The Ghost in the Freight Machine

Somewhere in the San Francisco Bay Area, on a weekend that should have belonged to the 49ers, a drayage operator sat at his desk with roughly 150 bills of lading stacked in front of him. He keyed them in by hand — one by one by one — while the playoffs played to an empty room. "Susan," he told the young computer scientist watching him, "I hate doing this." That man didn't know it, but he had just delivered the founding pitch for a company. The scientist was Sushanth Raman. The company is Pallet. And the gospel he's been preaching ever since lands like a slap: most AI in logistics is theater, and the few who skip the show are about to eat everyone's lunch.

On the latest episode of TPM Today — the Journal of Commerce's twice-monthly technology show — host Eric Johnson, broadcasting unusually from S&P Global's New York headquarters, brought Raman on to interrogate the single statistic haunting every logistics conference this fall: that 95% of enterprise AI initiatives have failed to do what they set out to do. Johnson has heard it everywhere. "That study has come up time and time again in my travels over the last two months," he said. "More people know about this than they do about any other aspect of AI right now." Raman didn't flinch. He'd built his entire company on knowing why.

The man who left Google to read bills of lading

Raman's résumé reads like a one-way ticket to a comfortable life in big tech. Computer science at Columbia. A stint at Google as a product manager on the AI team. Then Retool, where he worked shoulder-to-shoulder with the marquee names of modern logistics — Airbnb, Flexport, C.H. Robinson. It was there, surrounded by the best technology money could buy, that he noticed something absurd. "Everything from processing commercial invoices, packing lists, ISFs, arrival notices — all this stuff was still happening manually," he recalled. "Container tracking was still manual. Despite even having the best technology, this problem was ubiquitous across logistics."

So he quit. Not to build a demo, but to go look. He visited drayage and freight-forwarding operations across the Bay Area, and his grandparents — themselves industry veterans — had shown him the same thing years earlier. The 150-bills-of-lading weekend was the moment the abstraction became a face. "There's just so much more we could be doing to help you utilize your time better," Raman told the operator. To Johnson, he framed it as a calling rather than a market: "I just have to go solve the world of manual operations in global logistics."

Johnson, who has spent a career chronicling the industry's inefficiencies, recognized the convert's energy. "You're preaching to the choir here," he said. "We all find this industry endlessly fascinating, and we also know where all of the inefficiencies lie. We just haven't always had an idea of how to solve those." That, he suggested, is what makes this moment different — smart outsiders arriving to work the solution side rather than relitigate the problem.

Tribal knowledge: the thing the demos can't fake

Here is Raman's central, unfashionable claim. The reason 95% of AI agents fail has almost nothing to do with model quality and almost everything to do with ignorance. "The company that's coming to automate your business operation doesn't understand all the tribal rules that exist in your operation," he said. And those rules are everywhere, hiding in plain sight.

Consider the incoterm that's almost never written on the document. Or entity resolution: the email says "Procter and Gamble," but your system stores it as "PNG Incorporated," and no model trained on the open internet knows those are the same customer. Or freight class, which lives in someone's head and nowhere else. "There's like thousands of tribal rules that exist inside every single operation," Raman said, "and everyone watching this show knows their operators are doing things that are often not well documented."

Johnson translated it into something a non-technical reader can hold: every process and decision carries context. "It's not a yes or no," he offered. "Some of them are data. Some of them are people's gut instincts. Some of them are: I have a good relationship with them, and it's worth me paying $2 more — or I always know that this company sends this doc a day late, but they're so good at everything else that I just forgive them for this one transgression." Raman's answer to "how do you possibly capture a million unique circumstances" is methodical rather than magical: Pallet goes onsite, compares what was actually input against what landed in the system, infers the rules, then sits with operators who confirm or correct each one. The validated rules get chunked into discrete memories the system knows when to retrieve.

His analogy is disarmingly human. Onboarding an AI agent, he argued, is no different from onboarding a billing clerk. There are different billing cycles, different customers you'll chase harder than others. "If you didn't give that person any of that information, would they be successful at their job? Probably no," Raman said. "It's essentially the same thing with AI. It doesn't matter if the thing looks fancy or the widget looks cool — without context, much like your employees, it will not perform the job properly." With it, he claims, Pallet's copilot hits 97–98% accuracy.

"Every single implementation failed. Every single one."

Raman came armed with receipts. A well-known Canadian 3PL had tried four other AI vendors before Pallet. "Every single implementation failed," he said. "Every single one. They tried Pallet. It worked, and they were like: you guys are the only ones that captured our tribal knowledge." A Memphis-based freight forwarder ran the same gauntlet, then concluded the rival vendors plainly "did not understand CargoWise" or the nuances of forwarding. Pallet did, and became their exclusive automation partner.

