AI Hurdles:
AI Hurdles:
Transportation Capacity
At A Crossroads
Table of Contents
Executive Summary
Every fleet leader we talk to says they want AI. Almost none of them are struggling to find a vendor who offers it. The struggle is somewhere else entirely.
This paper is built from four internal interviews - with an operations veteran who came up on clipboards and spreadsheets, a business development lead who sits across the table from fleets evaluating AI, an applied scientist who builds BeyondTrucks' optimization engine, and a data and AI product leader who has spent a decade in this space. We asked each of them the same underlying question in different ways: what actually stops a fleet from getting value out of AI? Their answers converged on the same five hurdles, almost none of which are about the AI itself.
None of these hurdles are about whether a model is smart enough. They're about whether the fleet has spent the years before AI arrived capturing the right data, in the right structure, with the right trust and organizational will behind it. The fleets that clear these hurdles get compounding value. The fleets that skip them get an expensive bolt-on that quietly stops getting used.
The five hurdles, in the order fleets tend to hit them:
The data quality myth - believing more data beats the right data.
Legacy systems that record but don't decide - the gap between a system of record and a system of action.
The trust gap - why a plausible-sounding answer is not the same as a correct one.
The change-management wall - the human resistance that has nothing to do with loyalty.
Bolting AI onto old foundations - the fastest way to spend seven figures and still not get an answer you can use.
Introduction: The Wrong Question
Most fleets evaluating AI ask, "Which vendor has the best model?" It's the wrong question, according to every person we interviewed for this paper - and it's the wrong question for a specific, recurring reason: a model is only as good as the operational reality it's been given to reason about.
"Garbage in, garbage out. Over time, the conversation has evolved into how can I record data to feed my models? How can I surface this data to make decisions?"
- Matias, Business Development
A year ago, according to Matias, the conversation with prospective customers was simple and a little naive: put AI on something. Any process, any decision, any workflow - just put AI on it, because everyone was trying to catch up. That conversation has matured. Fleets now come in asking about data capture and value levers, not features. That's progress. But it also means the hurdles have moved from "does the AI work" to something harder to see and harder to fix: "does our operation actually produce the information the AI needs to be useful."
What follows are the five hurdles our own team encounters most often when they sit across the table from a fleet - in their own words, with the reasoning behind each one.
Executive Summary
Every fleet leader we talk to says they want AI. Almost none of them are struggling to find a vendor who offers it. The struggle is somewhere else entirely.
This paper is built from four internal interviews - with an operations veteran who came up on clipboards and spreadsheets, a business development lead who sits across the table from fleets evaluating AI, an applied scientist who builds BeyondTrucks' optimization engine, and a data and AI product leader who has spent a decade in this space. We asked each of them the same underlying question in different ways: what actually stops a fleet from getting value out of AI? Their answers converged on the same five hurdles, almost none of which are about the AI itself.
None of these hurdles are about whether a model is smart enough. They're about whether the fleet has spent the years before AI arrived capturing the right data, in the right structure, with the right trust and organizational will behind it. The fleets that clear these hurdles get compounding value. The fleets that skip them get an expensive bolt-on that quietly stops getting used.
The five hurdles, in the order fleets tend to hit them:
The data quality myth - believing more data beats the right data.
Legacy systems that record but don't decide - the gap between a system of record and a system of action.
The trust gap - why a plausible-sounding answer is not the same as a correct one.
The change-management wall - the human resistance that has nothing to do with loyalty.
Bolting AI onto old foundations - the fastest way to spend seven figures and still not get an answer you can use.
Our findings indicate that successful organizations must assertively position themselves in the market and pursue markedly different strategic paths based on their market position.
Private fleet operators must defend their control & quality premium while strategically outsourcing non-core movements.
Specialized and dedicated for-hire carriers must deepen customer integration, quality management, and emphasize risk management expertise.
Common carriers must singularly be focused on pursuing cost leadership.
We estimate that carriers who successfully implement these strategies could achieve 15-20% cost advantages over competitors while building stronger customer relationships that transcend price-based competition. Above all, developing strategic clarity about which of the three paths to pursue will be existential.
The Transportation Market at a Crossroads
The $800 billion U.S. freight transportation market has traditionally operated with three distinct business models:
Private fleets (~50% of US Transportation Spend) that prioritize service quality and control despite higher costs
Specialized for-hire carriers (~20% of US Transportation Spend) that create value through unique equipment and expertise for complex freight
Common for-hire carriers that compete primarily on cost efficiency in a commoditized market segment.
These traditional boundaries are increasingly blurring as each segment faces unique challenges. Private fleets must justify their cost premium, specialized carriers confront line haul commoditization from advancing autonomous technologies, and common carriers struggle with chronic margin pressure from brokerage disruption and market fragmentation. In this evolving landscape, strategic clarity on building competitive advantage becomes crucial for maintaining sustainable fleet operations.
Executive Summary
Every fleet leader we talk to says they want AI. Almost none of them are struggling to find a vendor who offers it. The struggle is somewhere else entirely.
This paper is built from four internal interviews - with an operations veteran who came up on clipboards and spreadsheets, a business development lead who sits across the table from fleets evaluating AI, an applied scientist who builds BeyondTrucks' optimization engine, and a data and AI product leader who has spent a decade in this space. We asked each of them the same underlying question in different ways: what actually stops a fleet from getting value out of AI? Their answers converged on the same five hurdles, almost none of which are about the AI itself.
None of these hurdles are about whether a model is smart enough. They're about whether the fleet has spent the years before AI arrived capturing the right data, in the right structure, with the right trust and organizational will behind it. The fleets that clear these hurdles get compounding value. The fleets that skip them get an expensive bolt-on that quietly stops getting used.
The five hurdles, in the order fleets tend to hit them:
The data quality myth - believing more data beats the right data.
Legacy systems that record but don't decide - the gap between a system of record and a system of action.
The trust gap - why a plausible-sounding answer is not the same as a correct one.
The change-management wall - the human resistance that has nothing to do with loyalty.
Bolting AI onto old foundations - the fastest way to spend seven figures and still not get an answer you can use.
Introduction: The Wrong Question
Most fleets evaluating AI ask, "Which vendor has the best model?" It's the wrong question, according to every person we interviewed for this paper - and it's the wrong question for a specific, recurring reason: a model is only as good as the operational reality it's been given to reason about.
"Garbage in, garbage out. Over time, the conversation has evolved into how can I record data to feed my models? How can I surface this data to make decisions?"
- Matias, Business Development
A year ago, according to Matias, the conversation with prospective customers was simple and a little naive: put AI on something. Any process, any decision, any workflow - just put AI on it, because everyone was trying to catch up. That conversation has matured. Fleets now come in asking about data capture and value levers, not features. That's progress. But it also means the hurdles have moved from "does the AI work" to something harder to see and harder to fix: "does our operation actually produce the information the AI needs to be useful."
What follows are the five hurdles our own team encounters most often when they sit across the table from a fleet - in their own words, with the reasoning behind each one.
HURDLE 1
More Data Is Not the Same as Better Data
The single most common misconception our team runs into is that data quantity equals AI readiness - that a fleet with years of historical records in a data warehouse is sitting on a goldmine, ready to point a model at it and get useful answers back.
"Data is only as good as it represents your business. It's not just about the quantity of data. It's understanding the relationships between those data, as well as between the data and the business processes themselves."
