The Big Haulers Put $100M Behind AI. Almost None of the Margin Comes From the Model.

On 10 June, in a Washington D.C. ballroom, the people who run North America's largest waste fleets did something they almost never do: they put real numbers on their AI. Not a vision-deck "journey." Actual figures, with timelines attached. At the inaugural Waste Leadership Summit, Waste Connections said it's spending about $100 million on seven AI projects through 2027. Republic Services told investors it can pull $100 million of extra EBITDA out of routing by 2028. GFL said one Toronto yard picked up three percentage points of hauling margin in three months. I build these systems for a living, computer vision and sensor stacks and the data pipelines underneath, and I read a disclosure deck the way a mechanic listens to an idling engine. The label says AI. What I want to know is which line items are a real model, and which are plumbing with a model's badge stuck on top.

Here's the deflating short version for anyone selling AI waste management ROI as a step change. The figures the haulers disclosed are worth taking seriously. But almost none of that margin comes from AI in the sense the headline means. It comes from operations research that predates deep learning by forty years, from cleaning up data nobody wanted to own, and from convincing a yard manager to actually use the tool. The model is the cheap part of all of it.

So let me take the five claims I keep hearing, from clients, from vendors, from the slide itself, and hold each one against what got disclosed at the Waste Leadership Summit 2026.

"Put $100 million into AI and you get $100 million of margin out."

Read the Waste Connections AI investment slowly. The company is spending roughly $100 million through 2027 to chase about the same figure in initial margin improvement into 2028 and 2029, with BCG on retainer to get there. A one-to-one first pass, spread across two to three years. Against adjusted EBITDA guidance near $3.3 billion for 2026 [per Waste Connections' own outlook], that lift is about three percent. Real money. Not a transformation. (Republic Services landed on an identical round number for its 2028 EBITDA target; when two boards independently pick the same suspiciously clean figure, you're looking at ambition, not a bottoms-up forecast.)

And that one-to-one return is the tell, not the disappointment. An MIT NANDA study from 2025 found that 95 percent of enterprise generative-AI pilots showed no measurable P&L impact at all, and only about 5 percent hit rapid returns. That same report put the money in back-office and operations automation, not the customer-facing chatbots most budgets chase. The haulers sit in that 5 percent precisely because they aimed at routing and billing, the unglamorous core, rather than a help-desk bot. $100 million sounds enormous until you remember a single front-load truck costs north of $350,000 to put on the road [industry estimate] and these fleets run tens of thousands of them. These are collection-and-billing gains, by the way; the heavier capital lives downstream in the waste-to-energy conversion technology the summit barely mentioned. That's the real shape of AI waste management ROI: small, compounding, operational. It adds up. It just doesn't print.

"This is the generative-AI wave finally hitting waste collection."

Mostly, it isn't. Strip the branding off those seven projects and what's underneath is operations research and statistics that predate the transformer by decades. Route optimization is a constrained-optimization problem you hand to a mixed-integer solver. The pricing engine is elasticity modeling, almost certainly gradient-boosted trees on customer features. The genuinely generative piece is the customer-service layer, and Republic was refreshingly specific about it: roughly 11 million calls a year, about half of them addressable by an AI agent [per Republic's summit remarks]. That layer works, and it's the lowest-margin item on the list. But here's why the distinction isn't pedantic. The MIT finding that 95 percent of pilots return nothing was specifically about generative AI; the haulers mostly are not doing generative AI for the margin items, which is a large part of why their numbers pencil. When a vendor pitches you "AI" for routing, ask whether there's a neural network anywhere in the loop or whether it's a solver with a forecasting model feeding it. The answer changes what you should pay and what can break. (Usually it's the solver, which is good news, because solvers don't hallucinate.)

"Route optimization is a solved problem. You just buy it."

AI route optimization for waste hauling is, underneath, the vehicle routing problem, and the solver that handles it is close to a commodity now. What isn't commodity is the last ten percent: a turn restriction the driver knows and the map doesn't, axle-weight limits, a container that's blocked on Tuesdays, and whether the crew trusts the sequence enough to actually run it. Hauling margin is thin to begin with (you're clearing a gate fee near $55/ton [industry estimate] before the driver's even paid), so the 50 to 60 basis points Waste Connections expects across its 14,000 trucks is a number worth chasing. But none of that gain lives in the algorithm. It lives in the cleanup around it.