Then came the line that will sting Sand Hill Road. Asked what else explains the carnage, Raman didn't reach for jargon. "A lot of AI agents are being built, no offense, by like Stanford dropouts who don't understand your industry really well," he said. "You don't want to teach them what's an ISF, what's an arrival notice, what's a BL, what's a packing list, how commercial invoices and packing lists are tied together, what's an incoterm." Beyond missing context, he listed the other classic failure modes: too few domain experts in the room, the wrong use cases chosen, and change management never thought through. The winners, he said, are the rare few who get all of it right at once.

Shippers, forwarders, and the broad bet

A viewer, Eduardo Ramirez, asked how AI agents help shippers specifically. Raman drew a clean line. Shippers run lean teams on big transportation budgets, so Pallet aims at spend. He described working with one of Europe's largest manufacturers to procure air freight on the spot market: Pallet reaches out to a handful of forwarders, collects quotes within a time window, runs the negotiation, and surfaces the cheapest option. Forwarders and 3PLs are a different animal — headcount-heavy operations where the metric that matters is net revenue per full-time employee. There, Pallet attacks G&A as a share of revenue.

One unnamed customer — "one of our mutual friends," Raman told Johnson — is automating commercial invoice and packing-list processing, ISFs, bills of lading, booking forms, letters of credit and track-and-trace, with a couple million dollars of EBITDA improvement expected on the bottom line. Which raised Johnson's sharpest strategic question: why go broad? Other startups pick one lane — customs entry generation, freight audit — and go deep. Raman's answer is that context, once captured, is reusable. "If you can capture the organizational context, it's really easy to spin up multiple agents in their business." And buyers, he argued, want a one-stop shop, not eight vendors to manage. He pointed to C.H. Robinson — which has gone all-in on AI across its business — as the public case study. "Look at the C.H. Robinson stock," he said. "That's an example of what happens when you do this automation successfully across your whole business." Johnson agreed it may become "the HBR case study for how this can actually impact business performance in logistics."

Organizational memory, and the AWS of work

The concept Raman keeps circling is organizational memory. A memory, he explained, is a specific rule tied to a customer or workflow. "Let's say you work with Apple, and Apple has a policy that all shipments are by default full container load, the incoterm is default Ex Works — there's a lot of specific rules around Apple. A memory is capturing one of those tribal rules." Johnson immediately saw the human implication: the shipper whose entire import operation lives in one overloaded person's head. If that person disappears for three months, the knowledge goes with them. Pallet's pitch is that the knowledge gets transferred to a digital workforce instead. "Yes," Raman confirmed. "That's how you should think about it."

Johnson floated his own framing — the "AWSification of work." Just as cloud computing lets companies scale capacity up and down, digital workers could let logistics firms surge through peak season or disruption and ease off in the lulls, without keeping everyone on 40-hour weeks year-round. Raman embraced it, then sharpened it with a real scenario: a private-equity-owned freight forwarder told to grow revenue by a fixed percentage next year without adding headcount. "That is physically impossible," they protested — the laws of physics don't allow covering more shipments with the same people. "The only answer," Raman said, "is using some of this agentic AI." Either you augment capacity, or you're in a rough spot.

How to call a bluff

With the market looking, in Johnson's words, "a little bit like the Wild West," he asked the most useful question of the episode: how should a buyer choose? Raman's advice was bracingly concrete. Ask blunt domain questions. "Do you have CargoWise experience? Could you point to what the CargoWise XML structure is like? What's an ISF? What's an arrival notice?" If the answer is hand-waving and "it's on the roadmap" — disqualify them. "They don't know what they're talking about."

Johnson added a piece of field-tested wisdom of his own: bring a technical person to the table. AI, he noted, has dragged non-technical buyers into technical decisions they never had to make when they simply bought SaaS. "Now there needs to be more understanding about what's under the hood." A technical colleague can smell whether a vendor is making sense, even without knowing the business cold.

The reassuring punchline: it's still early

For all the urgency, Raman's closing assessment was oddly calming. Are we at the transformative moment? "We're still very early," he said. "We're still in the phase of AI exploration. Very, very, very few companies have actually seen EBITDA gains from implementing AI." Johnson seized on it as the antidote to industry anxiety. "A lot of people feel very behind the curve already," he said. "You're not that far behind anybody else — and you may be ahead of people if you're even considering this." The mandate, then, is permanent research mode. You can't ignore it. But you're not late.

Then, as ritual demands, Johnson asked his closing question — favorite musician, and why. Raman went with EDM: Martin Garrix, for the dedication to craft that carried him from Europe to global fame, and John Summit, for the story. A fired accountant, freshly dumped, exiled to his parents' basement, written off by everyone — who spent the pandemic making music until the hits came. "Pain creates the best art," Johnson offered. It's a fitting coda for a founder who turned a drayage operator's miserable, 49ers-less weekend into a thesis. The ghost in the freight machine isn't a model. It's the thousands of unwritten rules in your operators' heads. Pallet's bet is that whoever learns to read them first wins.