- Mahriah, Data & AI Product
This shows up as two failure modes on opposite ends of the same problem. One kind of customer chases volume - more tables, more rows, more history - without ever documenting what any of it means or how it connects to a real decision. The other kind is intimidated by the scale of the problem and stalls before starting. Mahriah's advice to both is the same: start small, but start with quality. Identify the handful of relationships in the business that actually drive value, and document those first. That, on its own, is worth more than another terabyte of unlabeled history.
Why You Can't Just Dump the Data Lake Into an LLM
Two interviewees, working from different disciplines, arrived at the identical explanation for why this doesn't work. The applied science answer and the data architecture answer are the same answer.
"LLMs operate on language. Without language context or structural reference for what that data represents in the business, the model can't make sense of how those different data elements are related to one another."
- Mahriah, Data & AI Product
"A semantic layer, it's a translation, it lets the system understand that a customer and a receiver are the same thing. It's just something that lets you get around the fact that a database is basically like a spreadsheet, and the semantic layer is like giving somebody the company handbook alongside it."
- John, Applied Science
The company handbook analogy is worth sitting with. A spreadsheet full of loads, drivers, and dwell times tells a model what happened. It does not tell the model that a "receiver" and a "customer" are the same entity, that a certain exception code always means the same thing, or that a load marked late at one facility is treated differently than the same lateness at another. Without that context, the model is guessing - and, as both interviewees note independently, a model that has been trained to always produce an answer will guess confidently rather than say it doesn't know.
Mahriah is explicit that a semantic layer is necessary but not sufficient: "it cannot manufacture truth. It cannot manufacture new information. You still need to have the data to be able to answer those questions." A well-built semantic layer on top of thin data doesn't fix the data - it just makes the gaps easier to see.
Traditional Data Strategy vs. AI-Ready Dat
Mahriah draws a sharp line between how fleets have historically approached data and what AI actually needs from them:
Traditional approach: store everything, prioritize historical completeness, treat the TMS as a system of record that logs what already happened.
AI-ready approach: structure what you capture so a model can interpret it, and prioritize real-time operational data over historical archives - because the value of AI is mostly in decisions made now, not reports generated later.
That distinction - reporting versus deciding - recurs throughout this paper, because it is also the line between a system of record and a system of action, the subject of the next hurdle.
HURDLE 1
More Data Is Not the Same as Better Data
The single most common misconception our team runs into is that data quantity equals AI readiness - that a fleet with years of historical records in a data warehouse is sitting on a goldmine, ready to point a model at it and get useful answers back.
"Data is only as good as it represents your business. It's not just about the quantity of data. It's understanding the relationships between those data, as well as between the data and the business processes themselves."
- Mahriah, Data & AI Product
This shows up as two failure modes on opposite ends of the same problem. One kind of customer chases volume - more tables, more rows, more history - without ever documenting what any of it means or how it connects to a real decision. The other kind is intimidated by the scale of the problem and stalls before starting. Mahriah's advice to both is the same: start small, but start with quality. Identify the handful of relationships in the business that actually drive value, and document those first. That, on its own, is worth more than another terabyte of unlabeled history.
Why You Can't Just Dump the Data Lake Into an LLM
Two interviewees, working from different disciplines, arrived at the identical explanation for why this doesn't work. The applied science answer and the data architecture answer are the same answer.
"LLMs operate on language. Without language context or structural reference for what that data represents in the business, the model can't make sense of how those different data elements are related to one another."
- Mahriah, Data & AI Product
"A semantic layer, it's a translation, it lets the system understand that a customer and a receiver are the same thing. It's just something that lets you get around the fact that a database is basically like a spreadsheet, and the semantic layer is like giving somebody the company handbook alongside it."
- John, Applied Science
The company handbook analogy is worth sitting with. A spreadsheet full of loads, drivers, and dwell times tells a model what happened. It does not tell the model that a "receiver" and a "customer" are the same entity, that a certain exception code always means the same thing, or that a load marked late at one facility is treated differently than the same lateness at another. Without that context, the model is guessing - and, as both interviewees note independently, a model that has been trained to always produce an answer will guess confidently rather than say it doesn't know.
Mahriah is explicit that a semantic layer is necessary but not sufficient: "it cannot manufacture truth. It cannot manufacture new information. You still need to have the data to be able to answer those questions." A well-built semantic layer on top of thin data doesn't fix the data - it just makes the gaps easier to see.
Traditional Data Strategy vs. AI-Ready Dat
Mahriah draws a sharp line between how fleets have historically approached data and what AI actually needs from them:
Traditional approach: store everything, prioritize historical completeness, treat the TMS as a system of record that logs what already happened.
AI-ready approach: structure what you capture so a model can interpret it, and prioritize real-time operational data over historical archives - because the value of AI is mostly in decisions made now, not reports generated later.
That distinction - reporting versus deciding - recurs throughout this paper, because it is also the line between a system of record and a system of action, the subject of the next hurdle.
Our findings indicate that successful organizations must assertively position themselves in the market and pursue markedly different strategic paths based on their market position.
Private fleet operators must defend their control & quality premium while strategically outsourcing non-core movements.
Specialized and dedicated for-hire carriers must deepen customer integration, quality management, and emphasize risk management expertise.
Common carriers must singularly be focused on pursuing cost leadership.
We estimate that carriers who successfully implement these strategies could achieve 15-20% cost advantages over competitors while building stronger customer relationships that transcend price-based competition. Above all, developing strategic clarity about which of the three paths to pursue will be existential.
The Transportation Market at a Crossroads
The $800 billion U.S. freight transportation market has traditionally operated with three distinct business models:
Private fleets (~50% of US Transportation Spend) that prioritize service quality and control despite higher costs
Specialized for-hire carriers (~20% of US Transportation Spend) that create value through unique equipment and expertise for complex freight
Common for-hire carriers that compete primarily on cost efficiency in a commoditized market segment.
These traditional boundaries are increasingly blurring as each segment faces unique challenges. Private fleets must justify their cost premium, specialized carriers confront line haul commoditization from advancing autonomous technologies, and common carriers struggle with chronic margin pressure from brokerage disruption and market fragmentation. In this evolving landscape, strategic clarity on building competitive advantage becomes crucial for maintaining sustainable fleet operations.
HURDLE 1
More Data Is Not the Same as Better Data
The single most common misconception our team runs into is that data quantity equals AI readiness - that a fleet with years of historical records in a data warehouse is sitting on a goldmine, ready to point a model at it and get useful answers back.
"Data is only as good as it represents your business. It's not just about the quantity of data. It's understanding the relationships between those data, as well as between the data and the business processes themselves."
- Mahriah, Data & AI Product
This shows up as two failure modes on opposite ends of the same problem. One kind of customer chases volume - more tables, more rows, more history - without ever documenting what any of it means or how it connects to a real decision. The other kind is intimidated by the scale of the problem and stalls before starting. Mahriah's advice to both is the same: start small, but start with quality. Identify the handful of relationships in the business that actually drive value, and document those first. That, on its own, is worth more than another terabyte of unlabeled history.
Why You Can't Just Dump the Data Lake Into an LLM
Two interviewees, working from different disciplines, arrived at the identical explanation for why this doesn't work. The applied science answer and the data architecture answer are the same answer.
"LLMs operate on language. Without language context or structural reference for what that data represents in the business, the model can't make sense of how those different data elements are related to one another."
- Mahriah, Data & AI Product
"A semantic layer, it's a translation, it lets the system understand that a customer and a receiver are the same thing. It's just something that lets you get around the fact that a database is basically like a spreadsheet, and the semantic layer is like giving somebody the company handbook alongside it."
