Somewhere in a freight brokerage right now, an operator is on their fourth phone call of the morning, and it is not yet 7am. The call is a check call - a simple ask about whether a truck is where it is supposed to be. It will be repeated dozens of times today, alongside chasing a proof of delivery that a driver forgot to send, keying a rate confirmation into three systems, and auditing an invoice that does not match the quote. This is the job. It is also the thing Burt wants to take off their plate.
Burt, a company in Y Combinator's Winter 2026 batch, builds what it calls AI teammates for logistics. The pitch is narrow on purpose: it helps freight brokers and forwarders stand up AI coworkers that handle the repetitive, manual work end to end, so the humans can go back to the parts of the business that actually need a human - customers, carriers, and closing new lanes.
The problemThin margins, thick manual work
Freight brokerage is a relationship business that runs on razor-thin margins. The awkward part is that most of a brokerage team's day does not go into relationships at all. It goes into shipment tracking and check calls, document processing and data entry, chasing PODs and rate cons, finding and booking carriers, reconciling invoices, quoting, and building loads. As a brokerage grows, the volume of this work grows with it, and the usual answer is to hire more people to absorb it. That works until it doesn't - the team burns out, service slips, and growth stalls against the cost of headcount.
Burt's argument is that this pile of work is exactly the kind of thing software should have eaten years ago, and the reason it hasn't is that most software just adds another screen. A dashboard tells you a POD is missing. It does not go get the POD.
How it worksAn agent that works the way an operator does
The design choice that makes Burt interesting is that its agents do not ask a brokerage to change how it operates. They sit on top of the systems a team already uses and follow that team's specific processes. In the company's own description, the AI teammates work the way a human operator would: reading emails, making calls, logging into portals, and updating their transportation management system, then pulling in a person only when one is genuinely needed.
In practice, that loop covers a specific list of jobs. Burt's teammates run morning pickup audits, answer ETA requests, monitor shipments for delays and flag the ones going sideways, chase and process documents, cover loads, and reconcile invoices against quotes. The company frames it as "full visibility, full control," with every action logged so a supervisor can see what the agent did and why. The point is not a black box that takes over the desk. It is a coworker whose work you can read line by line.
There is a trust problem baked into any product that promises to act on your behalf, and Burt handles it with a staged rollout that reads like teaching someone to drive. First is Learn, where a team demonstrates its processes and connects its systems. Then Training Wheels, where the agent proposes actions and a human approves or corrects them. Only after it has earned it does the agent reach Autonomy and operate on its own. Each correction and exception, the company says, becomes permanent knowledge, so the same mistake does not come back next week.
The expertiseThey built the model before the product
Burt did not start as a logistics company. Its first public launch, in February 2026, was about something more foundational: training and deploying specialized models that outperform general, closed-source alternatives while being cheaper and faster to run. The team built an in-house post-training and inference stack for exactly this. In one case study, a customer had a model call running at very high volume that was too slow even on gemini-3-flash. Burt built a small vision language model for the task that came in roughly three times faster at the median while being more accurate than the frontier model it replaced.
That lineage matters, because it is the engine underneath the freight agents. Burt is not simply wrapping a general chatbot in a logistics costume. It is pointing a model-training capability at a domain where speed, cost, and reliability are not nice-to-haves but the difference between a tool a busy desk will use and one it will quietly abandon. The two domains still share a fingerprint in the company's URLs - it operates under both burthq.com and trainburt.com, a small fossil of the pivot.
The foundersRaised inside the warehouse
Most founders discover their market. Bobby Zhong and Kurt Sharma were, in a sense, raised in theirs. Zhong, the CEO, spent weekends at his father's third-party logistics warehouse. Sharma, the CTO, grew up watching his uncle sell on Amazon FBA, where half the stress was always coordinating shipments and chasing freight. That shared background is not decoration on a pitch deck. It is why the product is built around respect for the operator rather than a plan to replace them.
Studied CS at UC Irvine before dropping out to become the second engineer at Pirros (YC W23), then joined Replo (YC S21) building coding agents and wrangling LLMs.
Studied EECS at UC Berkeley. Previously at Replo (YC S21) building high-throughput data pipelines and sandboxed code-manipulation systems. Now leads Burt's model stack.
The two worked together at Replo before starting Burt, which explains how a two-person company ships both a model-training stack and an operations agent. When they describe the mission, they keep coming back to the same line: they are building for the people that keep the world moving, and there is no one else they would rather build for than the people in freight.
It is worth sitting with how unusual that focus is. The flashiest problem in technology right now is AI agents. Two engineers with the resumes to work on almost anything looked at that problem and pointed it at freight - an old-line industry of phone calls, spreadsheets, and PDFs that most of Silicon Valley has never had a reason to think about. The advantage is knowledge that does not show up in a pitch. They know what a rate con is, why a POD goes missing, and how a broker's day actually falls apart. That is the kind of detail you cannot fake, and it is the reason the product describes real work instead of a generic automation dream.
The marketWhere Burt fits
Burt is not alone in noticing that freight is a target-rich environment for AI. A cluster of startups is chasing pieces of the same workflow, and incumbent TMS platforms are bolting on their own automation. The traditional answer to back-office overload has been to offshore it to a BPO team. Burt's wedge against all of these is the combination of two things it can point to: agents that live on top of a broker's existing tools rather than asking for a rip-and-replace, and the specialized-model muscle to make those agents fast and cheap enough to run at freight-desk volume.
The business model follows the shape of the product. Burt sells AI teammates to freight brokers, forwarders, and carriers, onboarding them through the Learn, Training Wheels, and Autonomy path, with a demo-led motion. The company is early - a two-person team, a seed round of about $130,000 through YC, and a customer base it is still building lane by lane. What it has is a clear picture of a specific person's Tuesday and a plausible claim that it can hand a good chunk of that Tuesday to software.
Whether Burt becomes the default AI layer for freight or one of several in a crowded field is an open question. The more interesting bet is the underlying one: that in an industry running on thin margins and thick manual work, the winning product is not the one with the best dashboard, but the one that quietly does the work and knows when to ask for help.