GFL is the clean example. Patrick Dovigi credited the Toronto result to "the adoption that we're getting from that management team that runs that specific yard," not to a model. He's been the most measured voice in the room on this.

"Every day there's a new bright shiny object... you can only do so many things. In our perspective, let's do three to five well." - Patrick Dovigi, GFL Environmental

One yard, a bit over 200 residential routes, a manager who bought in. That's the GFL number. It's not 14,000 trucks across a continent, and the figure that survives the jump from pilot to fleet is almost never the pilot figure (route density and union work rules vary wildly market to market). If you want the texture of why the sensor-and-routing layer is harder than the pitch, I got into it in our look at what IoT route optimization actually delivers. Buying the solver is day one, not done.

"It's the model that's printing the margin."

No. The training set is the model; everything else is hyperparameters. The margin lives in the data feeding the thing, not the architecture, and I've paid for that lesson more than once. On an optical-sorter retrofit at a 600 TPD materials recovery facility in 2022, a Tomra autosort unit gave us 94 percent recall on PET, lovely numbers, right up until input moisture spiked above 18 percent and recall collapsed to 71 percent (though that threshold drifts with bale composition). Same model. Same weights. The world changed; the network didn't.

Then there's the failure I'm least proud of. In 2023 I lost six months to a slow precision drop on a classification line that I was sure was a model problem, retraining, tuning thresholds, second-guessing my own labels. It was a camera. Condensate from a poorly sealed enclosure gasket was fogging the lens on cold starts, and the model had been fine the entire time. I'd been wrong about where to look. Precision is easy; recall is where the model lies to you, and sensor drift always wins eventually if nobody owns calibration. So when a deck credits a margin gain to "the AI," the engineer's question is who keeps the inputs honest at 3 a.m. in February. That discipline, not the network, is the product behind any credible AI waste management software.

"The pricing engine is the AI that matters."

This one's half right, which is what makes it dangerous. The pricing engine is the highest-margin item the haulers disclosed. Waste Connections is reportedly generating, in Ron Mittelstaedt's words, "6,000 individual prices for the same services" by pulling very specific data on each customer, and it claims a 20 to 25 percent cut in churn alongside better price retention. Strip the AI label and you've got elasticity modeling and individualized dynamic pricing. Genuinely valuable, and carrying two things the slide leaves off.

First, churn reduction is the single most overstatable number in the deck. Retention attribution is badly confounded: a 20 to 25 percent cut relative to which control group, over what window, against what macro backdrop? I'd want to see the holdout before underwriting a dollar of it. Second, pricing 6,000 customers individually off their own data sits close to the practice the FTC's 2024 surveillance-pricing inquiry went looking for. That reputational and regulatory exposure is priced at zero on the slide, and zero is rarely where it settles.

Now notice what nobody put a number on. Predictive maintenance. Every vendor pitch I've sat through leads with predictive-maintenance ROI, and not one of the majors disclosed a hard figure on it at the summit. That silence is the most honest data point in the building. On a Hitachi Zosen line in 2024, my own predictive-maintenance pilot threw 11 false positives before we caught the labeling drift in the training data, and I'd overestimated the first-year payback by roughly half. Which is why my standing rule is to discount any predictive-maintenance ROI claim by half for its first 18 months, the labeling-drift recovery tax. The haulers know this. They led with routing and pricing, where the operations research is mature, and they stayed quiet on the one thing the brochures love most. Read the silence, not the slide.

Sources & Notes

• The disclosed figures behind this piece, the AI spend and project count, the routing basis points, the churn and EBITDA targets, and GFL's three-point Toronto gain, are drawn from Waste Dive's 18 June 2026 dispatch from the Waste Leadership Summit.
• Dovigi's "three to five well" framing and GFL's bottom-up posture come from a separate Waste Dive interview with the GFL chief executive.
• For the 95-percent-no-return figure and the finding that the durable returns sit in back-office automation, I leaned on MIT's NANDA "GenAI Divide" report, as written up by Fortune in August 2025.
• The 94 and 71 percent recall numbers, the 2023 fogged-enclosure chase, and the 2024 false-positive count are from my own commissioning and pilot work rather than any published study; the 18 percent moisture threshold is feedstock-specific and shouldn't be read as a spec.