- John, Applied Science
The company handbook analogy is worth sitting with. A spreadsheet full of loads, drivers, and dwell times tells a model what happened. It does not tell the model that a "receiver" and a "customer" are the same entity, that a certain exception code always means the same thing, or that a load marked late at one facility is treated differently than the same lateness at another. Without that context, the model is guessing - and, as both interviewees note independently, a model that has been trained to always produce an answer will guess confidently rather than say it doesn't know.
Mahriah is explicit that a semantic layer is necessary but not sufficient: "it cannot manufacture truth. It cannot manufacture new information. You still need to have the data to be able to answer those questions." A well-built semantic layer on top of thin data doesn't fix the data - it just makes the gaps easier to see.
Traditional Data Strategy vs. AI-Ready Dat
Mahriah draws a sharp line between how fleets have historically approached data and what AI actually needs from them:
Traditional approach: store everything, prioritize historical completeness, treat the TMS as a system of record that logs what already happened.
AI-ready approach: structure what you capture so a model can interpret it, and prioritize real-time operational data over historical archives - because the value of AI is mostly in decisions made now, not reports generated later.
That distinction - reporting versus deciding - recurs throughout this paper, because it is also the line between a system of record and a system of action, the subject of the next hurdle.
HURDLE 2
Systems That Record the Past Don't Support the Present
Every interviewee, unprompted, drew some version of the same distinction: a system built to log what happened is architecturally different from a system built to help decide what happens next. Legacy TMS platforms, almost universally, were built for the former.
"I think the legacy systems have always been about record keeping. I look back at my last careers and my teams of data analysts working on some of these legacy softwares - I wish I had known how hard their job really was, how much they were having to work around the product just to get answers."
- Natasha, Operations
Natasha's framing matters because she isn't speaking as a technologist evaluating architecture from the outside - she ran operations teams that lived inside these systems for years. Her point is that legacy TMS platforms are good at keeping data and bad at using it. Getting an answer out of them required an analyst to work around the product, not with it.
John, who builds optimization infrastructure, makes the same point from the engineering side, and resists an easy assumption in doing so: that the deciding factor is a system's age.
"It's not so much age, but architecture - a 20-year-old system that logs every status change and every exception with a timestamp is probably in better shape than a more modern system that only stores end states in rows. The dividing line is whether the system captured what happened or only what resulted."
- John, Applied Science
Asked to put a number on it anyway, John estimates roughly ten years before a legacy TMS's data foundation becomes genuinely incompatible with where AI is heading - while adding that many legacy providers are already visibly behind, and that smaller, newer competitors without decades of accumulated technical debt can move faster.
Systems of Record vs. Systems of Action
Mahriah frames the same gap in terms of what the data is being asked to do. A system of record only has to explain how data relates to backward-looking reporting. A system of action - one that's driving a live decision - needs that same understanding of the data available the moment the data arrives, because the decision can't wait for a human to translate a report into a recommendation.
"Once you move to a system of action, which is driving business decisions, by definition those need to be real time. You need to have an understanding of that data as it comes in, because you're driving direct action versus having somebody translate a system of record into somebody's interpretation of what we should do going forward."
- Mahriah, Data & AI Product
This is precisely the gap Natasha describes from the operator's chair, and precisely the gap John is describing from the optimization engine's chair. Legacy TMS platforms were built to answer "what happened," and fleets are now asking them to answer "what should happen next" - a question their architecture was never designed to support.
HURDLE 2
Systems That Record the Past Don't Support the Present
Every interviewee, unprompted, drew some version of the same distinction: a system built to log what happened is architecturally different from a system built to help decide what happens next. Legacy TMS platforms, almost universally, were built for the former.
"I think the legacy systems have always been about record keeping. I look back at my last careers and my teams of data analysts working on some of these legacy softwares - I wish I had known how hard their job really was, how much they were having to work around the product just to get answers."
- Natasha, Operations
Natasha's framing matters because she isn't speaking as a technologist evaluating architecture from the outside - she ran operations teams that lived inside these systems for years. Her point is that legacy TMS platforms are good at keeping data and bad at using it. Getting an answer out of them required an analyst to work around the product, not with it.
John, who builds optimization infrastructure, makes the same point from the engineering side, and resists an easy assumption in doing so: that the deciding factor is a system's age.
"It's not so much age, but architecture - a 20-year-old system that logs every status change and every exception with a timestamp is probably in better shape than a more modern system that only stores end states in rows. The dividing line is whether the system captured what happened or only what resulted."
- John, Applied Science
Asked to put a number on it anyway, John estimates roughly ten years before a legacy TMS's data foundation becomes genuinely incompatible with where AI is heading - while adding that many legacy providers are already visibly behind, and that smaller, newer competitors without decades of accumulated technical debt can move faster.
Systems of Record vs. Systems of Action
Mahriah frames the same gap in terms of what the data is being asked to do. A system of record only has to explain how data relates to backward-looking reporting. A system of action - one that's driving a live decision - needs that same understanding of the data available the moment the data arrives, because the decision can't wait for a human to translate a report into a recommendation.
"Once you move to a system of action, which is driving business decisions, by definition those need to be real time. You need to have an understanding of that data as it comes in, because you're driving direct action versus having somebody translate a system of record into somebody's interpretation of what we should do going forward."
- Mahriah, Data & AI Product
This is precisely the gap Natasha describes from the operator's chair, and precisely the gap John is describing from the optimization engine's chair. Legacy TMS platforms were built to answer "what happened," and fleets are now asking them to answer "what should happen next" - a question their architecture was never designed to support.
Our findings indicate that successful organizations must assertively position themselves in the market and pursue markedly different strategic paths based on their market position.
Private fleet operators must defend their control & quality premium while strategically outsourcing non-core movements.
Specialized and dedicated for-hire carriers must deepen customer integration, quality management, and emphasize risk management expertise.
Common carriers must singularly be focused on pursuing cost leadership.
We estimate that carriers who successfully implement these strategies could achieve 15-20% cost advantages over competitors while building stronger customer relationships that transcend price-based competition. Above all, developing strategic clarity about which of the three paths to pursue will be existential.
The Transportation Market at a Crossroads
The $800 billion U.S. freight transportation market has traditionally operated with three distinct business models:
Private fleets (~50% of US Transportation Spend) that prioritize service quality and control despite higher costs
Specialized for-hire carriers (~20% of US Transportation Spend) that create value through unique equipment and expertise for complex freight
Common for-hire carriers that compete primarily on cost efficiency in a commoditized market segment.
These traditional boundaries are increasingly blurring as each segment faces unique challenges. Private fleets must justify their cost premium, specialized carriers confront line haul commoditization from advancing autonomous technologies, and common carriers struggle with chronic margin pressure from brokerage disruption and market fragmentation. In this evolving landscape, strategic clarity on building competitive advantage becomes crucial for maintaining sustainable fleet operations.
HURDLE 2
Systems That Record the Past Don't Support the Present
Every interviewee, unprompted, drew some version of the same distinction: a system built to log what happened is architecturally different from a system built to help decide what happens next. Legacy TMS platforms, almost universally, were built for the former.
"I think the legacy systems have always been about record keeping. I look back at my last careers and my teams of data analysts working on some of these legacy softwares - I wish I had known how hard their job really was, how much they were having to work around the product just to get answers."
- Natasha, Operations
Natasha's framing matters because she isn't speaking as a technologist evaluating architecture from the outside - she ran operations teams that lived inside these systems for years. Her point is that legacy TMS platforms are good at keeping data and bad at using it. Getting an answer out of them required an analyst to work around the product, not with it.
John, who builds optimization infrastructure, makes the same point from the engineering side, and resists an easy assumption in doing so: that the deciding factor is a system's age.
