AI Changes FTL Technology Assumptions

Satvik Agnihotri
4 min readFeb 1, 2024
Photo by Bailey Alexander on Unsplash

For anyone vaguely familiar with the trucking industry, the complexity of the industry is alarmingly obvious. Streamlined systems are sparse because edge-cases are the majority of operational complexity.

Any enterprise broker or carrier knows the difficulty of market pricing, let alone trying to identify the cost of fulfillment for their organization. Nihalists would made great freight brokers: chaos is omnipresent.

For the last 30 years, industry has flirted with AI but has been skeptical to invest. It’s difficult to justify: AI has staggering up-front costs and is often naive; It struggles to adapt to local-context. Companies have had no choice but to invest heavily in dispatchers and load-planners to manage operations. No one, let alone AI, could continuously switch between the vast expanse of different contexts. That is, until now.

AI has never seized the reigns of operations because, historically, complexity wins.

Thus, the status quo has resolved to balancing weighted-set-cover-theory models with empty miles, utilization levels, and a dispatcher’s gut on their drivers’ needs.

There is no COGS breakdown, profitability analysis, or feed-back-loop to find operational asymmetries.

This system is brutally limiting and drowned in chaos. It is impossible for a carrier operate an efficient network when it doesn’t know how individual decisions contribute to it.

It’s also important to take a step back, and note that carriers are not to blame. Running AI agents to parse data and continuously segment different KPIs by various granularities, is both complex and quickly becomes expensive.

Many carriers, especially in down-markets, sit on the verge of violating their covenants, operating at 1–2% margins.

It would be unwarranted for a board room to get together and decide to investigate the underlying technology that they outsource, instead of focusing on their strategic procurement priorities over the next quarter.

At the same time, this leaves a very significant gap in the market. Existing incumbents: Platform Sciences, Manhattan Associates, Samsara, often focus on the integration and utilization part of the technology stack.

Existing frameworks essentially splicing historical data and looking at the high-level KPIs. How’s my utilization? Empty miles? Revenue per mile? If nothing is broken, we’re in the clear.

Herein lies a false fundamental assumption. They fabric used to stitch complex operations together, loses granularity while being weaved.

What we need is proactive software: something intelligent enough to explore statistical patterns by itself. Software shouldn’t just display your data, but flags when a given shipper is consistently tardy, anticipates seasonal shifts in demand, and proactively gives your operations teams suggestions while delicately balancing all relevant KPIs.

This has only recently become possible because of two evolutions:

1. Generative AI

Traditional AI has always been task-specific. It does very well in structured environments — with clearly defined rules, and limited edge-cases.

Generative AI, on the other hand, tends to be far more dynamic in its output. While it struggles to be precise, it is better at applying an “intuition”. It explores abstract vector spaces and connect patterns that otherwise lack context.

The development of GenAI unlocks the possibility of a hybrid approach: Where the dynamic nature of GenAI can be channeled into the rigidity of traditional-AI solutions, to offer an end-end operational platform.

In this way, we unlock more abstract context for traditional-AI solutions to take into account: macro-economic factors, real-time weather conditions, shipper-specific seasonality trends, in a far more sophisticated way.

2. Availability of AI Resources

Over the last 17 years, GPU performance per dollar has been steadily increasing. As such, the barriers to run increasingly complex AI operations is decreasing.

Far-less CapEx is needed to create the same results, meaning board-rooms that previously passed on AI initiatives are finding themselves facing increasingly appealing decisions.

In the past, the amount of compute required to properly understand these systems would’ve been far to expensive and difficult to implement. Rightfully, so.

No amount of math or software could create the structure needed to model an enterprise-trucking company. That is, until now.

We are reaching an inflection point, both through technological advancements in GenAI, and available compute, such that AI can be used to model and execute decisions in far more complicated and abstract environments.

We can build models that correctly map the complexity: internalizing driver preferences, load:capacity ratios, geographic fuel-premiums relative to one-way-pricing fuel-surcharges, seasonality trends, expected maintenance, the list goes on.

For the first time, we have the tools to build effective infrastructure that captures the complexity of the businesses and work within it.

The world needs platforms that are elastic to the environment and sophisticated enough to handle the true operational complexity of real-businesses.

We’re building this for enterprise FTL carriers at Stealth Logistix.