Three fleet leaders managing a combined 47,000 assets sat down to discuss where AI is actually showing up in fleet operations today, what’s proving valuable, and why adoption is harder than most companies expect.

The conversation wasn’t about futuristic demos or replacing mechanics with automation.

It was about workflows, operational bottlenecks, bad data, and the reality that most AI projects fail long before the model becomes the problem.

Because across all three organizations, one theme kept emerging:

‍AI doesn’t fix operational discipline gaps. It exposes them.

The Pressure Is Real, and So Is the Confusion

Here’s how AI conversations usually begin inside a fleet organization:

Someone sees a demo at a trade show. A vendor says their platform is now “AI-powered.” A LinkedIn post starts circulating internally.

Then IT gets asked:

Can we use this?

The problem is the conversation usually starts with a solution instead of a problem.

ChaChi Gallo, Vice President of IT at Michels Corporation, has been fielding these conversations for years. Michels operates roughly 18,000 pieces of equipment across heavy utility and infrastructure operations globally.

His response is consistent:

“Come to IT with your problems, not your solutions.”

What are you trying to do?

What’s broken or slow?

What would actually change if this worked?

That question stops many conversations cold.

Not because teams are resistant to AI, but because identifying the actual operational bottleneck is harder than evaluating software.

And the companies making meaningful progress with AI right now are not necessarily the ones moving the fastest.

They’re the ones that understand:

• what problem they are solving
• what workflow they are changing
• whether their data is clean enough to support it
• and whether their workforce will realistically adopt it

That sequencing matters more than the technology itself.

What These Teams Are Actually Doing

Voice Agents for Fleet Support Calls

Eric Amlee oversees fleet operations for Primoris Services Corporation, managing roughly 24,000 assets across North America, ranging from light towers and arrow boards to crawler cranes and line equipment.

His team handles hundreds of support calls daily from field crews and internal teams.

That call volume creates a major operational drag:

• answering repetitive questions
• routing requests
• documenting issues
• managing communication loops
• escalating problems manually

The result is that skilled staff spend a significant portion of their day reacting instead of solving higher-value operational problems.

So Primoris began exploring AI voice agents.

The goal wasn’t replacing people. It was removing low-value coordination work.

The goal of the voice agent is to:

• answer incoming calls
• provide basic troubleshooting
• route escalations
• capture information
• and handle repetitive communication workflows simultaneously

As Amlee explained:

“That voice agent is infinitely scalable. It can be answering ten or twenty calls at one time.”

Importantly, this was not treated as an experimental innovation initiative detached from operations.

The business case was calculated against real operational metrics like burden rates, call volume, staffing costs and the cost of the agent itself. 

The economics justified the investment.

And parts of the workflow are already live.

Predictive Maintenance and the Reality of Bad Data

David Heppe leads Data and Technology Solutions at Teichert Construction, a heavy civil contractor operating approximately 5,000 assets, most maintained internally by in-house mechanics.

Their AI focus has been predictive maintenance.

Teichert ran a pilot with a consulting firm using historical work order data to surface patterns that could improve maintenance decision-making.

The outcome was honest and refreshingly direct.

“I think it was moderately successful,” Heppe said. “But in looking at growing it, we recognize the need to get better data in the system and have a better process for all our mechanics to be using because some are better than others.”

That assessment matters because it reflects the reality many organizations discover once AI projects begin:

The limiting factor is often not the model. It’s the operational consistency behind the data.

When work orders are documented differently across mechanics:

• some detailed
• some minimal
• some left open across multiple jobs

The historical record loses reliability.

And without reliable historical records, predictive systems struggle to generate trustworthy insights.

You cannot train a meaningful maintenance model on inconsistent operational behavior.

Deliberate Adoption Instead of Fast Adoption

At Michels Corporation, the approach has been intentionally cautious.

Gallo’s team is still in the discovery and education phase, focusing heavily on:

• standardization
• workforce education
• governance
• and internal readiness

That caution comes from experience.

Before joining Michels, Gallo spent years in manufacturing environments where organizations rushed into emerging technologies before the operational foundation existed to support them.

“We had a lot of failures there because we rushed into a lot of things.”

That lesson now shapes how Michels evaluates AI initiatives.

The emphasis is not:

 “What AI should we buy?”

It is:

“What operational problem are we solving, and are we organizationally prepared to support it?”

The Data Problem Nobody Is Working Around

Every discussion eventually came back to the same issue:

data quality.

Not in the abstract sense of “data matters.”

In the very practical sense that poor operational data becomes the primary blocker to meaningful AI adoption.

At Primoris, the challenge comes from acquisitions.

As the company grows through acquired businesses, it inherits fragmented operational systems, inconsistent standards, and incompatible data structures.

Before that information can support AI initiatives, it must be standardized, validated and integrated.

That process can take years.

At Teichert, the challenge is less about acquisitions and more about workflow consistency.

When mechanics document work differently, the organization loses the signal needed to identify meaningful maintenance patterns.

The solutions discussed across all three organizations were remarkably similar: applying data capture standards.

That’s mandatory fields where possible, dropdown inputs instead of free text with standard operating procedures and ongoing audits to ensure compliance.

None of these ideas are glamorous.

But they are foundational.

Because AI systems inherit the operational maturity of the organization implementing them.

As Gallo put it directly:

“AI is not going to fix your data problem. It’s not going to fix your people’s problems. You still need the people to put in good data, and you still need a good data architecture.”

What AI can do is improve consistency at the point of entry.

It can:

• flag incomplete information
• suggest corrections
• identify anomalies
• prompt missing fields
• and reduce administrative friction

But the underlying operational discipline still has to exist.

The Real Opportunity

One of the most useful ideas from the discussion came from Amlee:

“Don’t build AI to think for you. Build AI to think like you.”

That framing shifts AI from replacement technology to operational extension.

The most valuable implementations are often the least flashy, augmenting tedious and repetitive tasks and documentation or standardizing workflows.

In practice, that could look like a mechanic calling a number, verbally describing an issue, and having the system automatically:

• generate the ticket
• identify the asset
• recommend likely parts
• route the request
• and log the maintenance history

All without the mechanic stopping work, opening a laptop, or manually entering information after a long shift.

For the fleet manager the workflow is documented, the maintenance data becomes searchable and operational visibility improves

And for the mechanic, the end user, the process simply becomes easier.

That distinction matters.

Because adoption rarely happens when AI only benefits leadership dashboards.

It happens when the frontline user saves time, reduces friction, or avoids tedious work.

And that may be the clearest takeaway from this conversation:

The companies seeing the most traction with AI in fleet operations are not treating AI as a standalone technology initiative.

They are treating it as workflow infrastructure layered on top of disciplined operational processes.

Want to watch the full Fleet Operations Shop Talk discussion?

YouTube discussion recording

Speakers

• Eric Amlee, Senior Vice President of Fleet, Primoris Services Corporation
• David Heppe, Data and Technology Solutions, Teichert Construction
• ChaChi Gallo, Vice President of IT, Michels Corporation
• Ben Preston, Co-founder, Gearflow

‍