"It's not so much age, but architecture - a 20-year-old system that logs every status change and every exception with a timestamp is probably in better shape than a more modern system that only stores end states in rows. The dividing line is whether the system captured what happened or only what resulted."
- John, Applied Science
Asked to put a number on it anyway, John estimates roughly ten years before a legacy TMS's data foundation becomes genuinely incompatible with where AI is heading - while adding that many legacy providers are already visibly behind, and that smaller, newer competitors without decades of accumulated technical debt can move faster.
Systems of Record vs. Systems of Action
Mahriah frames the same gap in terms of what the data is being asked to do. A system of record only has to explain how data relates to backward-looking reporting. A system of action - one that's driving a live decision - needs that same understanding of the data available the moment the data arrives, because the decision can't wait for a human to translate a report into a recommendation.
"Once you move to a system of action, which is driving business decisions, by definition those need to be real time. You need to have an understanding of that data as it comes in, because you're driving direct action versus having somebody translate a system of record into somebody's interpretation of what we should do going forward."
- Mahriah, Data & AI Product
This is precisely the gap Natasha describes from the operator's chair, and precisely the gap John is describing from the optimization engine's chair. Legacy TMS platforms were built to answer "what happened," and fleets are now asking them to answer "what should happen next" - a question their architecture was never designed to support.
HURDLE 3
Confident and Correct Are Not the Same Thing
The most viscerally understood hurdle among everyone interviewed is hallucination - not as an abstract AI risk, but as the specific moment a fleet stops trusting the system.
"LLMs will make things up because they've been trained to give you an answer. What I've trained to do is to give you an answer that sounds plausible. Quickly you'll see it's inventing cars that don't exist, or ignoring a customer completely, or computing dwell times using information that's not consistent with what we actually have."
- John, Applied Science
John offers a concrete illustration from outside the trucking world that lands the point precisely: he asked an AI search tool whether a well-known Los Angeles sushi restaurant had a San Francisco location. It confidently answered yes, with a street address and hours - none of which existed. The restaurant chain, in the model's training data, looked like the kind of business that would have a San Francisco location, so it generated one. Natasha's version of the same failure mode is a self-driving car that dropped her in a back alley because the routing model didn't know something a local driver would have known.
"AI agent can route a truck into certain areas, but it may not know that there happens to be construction today, or that there's a school zone during a certain time. A dispatcher may know something better, or even a driver may know better."
- Natasha, Customer Excellence
Optimization Is a Guarantee. AI Is a Category Label.
John draws a distinction that he argues fleets should insist on, and that most vendors blur on purpose: AI is a broad family of tools, useful for understanding what a person wants and generating a plausible-sounding response. Optimization, by contrast, is a deterministic solver that can certify a plan against hard constraints - hours of service, weight limits, appointment windows - before anyone acts on it.
"A language model can say, this route looks reasonable, don't you think? But the solver can actually certify that yes, this route will respect your hours of service, your weight limits, your appointment windows. You have to have that ability to trace every decision to some deterministic optimization process."
- John, Applied Science
The distinction matters to a customer, John notes, "the first time a plan violates a constraint and someone has to call a shipper and explain it." Until that moment, most buyers treat AI and optimization as interchangeable marketing language. Afterward, they never do again.
Trust Has Two Layers
Mahriah separates trust into two distinct problems that require different solutions: whether the output itself can be believed, and whether using AI introduces new security or access risk. The first, she notes, predates the current wave of AI entirely - she saw the same resistance from expert planners a decade ago with early machine learning forecasts.
"If somebody gives you a demand forecast of 10 units, a demand planner is going to be like, no, I know it's going to be 15, I don't believe it. How do you give the information such that the user who is an expert in that space can take it and trust it and move forward with it?"
- Mahriah, Data & AI Product
Her observation is that modern AI is, ironically, part of the fix for a problem it also causes: today's models can explain their own reasoning in plain language in a way older machine-learning systems couldn't, which gives an expert dispatcher something to argue with, agree with, or override - rather than a black-box number to blindly accept or reject.
HURDLE 3
Confident and Correct Are Not the Same Thing
The most viscerally understood hurdle among everyone interviewed is hallucination - not as an abstract AI risk, but as the specific moment a fleet stops trusting the system.
"LLMs will make things up because they've been trained to give you an answer. What I've trained to do is to give you an answer that sounds plausible. Quickly you'll see it's inventing cars that don't exist, or ignoring a customer completely, or computing dwell times using information that's not consistent with what we actually have."
- John, Applied Science
John offers a concrete illustration from outside the trucking world that lands the point precisely: he asked an AI search tool whether a well-known Los Angeles sushi restaurant had a San Francisco location. It confidently answered yes, with a street address and hours - none of which existed. The restaurant chain, in the model's training data, looked like the kind of business that would have a San Francisco location, so it generated one. Natasha's version of the same failure mode is a self-driving car that dropped her in a back alley because the routing model didn't know something a local driver would have known.
"AI agent can route a truck into certain areas, but it may not know that there happens to be construction today, or that there's a school zone during a certain time. A dispatcher may know something better, or even a driver may know better."
- Natasha, Customer Excellence
Optimization Is a Guarantee. AI Is a Category Label.
John draws a distinction that he argues fleets should insist on, and that most vendors blur on purpose: AI is a broad family of tools, useful for understanding what a person wants and generating a plausible-sounding response. Optimization, by contrast, is a deterministic solver that can certify a plan against hard constraints - hours of service, weight limits, appointment windows - before anyone acts on it.
"A language model can say, this route looks reasonable, don't you think? But the solver can actually certify that yes, this route will respect your hours of service, your weight limits, your appointment windows. You have to have that ability to trace every decision to some deterministic optimization process."
- John, Applied Science
The distinction matters to a customer, John notes, "the first time a plan violates a constraint and someone has to call a shipper and explain it." Until that moment, most buyers treat AI and optimization as interchangeable marketing language. Afterward, they never do again.
Trust Has Two Layers
Mahriah separates trust into two distinct problems that require different solutions: whether the output itself can be believed, and whether using AI introduces new security or access risk. The first, she notes, predates the current wave of AI entirely - she saw the same resistance from expert planners a decade ago with early machine learning forecasts.
"If somebody gives you a demand forecast of 10 units, a demand planner is going to be like, no, I know it's going to be 15, I don't believe it. How do you give the information such that the user who is an expert in that space can take it and trust it and move forward with it?"
- Mahriah, Data & AI Product
Her observation is that modern AI is, ironically, part of the fix for a problem it also causes: today's models can explain their own reasoning in plain language in a way older machine-learning systems couldn't, which gives an expert dispatcher something to argue with, agree with, or override - rather than a black-box number to blindly accept or reject.
Our findings indicate that successful organizations must assertively position themselves in the market and pursue markedly different strategic paths based on their market position.
Private fleet operators must defend their control & quality premium while strategically outsourcing non-core movements.
Specialized and dedicated for-hire carriers must deepen customer integration, quality management, and emphasize risk management expertise.
Common carriers must singularly be focused on pursuing cost leadership.
We estimate that carriers who successfully implement these strategies could achieve 15-20% cost advantages over competitors while building stronger customer relationships that transcend price-based competition. Above all, developing strategic clarity about which of the three paths to pursue will be existential.
The Transportation Market at a Crossroads
The $800 billion U.S. freight transportation market has traditionally operated with three distinct business models:
Private fleets (~50% of US Transportation Spend) that prioritize service quality and control despite higher costs
Specialized for-hire carriers (~20% of US Transportation Spend) that create value through unique equipment and expertise for complex freight
Common for-hire carriers that compete primarily on cost efficiency in a commoditized market segment.
