance,” he says. “It rewards systems that show up, write down what they did, and make the next day’s work easier.”
As logistics technology makes bigger and bigger promises, like self-driving trucks, blockchain provenance, and predictive everything, Kumar’s vision stays focused on what he calls “narrow agents with audit trails.” Any business that moves $30 million worth of freight a year and has a problem that costs $40,000 a month can use this method. And in an industry where margins are always getting smaller and workers are hard to find, that kind of practical innovation might be more important than the big ideas.
Kumar ends by saying, “We’re not trying to reinvent freight. We’re just trying to give our team back their time so they can do the work that really needs judgment. The agents keep the work going. People are still in charge of the relationships. That’s the way it should be.”
