AI agents that learn each customer's unwritten rules and do logistics back office work.

Coordinators retyping what someone already wrote

Pallet starts from a scene its founder describes in the company's Series B letter. In the early days he visited a trucking company whose owner had a stack of more than 100 paper work orders to key into a computer, six hours every weekend. Pallet's account is that this is ordinary in logistics. A single shipment passes between shippers, carriers, drivers, warehouses and brokers, each working in a different system, and the coordination happens through email, PDFs and web portals. The company says about 10% of supply chain spend goes to that administrative work, and its company page says two thirds of operations teams still run on spreadsheets and tribal knowledge, the unwritten rules experienced staff carry in their heads.

An order starts and gets sent to you, and you can deliver it without a human in the loop. I think that’s where the whole industry is headed.

Sushanth Raman · CEO and co-founder · TechCrunch, at the Series A · 2024-10-02

Why the old automation stalled

Sushanth Raman, the CEO, says he got his start in supply chain building an OCR system, software that reads documents by fixed templates. In an October 2025 post he writes that one mid sized carrier needed 500 bill of lading templates, and that the real work began after extraction: customer names coded differently in the TMS, liftgates no one writes down, loads that must be consolidated when they go to the same place in the same week. When GPT-4 arrived in 2023, he writes, he concluded that template tools would become obsolete, because a language model can read open formats and follow instructions written like training for a new employee. That is the belief that sets Pallet apart: the scarce skill is not reading the document but knowing which rule applies to which customer, and software can capture and apply those rules.

The company has turned that belief into named layers. Its CTO, Nilkanth Patel, wrote in September 2025 that the Memory Layer stores rules as plain English facts, such as a chassis rental price or an approval threshold for a discount, and that a second model acts as a judge before an agent's work goes out. Pallet's posts describe brokerages averaging 20 custom rules per customer across 200 customers, about 4,000 rules for one operation. The reason this works now is partly the models and partly the plumbing. By 2026 the agents could also work inside AS/400 green screens, remote desktops and browser portals, the places where much logistics work still happens.

One order, followed through an agent

Take an illustrative case: a freight forwarder that receives shipping requests by email. The agent reads the email and its attachments, whether PDFs, spreadsheets or scanned photos. It pulls the memories tied to that customer, such as default terms the customer never states, the facility's receiving hours and the carrier to prefer on a lane. It drafts the order, checks it against those rules, and writes it into the system of record, CargoWise for a forwarder or a TMS through EDI for a carrier. Pallet says a task can run through several models from different providers, and that disagreement sends the item to a person. When an operator fixes an exception and explains why, Pallet turns the explanation into a new memory, tests it against past cases and applies it the next time.

People's judgement stays in two places. Operators handle the items the system flags and decide which rules are true. During deployment, Pallet's own engineers sit with those operators: the Eassons story describes the team working on site in Nova Scotia, and the Prism story says Pallet was in Prism's Poland office within days of signing. The commercial promise rests on that last step. Pallet's product pages say customers start with a proof of concept and pay for what the agents deliver, not for seats or licenses.

What a buyer actually checks

The case Pallet makes is arithmetic, not novelty: minutes per order, people per desk, error rates on filings a bank or customs office will reject. A buyer can check that arithmetic in a pilot, which is why the Seoul posting is built around discovery and structured pilots rather than demos.

The Seoul Sales Director sells this loop to Korea's largest logistics providers. The posting asks for discovery sessions that map a customer's workflows to agent capabilities, then pilots that show measurable ROI. In practice that means learning a forwarder's or a 3PL's order intake well enough to name the minutes and errors a pilot should move.

Agents: the work customers pay for

Pallet sells one core product, its agents, and two products that sit around them. The agents, sold in 2025 under the name CoPallet, take on order entry, shipment execution, customer service, billing and settlement, and customs filings. They work in the channels operators already use: email, phone, documents, APIs and EDI, browsers and remote desktops. Underneath sit three parts the company names. Memory holds each customer's rules. Fabric, which Pallet calls its MCP layer, connects agents to systems of record, down to AS/400. Forge, announced in May 2026, builds an agent by learning from a customer's historical emails, chats and operational data, then runs thousands of simulations before go live. Pallet says most custom agents go live in about six weeks.

Atlas, launched in March 2026, is the second product. It pulls data from a customer's TMS, WMS, ERP, email and documents, and links customers, subsidiaries, facilities, lanes and contracts into one model of the network. The launch post sold it as a way to find growth, such as lanes a carrier could win near customers it already serves. The current product page frames it as planning: whether to buy, transfer or expedite parts, how to allocate scarce inventory, which production slot to promise. An approved decision can be handed to agents to carry out. Atlas competes with business intelligence tools and with the spreadsheets planners build by hand.

The third is Custom Models, announced in July 2026: a model trained on one customer's workflow data, never shared with another customer, and able to run on infrastructure the customer controls. Pallet's research group reported that its first fine tuned production model reached frontier level accuracy at roughly one third of the cost, and that for that deployment the break even point was about $750 a day of frontier model spend. The pitch is cost and ownership. As an operation adds rules, a general model has to carry more context on every call, while a dedicated model learns the stable patterns once.

Who has bought it, and what they report

Pallet publishes five named customer stories, each quoting the customer's own executives. Lineage, which its story describes as the largest temperature controlled warehouse operator, with more than 480 facilities, runs three agents across receiving, freight forwarding and warehouse orders. The page reports 2 million AI runs a year at 99% average accuracy, order entry cut from 15 minutes to 2, and a forwarding agent live in 49 days, and it expects more than 3 million shipments in 2026. STG Logistics, a port to door provider, reports 95% of orders processed with no human touch and an 80% cost saving in order entry. Eassons Transport, a Canadian refrigerated carrier, went live in 40 days at 98% touchless load building. Prism Logistix runs 20 agents and reports a 10% margin improvement.

With our custom model, we deliver faster and better service to our customers. The model just knows every customer and lane preference, without having to look it up in our system every time.

Gene Welsh · Chief Transportation Officer, Mallory Alexander · Pallet Custom Models announcement · 2026-07-15

The fifth, Mallory Alexander, is the clearest test of accuracy. Its CEO, Paul Svindland, had paused an AI project with another vendor and chose Pallet for its freight forwarding knowledge and its CargoWise integration. The story reports three times more import files per operator and 100% accuracy on importer security filings and letters of credit, documents that banks and governments require. It also quotes him putting the enterprise value Pallet created at $18 million.

How to read these numbers

Every result here comes from Pallet's own case pages, with customer executives quoted. The measures differ: Lineage counts AI runs, STG counts touchless orders, Prism counts margin and ramp time for new hires, Mallory counts files per operator. Each figure describes one deployment as the customer and Pallet report it, so in an interview ask which measure Pallet tracks across all accounts.