These traditional boundaries are increasingly blurring as each segment faces unique challenges. Private fleets must justify their cost premium, specialized carriers confront line haul commoditization from advancing autonomous technologies, and common carriers struggle with chronic margin pressure from brokerage disruption and market fragmentation. In this evolving landscape, strategic clarity on building competitive advantage becomes crucial for maintaining sustainable fleet operations.
HURDLE 3
Confident and Correct Are Not the Same Thing
The most viscerally understood hurdle among everyone interviewed is hallucination - not as an abstract AI risk, but as the specific moment a fleet stops trusting the system.
"LLMs will make things up because they've been trained to give you an answer. What I've trained to do is to give you an answer that sounds plausible. Quickly you'll see it's inventing cars that don't exist, or ignoring a customer completely, or computing dwell times using information that's not consistent with what we actually have."
- John, Applied Science
John offers a concrete illustration from outside the trucking world that lands the point precisely: he asked an AI search tool whether a well-known Los Angeles sushi restaurant had a San Francisco location. It confidently answered yes, with a street address and hours - none of which existed. The restaurant chain, in the model's training data, looked like the kind of business that would have a San Francisco location, so it generated one. Natasha's version of the same failure mode is a self-driving car that dropped her in a back alley because the routing model didn't know something a local driver would have known.
"AI agent can route a truck into certain areas, but it may not know that there happens to be construction today, or that there's a school zone during a certain time. A dispatcher may know something better, or even a driver may know better."
- Natasha, Customer Excellence
Optimization Is a Guarantee. AI Is a Category Label.
John draws a distinction that he argues fleets should insist on, and that most vendors blur on purpose: AI is a broad family of tools, useful for understanding what a person wants and generating a plausible-sounding response. Optimization, by contrast, is a deterministic solver that can certify a plan against hard constraints - hours of service, weight limits, appointment windows - before anyone acts on it.
"A language model can say, this route looks reasonable, don't you think? But the solver can actually certify that yes, this route will respect your hours of service, your weight limits, your appointment windows. You have to have that ability to trace every decision to some deterministic optimization process."
- John, Applied Science
The distinction matters to a customer, John notes, "the first time a plan violates a constraint and someone has to call a shipper and explain it." Until that moment, most buyers treat AI and optimization as interchangeable marketing language. Afterward, they never do again.
Trust Has Two Layers
Mahriah separates trust into two distinct problems that require different solutions: whether the output itself can be believed, and whether using AI introduces new security or access risk. The first, she notes, predates the current wave of AI entirely - she saw the same resistance from expert planners a decade ago with early machine learning forecasts.
"If somebody gives you a demand forecast of 10 units, a demand planner is going to be like, no, I know it's going to be 15, I don't believe it. How do you give the information such that the user who is an expert in that space can take it and trust it and move forward with it?"
- Mahriah, Data & AI Product
Her observation is that modern AI is, ironically, part of the fix for a problem it also causes: today's models can explain their own reasoning in plain language in a way older machine-learning systems couldn't, which gives an expert dispatcher something to argue with, agree with, or override - rather than a black-box number to blindly accept or reject.
HURDLE 4
It's Not Loyalty. It's Fear - and It's Rational.
Ask a fleet leader why they haven't switched off a twenty-year-old TMS, and "loyalty to the vendor" is rarely the honest answer, according to the people who have those conversations for a living.
"There's definitely nothing that has to do with loyalty. These legacy TMSs - there is no loyalty for them or from the company. I would think a lot of it is more risk. Changing a large system is almost like open heart surgery. You have to take it in steps and be prepared to do it."
- Natasha, Customer Excellence
Natasha's point is that switching systems touches every role that interacts with the product - the data-entry clerk, the dispatcher, the driver - which makes it feel like an operational risk even when the underlying software relationship carries no real loyalty at all. That fear is a rational response to a real cost, not an irrational attachment to an old vendor.
Matias sees the same dynamic from the sales seat, and adds a structural reason it persists: legacy systems often come with an in-house team whose job is specifically to keep that system running.
"Systems that have been in place for 20 years, that have a whole in-house team supporting it - there's a lot of inertia. There's often people whose job is to support that system, so there's not a lot of interest in changing and doing things better."
- Matias, Business Development
The Gap Between What Leadership Believes and What's Happening on the Floor
Perhaps the most concrete story in this entire research effort comes from Matias, describing what BeyondTrucks routinely finds when it visits a prospective customer's operation in person rather than taking leadership's word for how things run.
"Leadership believes everything is solved, our processes are working well, if it's not broken don't fix it. But you see offices full of post-it notes, critical data stored on a paper sheet lost somewhere in a corner, boards, old IT systems that take up to one hour to refresh to show you where your trucks are. Half of my dispatchers just waste two or three hours on this thing every day, and they don't know it."
- Matias, Business Development
This is not a data problem or a model problem. It's a visibility problem at the leadership level, and it means the AI conversation often has to start with a walk around the floor before it can start with a conversation about architecture.
Natasha's advice to dispatchers and operations staff who fear AI is aimed at exactly this anxiety - the belief that institutional knowledge is a job-security asset that AI threatens, rather than an asset AI can help scale:
"That knowledge that lives in your head - if portions of that live into software that can help you be way more efficient, now maybe you're not working twelve hours a day. Think about how much easier and more efficient your day can be, and what you can really change that's impactful to a driver or a customer, versus just putting out fires all day long."
- Natasha, Customer Excellence
HURDLE 4
It's Not Loyalty. It's Fear - and It's Rational.
Ask a fleet leader why they haven't switched off a twenty-year-old TMS, and "loyalty to the vendor" is rarely the honest answer, according to the people who have those conversations for a living.
"There's definitely nothing that has to do with loyalty. These legacy TMSs - there is no loyalty for them or from the company. I would think a lot of it is more risk. Changing a large system is almost like open heart surgery. You have to take it in steps and be prepared to do it."
- Natasha, Customer Excellence
Natasha's point is that switching systems touches every role that interacts with the product - the data-entry clerk, the dispatcher, the driver - which makes it feel like an operational risk even when the underlying software relationship carries no real loyalty at all. That fear is a rational response to a real cost, not an irrational attachment to an old vendor.
Matias sees the same dynamic from the sales seat, and adds a structural reason it persists: legacy systems often come with an in-house team whose job is specifically to keep that system running.
"Systems that have been in place for 20 years, that have a whole in-house team supporting it - there's a lot of inertia. There's often people whose job is to support that system, so there's not a lot of interest in changing and doing things better."
- Matias, Business Development
The Gap Between What Leadership Believes and What's Happening on the Floor
Perhaps the most concrete story in this entire research effort comes from Matias, describing what BeyondTrucks routinely finds when it visits a prospective customer's operation in person rather than taking leadership's word for how things run.
"Leadership believes everything is solved, our processes are working well, if it's not broken don't fix it. But you see offices full of post-it notes, critical data stored on a paper sheet lost somewhere in a corner, boards, old IT systems that take up to one hour to refresh to show you where your trucks are. Half of my dispatchers just waste two or three hours on this thing every day, and they don't know it."
- Matias, Business Development
This is not a data problem or a model problem. It's a visibility problem at the leadership level, and it means the AI conversation often has to start with a walk around the floor before it can start with a conversation about architecture.
Natasha's advice to dispatchers and operations staff who fear AI is aimed at exactly this anxiety - the belief that institutional knowledge is a job-security asset that AI threatens, rather than an asset AI can help scale:
"That knowledge that lives in your head - if portions of that live into software that can help you be way more efficient, now maybe you're not working twelve hours a day. Think about how much easier and more efficient your day can be, and what you can really change that's impactful to a driver or a customer, versus just putting out fires all day long."
