The war on complexity starts at a portal nobody wanted to feed

The war on complexity starts at a portal nobody wanted to feed. Traditional TMS fails when it demands that third parties bring the data. The loose agent fails when it demands that the human disappear. Both were built against how the work actually happens.

In 30 seconds

One generation of transport software built expensive screens and then asked carriers and clerks to feed them by hand. The next generation promised agents that resolve everything with nobody intervening. Both share the same design sin: they do not go where the action is.

There is a third architecture. The system collects the data in the channel where the work already happens. It lands in a single shipment record. It automates the repetitive. And what it cannot resolve with certainty it escalates as a ticket, with a ready file, to a human who decides.

The portal nobody wanted to feed

I saw it up close. The UniGIS implementation at Alpura, one of the largest dairy companies in the country, depended on an assumption that looked reasonable on paper and proved lethal in operations. Every carrier provider had to enter the system and capture, by hand, every status of every trip. Loaded. In transit. Delivered. Without that voluntary third-party capture, the most expensive screen in the boardroom had nothing to show.

Consider the incentive. Hundreds of providers, each with their own operation on top, feeding their customer's system for free. Data arrived late, incomplete, or not at all. And it was not a defect of that brand in particular. It was the defect of an entire generation of software. That is the canonical example of how not to build a system. The tool did not go get the data. It demanded that third parties bring it.

Anatomy of the luxury database

The industry has an elegant name for that generation: Transportation Management System. I prefer a more honest one. Luxury databases. Extremely expensive systems that do not capture reality by themselves, require dozens of people to stay current, are not intuitive for whoever operates, and need armies of consultants to adapt to each company.

The category numbers are public and brutal. The classic Standish Group analysis of about 50,000 projects found that 66% of enterprise technology projects end in partial or total failure. A mid-market TMS implementation takes 6 to 12 months in a standard configuration. And the license is only 20 to 25% of total cost of ownership, with three out of four implementations exceeding their budget. The rest goes to integrations, consulting, and maintenance. In other words, to administering the complexity the system itself brought in.

The opposite promise: the loose agent

Against that inheritance, the software industry offered this year an opposite promise that is equally defective. The autonomous AI agent that executes everything end to end with nobody intervening. Gartner projections already put a date on that promise. More than 40% of agentic AI projects will be canceled before the end of 2027, due to rising costs, unclear value, and inadequate risk controls. Meanwhile adoption runs. Enterprise applications with embedded agents will go from under 5% in 2025 to about 40% by the end of 2026.

Both enemies share the same design sin. They were built against how the work actually happens. Traditional TMS demands that the human go to the system. The loose agent demands that the human disappear. Neither goes where the action is.

The third architecture

There is a third architecture, and it is winning quietly. It has four traits. First, the system goes for the data, not the data to the system. Capture happens where the action happens, and in Mexico the action happens on WhatsApp, where penetration reaches 93% and more than 200 million businesses already operate on the platform. The operator sends the delivery photo through the channel they already use. Current technology can collect it there, along with GPS and email, without asking anyone to change tools or capture anything twice.

Second, everything lands in a single central shipment record, with evidence attached. Third, AI absorbs the repetitive first: matches, reconciliations, statuses, classification. And fourth, what AI cannot resolve with certainty it does not invent. It escalates as a ticket, with a ready file, to a human who decides.

Why a wrong yes is expensive

The difference with the loose agent is not cosmetic. It is economic. In the autonomous model, the error travels straight to payment and is discovered weeks later, with no defense. In the ticket model, the error stops in a review queue and costs minutes. As Pickaxe's human-in-the-loop systems guide documents, the principle fits in one sentence. The cost of a wrong yes is far higher than the cost of a small delay.

What Decagon already sells

The argument does not come from theory alone. It comes from the most aggressive player in the agent market. Jesse Zhang, founder of Decagon, the platform valued at $4.5 billion, puts it without romance on the Founders in Arms podcast. A good agent should resolve between 70 and 90% of volume. The rest lives as a ticket and reaches a person. The company that serves Hertz and Notion does not sell total autonomy. It sells a system where the human is part of the product.

The boardroom lesson

I return to the Alpura boardroom because the full moral is there. That system failed from excess human friction. The autonomous agent fails from excess faith in the machine. The architecture that wins does not choose between the two extremes. It captures without friction where the action happens, automates the routine, and reserves the human exactly for the case that needs them, with the evidence already served.

The right question for any vendor is no longer how many modules they have or how smart their agent is. It is double and simple. Who feeds your system, and what happens to the case your system cannot resolve. If the answer to the first is "your providers, by hand," we have seen that movie. If the answer to the second is silence, we have seen that one too.

At OCL Cargo that architecture is operational: capture at source, voyage file, agents that match, tickets to Finance when money or certainty is at stake. It can stamp invoice and Carta Porte. Typical pilot: 6 to 8 weeks.

Sources

• Standish Group CHAOS / enterprise project failure rates.
• Gartner: +40% agentic projects canceled by 2027
• Pickaxe: human-in-the-loop systems guide.
• Founders in Arms / Decagon public profiles on valuation and coverage.

Related reading

• The math vendors do not multiply
• What Decagon understood before the market
• A 1983 irony that transport software perfected

FAQ

By Gibrán Ramírez, CEO of OCL Cargo.