- Natasha, Customer Excellence
Our findings indicate that successful organizations must assertively position themselves in the market and pursue markedly different strategic paths based on their market position.
Private fleet operators must defend their control & quality premium while strategically outsourcing non-core movements.
Specialized and dedicated for-hire carriers must deepen customer integration, quality management, and emphasize risk management expertise.
Common carriers must singularly be focused on pursuing cost leadership.
We estimate that carriers who successfully implement these strategies could achieve 15-20% cost advantages over competitors while building stronger customer relationships that transcend price-based competition. Above all, developing strategic clarity about which of the three paths to pursue will be existential.
The Transportation Market at a Crossroads
The $800 billion U.S. freight transportation market has traditionally operated with three distinct business models:
Private fleets (~50% of US Transportation Spend) that prioritize service quality and control despite higher costs
Specialized for-hire carriers (~20% of US Transportation Spend) that create value through unique equipment and expertise for complex freight
Common for-hire carriers that compete primarily on cost efficiency in a commoditized market segment.
These traditional boundaries are increasingly blurring as each segment faces unique challenges. Private fleets must justify their cost premium, specialized carriers confront line haul commoditization from advancing autonomous technologies, and common carriers struggle with chronic margin pressure from brokerage disruption and market fragmentation. In this evolving landscape, strategic clarity on building competitive advantage becomes crucial for maintaining sustainable fleet operations.
HURDLE 4
It's Not Loyalty. It's Fear - and It's Rational.
Ask a fleet leader why they haven't switched off a twenty-year-old TMS, and "loyalty to the vendor" is rarely the honest answer, according to the people who have those conversations for a living.
"There's definitely nothing that has to do with loyalty. These legacy TMSs - there is no loyalty for them or from the company. I would think a lot of it is more risk. Changing a large system is almost like open heart surgery. You have to take it in steps and be prepared to do it."
- Natasha, Customer Excellence
Natasha's point is that switching systems touches every role that interacts with the product - the data-entry clerk, the dispatcher, the driver - which makes it feel like an operational risk even when the underlying software relationship carries no real loyalty at all. That fear is a rational response to a real cost, not an irrational attachment to an old vendor.
Matias sees the same dynamic from the sales seat, and adds a structural reason it persists: legacy systems often come with an in-house team whose job is specifically to keep that system running.
"Systems that have been in place for 20 years, that have a whole in-house team supporting it - there's a lot of inertia. There's often people whose job is to support that system, so there's not a lot of interest in changing and doing things better."
- Matias, Business Development
The Gap Between What Leadership Believes and What's Happening on the Floor
Perhaps the most concrete story in this entire research effort comes from Matias, describing what BeyondTrucks routinely finds when it visits a prospective customer's operation in person rather than taking leadership's word for how things run.
"Leadership believes everything is solved, our processes are working well, if it's not broken don't fix it. But you see offices full of post-it notes, critical data stored on a paper sheet lost somewhere in a corner, boards, old IT systems that take up to one hour to refresh to show you where your trucks are. Half of my dispatchers just waste two or three hours on this thing every day, and they don't know it."
- Matias, Business Development
This is not a data problem or a model problem. It's a visibility problem at the leadership level, and it means the AI conversation often has to start with a walk around the floor before it can start with a conversation about architecture.
Natasha's advice to dispatchers and operations staff who fear AI is aimed at exactly this anxiety - the belief that institutional knowledge is a job-security asset that AI threatens, rather than an asset AI can help scale:
"That knowledge that lives in your head - if portions of that live into software that can help you be way more efficient, now maybe you're not working twelve hours a day. Think about how much easier and more efficient your day can be, and what you can really change that's impactful to a driver or a customer, versus just putting out fires all day long."
- Natasha, Customer Excellence
HURDLE 5
Buying AI Before You've Bought Architecture
The fastest way to spend real money and get nothing usable back, according to Mahriah, is to bolt AI agents onto a legacy TMS [https://www.beyondtrucks.com/ai-ready-vs-ai-native-tms] without rebuilding the data foundation underneath it. She describes a pattern she has seen repeat, almost identically, across customer conversations.
"If a CIO came to me having just spent seven figures bolting AI agents into their existing system, I would say you are leaving a lot of money on the table. Ask them how they're measuring their ROI. A lot of times, without significant rework, you will not be getting the answers you expect, and there's going to be a lot more exceptions that a human has to come and fix."
- Mahriah, Data & AI Product
The project she describes starts small and grows quietly out of control: what was pitched as a month-long AI agent implementation stretches to two months, then three, then six - and even once it's technically live, the fleet doesn't realize the value it expected, because people are still making decisions outside the system. The system doesn't understand the business well enough to be trusted with the decision.
How Much Patching Is Too Much?
Mahriah draws a clear line between patching that works and patching that doesn't: a legacy system can absolutely be patched to solve one confined, specific use case. What it cannot do is support a genuine step change in how the business runs.
"You can patch for confined use cases. If you're trying to unlock the potential of AI in your business, solving for one specific use case is not going to do that - that's where patching becomes untenable. I have not seen people drive real P&L change without completely building a separate entire data landscape to support it."
- Mahriah, Data & AI Product
Buy Architecture, Not a Model
Asked directly what a mid-market or enterprise fleet should actually be shopping for, Mahriah's answer reframes the entire buying decision away from which AI is best:
"I would suggest you don't buy AI. You should focus on buying architecture. AI is going to change every six months - probably every month at this point - so you want platforms that understand your business data. Models are becoming commoditized. Your proprietary operational data is your real strategic asset as a company."
- Mahriah, Data & AI Product
John's caution to the industry as a whole is the mirror image of the same warning, aimed at vendors rather than buyers: AI "demos beautifully" and is easy to fund, while the unglamorous work of event capture - instrumenting exactly when a truck arrived, when a load got reassigned, why a dispatcher overrode a plan - doesn't demo at all and is easy to underfund.
"My worry is not so much that the industry spends a lot of money on AI, but that money gets spent building sophisticated systems on top of data that can't support them - and then everyone throws the baby out with the bathwater and says the AI didn't work. No - you had to give it a foundation, and that's where things really went wrong."
- John, Applied Science
Clearing the Hurdles: What Actually Works
Across all four interviews, the same handful of practices came up as the difference between fleets who get real value from AI and fleets who don't. None of them require replacing every system overnight.
1. Capture every override, with a reason
"Every override that occurs, capture a reason for it. Anytime a plan gets changed, that is the single most valuable data point the operation generates - it's a human expert saying the system's model of the business is wrong in this specific way. If you're not recording those overrides, it never learns, and dispatchers are stuck fighting against their software forever."
- John, Applied Science
John calls this out as the one wish-list item he'd give any fleet on the data front, and describes it as the fastest practical way to capture tribal knowledge - the countless small rules a dispatcher knows but would never think to recite unprompted - without a multi-month requirements process.
2. Prioritize operational, real-time data over historical archives
Mahriah's advice is to stop optimizing for the biggest possible data warehouse and start optimizing for the data that's actually flowing through decisions right now - because that's the data a system of action needs, and it's usually thinner and more specific than fleets expect.
3. Keep a human in the loop, especially on judgment calls
Every operational interviewee independently landed on the same position: AI should narrow the gap between what a dispatcher wants and what the software can do, not replace the dispatcher's judgment about construction, school zones, or a shipper relationship the system has no way to know about.
4. Insist on optimization guarantees, not just AI-generated suggestions
For anything with a hard constraint - hours of service, weight limits, appointment windows - a deterministic solver should be certifying the plan, not a language model guessing at plausibility.
5. Build or buy for architecture, not for this quarter's model
Because models are commoditizing quickly, the durable asset is the fleet's own well-structured operational data and a platform architected to use it - not whichever model happens to be state of the art this month.
6. Walk the floor before trusting the leadership narrative
The single biggest unlock in several of BeyondTrucks' customer engagements has come from simply observing operations directly - post-it notes, paper trip sheets, and manual workarounds that leadership didn't know existed - rather than accepting "our processes are working fine" at face value.
Closing Thought
None of the people interviewed for this paper described AI adoption as primarily a technology problem. Natasha described it as a fear-and-readiness problem. Matias described it as a discovery-and-trust problem. John described it as a data-architecture problem wearing an AI costume. Mahriah described it as an architecture-first buying decision that most companies get backwards.
The common thread is this: the fleets that get real value from AI are not the ones with the newest model. They're the ones that did the less exciting work first - capturing what actually happened in the field, structuring it so a system can use it, and building enough trust in the output that a dispatcher is willing to act on it. Everything else follows from that foundation.
This paper draws on internal interviews conducted as part of BeyondTrucks' ongoing "AI Hurdles" research series. Quotes have been lightly edited for clarity and length while preserving the speakers' meaning.
HURDLE 5
Buying AI Before You've Bought Architecture
The fastest way to spend real money and get nothing usable back, according to Mahriah, is to bolt AI agents onto a legacy TMS [https://www.beyondtrucks.com/ai-ready-vs-ai-native-tms] without rebuilding the data foundation underneath it. She describes a pattern she has seen repeat, almost identically, across customer conversations.
"If a CIO came to me having just spent seven figures bolting AI agents into their existing system, I would say you are leaving a lot of money on the table. Ask them how they're measuring their ROI. A lot of times, without significant rework, you will not be getting the answers you expect, and there's going to be a lot more exceptions that a human has to come and fix."
- Mahriah, Data & AI Product
The project she describes starts small and grows quietly out of control: what was pitched as a month-long AI agent implementation stretches to two months, then three, then six - and even once it's technically live, the fleet doesn't realize the value it expected, because people are still making decisions outside the system. The system doesn't understand the business well enough to be trusted with the decision.
How Much Patching Is Too Much?
Mahriah draws a clear line between patching that works and patching that doesn't: a legacy system can absolutely be patched to solve one confined, specific use case. What it cannot do is support a genuine step change in how the business runs.
"You can patch for confined use cases. If you're trying to unlock the potential of AI in your business, solving for one specific use case is not going to do that - that's where patching becomes untenable. I have not seen people drive real P&L change without completely building a separate entire data landscape to support it."
- Mahriah, Data & AI Product
Buy Architecture, Not a Model
Asked directly what a mid-market or enterprise fleet should actually be shopping for, Mahriah's answer reframes the entire buying decision away from which AI is best:
"I would suggest you don't buy AI. You should focus on buying architecture. AI is going to change every six months - probably every month at this point - so you want platforms that understand your business data. Models are becoming commoditized. Your proprietary operational data is your real strategic asset as a company."
- Mahriah, Data & AI Product
John's caution to the industry as a whole is the mirror image of the same warning, aimed at vendors rather than buyers: AI "demos beautifully" and is easy to fund, while the unglamorous work of event capture - instrumenting exactly when a truck arrived, when a load got reassigned, why a dispatcher overrode a plan - doesn't demo at all and is easy to underfund.
"My worry is not so much that the industry spends a lot of money on AI, but that money gets spent building sophisticated systems on top of data that can't support them - and then everyone throws the baby out with the bathwater and says the AI didn't work. No - you had to give it a foundation, and that's where things really went wrong."
- John, Applied Science
Clearing the Hurdles: What Actually Works
Across all four interviews, the same handful of practices came up as the difference between fleets who get real value from AI and fleets who don't. None of them require replacing every system overnight.
1. Capture every override, with a reason
"Every override that occurs, capture a reason for it. Anytime a plan gets changed, that is the single most valuable data point the operation generates - it's a human expert saying the system's model of the business is wrong in this specific way. If you're not recording those overrides, it never learns, and dispatchers are stuck fighting against their software forever."
- John, Applied Science
John calls this out as the one wish-list item he'd give any fleet on the data front, and describes it as the fastest practical way to capture tribal knowledge - the countless small rules a dispatcher knows but would never think to recite unprompted - without a multi-month requirements process.
2. Prioritize operational, real-time data over historical archives
Mahriah's advice is to stop optimizing for the biggest possible data warehouse and start optimizing for the data that's actually flowing through decisions right now - because that's the data a system of action needs, and it's usually thinner and more specific than fleets expect.
3. Keep a human in the loop, especially on judgment calls
Every operational interviewee independently landed on the same position: AI should narrow the gap between what a dispatcher wants and what the software can do, not replace the dispatcher's judgment about construction, school zones, or a shipper relationship the system has no way to know about.
4. Insist on optimization guarantees, not just AI-generated suggestions
For anything with a hard constraint - hours of service, weight limits, appointment windows - a deterministic solver should be certifying the plan, not a language model guessing at plausibility.
5. Build or buy for architecture, not for this quarter's model
Because models are commoditizing quickly, the durable asset is the fleet's own well-structured operational data and a platform architected to use it - not whichever model happens to be state of the art this month.
6. Walk the floor before trusting the leadership narrative
The single biggest unlock in several of BeyondTrucks' customer engagements has come from simply observing operations directly - post-it notes, paper trip sheets, and manual workarounds that leadership didn't know existed - rather than accepting "our processes are working fine" at face value.
Closing Thought
None of the people interviewed for this paper described AI adoption as primarily a technology problem. Natasha described it as a fear-and-readiness problem. Matias described it as a discovery-and-trust problem. John described it as a data-architecture problem wearing an AI costume. Mahriah described it as an architecture-first buying decision that most companies get backwards.
The common thread is this: the fleets that get real value from AI are not the ones with the newest model. They're the ones that did the less exciting work first - capturing what actually happened in the field, structuring it so a system can use it, and building enough trust in the output that a dispatcher is willing to act on it. Everything else follows from that foundation.
This paper draws on internal interviews conducted as part of BeyondTrucks' ongoing "AI Hurdles" research series. Quotes have been lightly edited for clarity and length while preserving the speakers' meaning.
Our findings indicate that successful organizations must assertively position themselves in the market and pursue markedly different strategic paths based on their market position.
Private fleet operators must defend their control & quality premium while strategically outsourcing non-core movements.
Specialized and dedicated for-hire carriers must deepen customer integration, quality management, and emphasize risk management expertise.
Common carriers must singularly be focused on pursuing cost leadership.
We estimate that carriers who successfully implement these strategies could achieve 15-20% cost advantages over competitors while building stronger customer relationships that transcend price-based competition. Above all, developing strategic clarity about which of the three paths to pursue will be existential.
The Transportation Market at a Crossroads
The $800 billion U.S. freight transportation market has traditionally operated with three distinct business models:
Private fleets (~50% of US Transportation Spend) that prioritize service quality and control despite higher costs
Specialized for-hire carriers (~20% of US Transportation Spend) that create value through unique equipment and expertise for complex freight
Common for-hire carriers that compete primarily on cost efficiency in a commoditized market segment.
These traditional boundaries are increasingly blurring as each segment faces unique challenges. Private fleets must justify their cost premium, specialized carriers confront line haul commoditization from advancing autonomous technologies, and common carriers struggle with chronic margin pressure from brokerage disruption and market fragmentation. In this evolving landscape, strategic clarity on building competitive advantage becomes crucial for maintaining sustainable fleet operations.
HURDLE 5
Buying AI Before You've Bought Architecture
The fastest way to spend real money and get nothing usable back, according to Mahriah, is to bolt AI agents onto a legacy TMS [https://www.beyondtrucks.com/ai-ready-vs-ai-native-tms] without rebuilding the data foundation underneath it. She describes a pattern she has seen repeat, almost identically, across customer conversations.
"If a CIO came to me having just spent seven figures bolting AI agents into their existing system, I would say you are leaving a lot of money on the table. Ask them how they're measuring their ROI. A lot of times, without significant rework, you will not be getting the answers you expect, and there's going to be a lot more exceptions that a human has to come and fix."
- Mahriah, Data & AI Product
The project she describes starts small and grows quietly out of control: what was pitched as a month-long AI agent implementation stretches to two months, then three, then six - and even once it's technically live, the fleet doesn't realize the value it expected, because people are still making decisions outside the system. The system doesn't understand the business well enough to be trusted with the decision.
How Much Patching Is Too Much?
Mahriah draws a clear line between patching that works and patching that doesn't: a legacy system can absolutely be patched to solve one confined, specific use case. What it cannot do is support a genuine step change in how the business runs.
"You can patch for confined use cases. If you're trying to unlock the potential of AI in your business, solving for one specific use case is not going to do that - that's where patching becomes untenable. I have not seen people drive real P&L change without completely building a separate entire data landscape to support it."
- Mahriah, Data & AI Product
Buy Architecture, Not a Model
Asked directly what a mid-market or enterprise fleet should actually be shopping for, Mahriah's answer reframes the entire buying decision away from which AI is best:
"I would suggest you don't buy AI. You should focus on buying architecture. AI is going to change every six months - probably every month at this point - so you want platforms that understand your business data. Models are becoming commoditized. Your proprietary operational data is your real strategic asset as a company."
- Mahriah, Data & AI Product
John's caution to the industry as a whole is the mirror image of the same warning, aimed at vendors rather than buyers: AI "demos beautifully" and is easy to fund, while the unglamorous work of event capture - instrumenting exactly when a truck arrived, when a load got reassigned, why a dispatcher overrode a plan - doesn't demo at all and is easy to underfund.
"My worry is not so much that the industry spends a lot of money on AI, but that money gets spent building sophisticated systems on top of data that can't support them - and then everyone throws the baby out with the bathwater and says the AI didn't work. No - you had to give it a foundation, and that's where things really went wrong."
- John, Applied Science
Clearing the Hurdles: What Actually Works
Across all four interviews, the same handful of practices came up as the difference between fleets who get real value from AI and fleets who don't. None of them require replacing every system overnight.
1. Capture every override, with a reason
"Every override that occurs, capture a reason for it. Anytime a plan gets changed, that is the single most valuable data point the operation generates - it's a human expert saying the system's model of the business is wrong in this specific way. If you're not recording those overrides, it never learns, and dispatchers are stuck fighting against their software forever."
- John, Applied Science
John calls this out as the one wish-list item he'd give any fleet on the data front, and describes it as the fastest practical way to capture tribal knowledge - the countless small rules a dispatcher knows but would never think to recite unprompted - without a multi-month requirements process.
2. Prioritize operational, real-time data over historical archives
Mahriah's advice is to stop optimizing for the biggest possible data warehouse and start optimizing for the data that's actually flowing through decisions right now - because that's the data a system of action needs, and it's usually thinner and more specific than fleets expect.
3. Keep a human in the loop, especially on judgment calls
Every operational interviewee independently landed on the same position: AI should narrow the gap between what a dispatcher wants and what the software can do, not replace the dispatcher's judgment about construction, school zones, or a shipper relationship the system has no way to know about.
4. Insist on optimization guarantees, not just AI-generated suggestions
For anything with a hard constraint - hours of service, weight limits, appointment windows - a deterministic solver should be certifying the plan, not a language model guessing at plausibility.
5. Build or buy for architecture, not for this quarter's model
Because models are commoditizing quickly, the durable asset is the fleet's own well-structured operational data and a platform architected to use it - not whichever model happens to be state of the art this month.
6. Walk the floor before trusting the leadership narrative
The single biggest unlock in several of BeyondTrucks' customer engagements has come from simply observing operations directly - post-it notes, paper trip sheets, and manual workarounds that leadership didn't know existed - rather than accepting "our processes are working fine" at face value.
Closing Thought
None of the people interviewed for this paper described AI adoption as primarily a technology problem. Natasha described it as a fear-and-readiness problem. Matias described it as a discovery-and-trust problem. John described it as a data-architecture problem wearing an AI costume. Mahriah described it as an architecture-first buying decision that most companies get backwards.
The common thread is this: the fleets that get real value from AI are not the ones with the newest model. They're the ones that did the less exciting work first - capturing what actually happened in the field, structuring it so a system can use it, and building enough trust in the output that a dispatcher is willing to act on it. Everything else follows from that foundation.
This paper draws on internal interviews conducted as part of BeyondTrucks' ongoing "AI Hurdles" research series. Quotes have been lightly edited for clarity and length while preserving the speakers' meaning.
About the Authors
Natasha Martinez
Head of Fleet Success, BeyondTrucks
Matias Oberpaur
Director of Business Development, BeyondTrucks
Dr. John Carlsson
Chief Data Scientist, BeyondTrucks
Dr. Mahriah Alf
Head of Product, BeyondTrucks
About the Authors
Natasha Martinez
Head of Fleet Success, BeyondTrucks
Matias Oberpaur
Director of Business Development, BeyondTrucks
Dr. John Carlsson
Chief Data Scientist, BeyondTrucks
Dr. Mahriah Alf
Head of Product, BeyondTrucks
Our findings indicate that successful organizations must assertively position themselves in the market and pursue markedly different strategic paths based on their market position.
Private fleet operators must defend their control & quality premium while strategically outsourcing non-core movements.
Specialized and dedicated for-hire carriers must deepen customer integration, quality management, and emphasize risk management expertise.
Common carriers must singularly be focused on pursuing cost leadership.
We estimate that carriers who successfully implement these strategies could achieve 15-20% cost advantages over competitors while building stronger customer relationships that transcend price-based competition. Above all, developing strategic clarity about which of the three paths to pursue will be existential.
The Transportation Market at a Crossroads
The $800 billion U.S. freight transportation market has traditionally operated with three distinct business models:
Private fleets (~50% of US Transportation Spend) that prioritize service quality and control despite higher costs
Specialized for-hire carriers (~20% of US Transportation Spend) that create value through unique equipment and expertise for complex freight
Common for-hire carriers that compete primarily on cost efficiency in a commoditized market segment.
These traditional boundaries are increasingly blurring as each segment faces unique challenges. Private fleets must justify their cost premium, specialized carriers confront line haul commoditization from advancing autonomous technologies, and common carriers struggle with chronic margin pressure from brokerage disruption and market fragmentation. In this evolving landscape, strategic clarity on building competitive advantage becomes crucial for maintaining sustainable fleet operations.
About the Authors
Natasha Martinez
Head of Fleet Success, BeyondTrucks
Matias Oberpaur
Director of Business Development, BeyondTrucks
Dr. John Carlsson
Chief Data Scientist, BeyondTrucks
Dr. Mahriah Alf
Head of Product, BeyondTrucks