AI Agents for Waste ManagementOperations in 2026

Before anything about models, agents or robots, the arithmetic that governs a hauling business. It is simple, it has not changed in forty years, and every credible AI investment in this industry is an attempt to move one number in it.

Once a collection truck leaves the yard, almost all of its cost is committed. The driver is paid for the shift. The fuel burn is a function of the route, not the number of carts serviced on it. Insurance, maintenance, licensing and truck depreciation are fixed per truck-day whether the vehicle services 300 homes or 500. Revenue, meanwhile, scales with billable stops completed. That asymmetry is the entire business: revenue per truck-day is the margin variable, and route density is what moves it.

The consolidation you have watched for a decade is a density strategy. When a national buys the independent two towns over, the acquisition thesis is usually not the customer list — it is the ability to fold those stops into routes that already exist, and to run the combined book with fewer trucks. The big four spent roughly $3.3 billion on M&A in 2025 and roughly $700 million in the first quarter of 2026 alone. That is what buying density costs when you cannot engineer it.

Here is the FY2025 picture from the filings, which is the only place these numbers should come from.

FY2025 results as reported by the four large publicly traded North American solid waste companies.

One thing you will not find in this article: a cost per stop, a stops-per-shift benchmark, or a percentage density improvement attributed to route optimization software. Those numbers circulate widely and they look authoritative. Every version we could trace resolves to a routing-software vendor's blog or a business-valuation marketing page rather than a filing, an agency dataset, or an operator disclosure. Putting one of them in your board deck is putting a vendor's marketing assumption into your capital plan. The mechanism above is defensible. The precise figures are not, and you do not need them to make the decision.

If a number in your ROI model cannot be traced to a filing, an agency dataset, or a named operator saying it about their own business, take it out. The model is usually still positive without it — and it survives the first hard question.

— Frenchy Digital operating principle

This is the strongest material available in this industry and it is systematically under-read, because it lives in investor communications and trade-press conference coverage rather than in vendor case studies. The large public haulers have quantified their AI programmes in public. Not vaguely — with basis points, dollar commitments and truck counts.

Every figure in the table below is a company statement. None of it is audited, independently replicated, or presented with methodology. Some of it is a forward projection rather than a result. Treated as what it is — large operators telling their investors what they expect from money they have already committed — it is far better evidence than anything a vendor will show you.

AI programmes as described publicly by the large North American haulers, per Waste Dive's coverage of the Waste Leadership Summit and company investor communications. All figures are company statements.

Now read those numbers the way an operator should, which is more sceptically than the headline invites.

Fifty to sixty basis points is a small number that means something. Waste Connections runs AI-enabled routing across 14,000 trucks, built on internal software fed by Google and Waze data, and expects 50 to 60 basis points of margin improvement from it. Against a hauler margin in the low thirties, that is roughly two percent of the margin. It is not a transformation. It is a permanent, compounding improvement that a company operating 14,000 trucks considered worth roughly $100 million of programme spend through 2027 across seven projects. Both readings are correct at once, and the second is the one that should shape your expectations.

Republic's $100 million by 2028 is a projection, not a result. The company has said routing and disposal optimization is expected to add $100 million in EBITDA by 2028, with real-time disposal optimization reaching scale in 2027. Disposal optimization is the interesting half: deciding, in real time, which facility a given load should go to given tipping fees, haul distance, queue times and contractual commitments. That is a genuinely hard optimization problem, it is exactly the kind of thing modern systems are good at, and it requires data plumbing that most regional haulers do not have yet.

GFL's three points is the most striking figure and the weakest evidence. The company says a Toronto yard that automated more than 200 residential routes saw a three-percentage-point increase in hauling margin after three months. Three points is enormous in this industry. It is also one yard, one market, one three-month window, and no published methodology — which does not mean it is wrong, only that it should be read as an existence proof rather than a planning assumption. If you take one thing from it, take the unit of deployment: a yard, not an enterprise.

Republic Services has said it handles roughly 11 million customer calls annually and that it considers about half of them AI-addressable. Of every figure in this article, that is the one that scales down most cleanly to a regional hauler or a municipal collections department, because the reason for the ratio is structural rather than a function of size.

Waste customer service is unusually repetitive and unusually verifiable. A resident calls because the truck did not come, because they want a bigger cart, because they have a mattress, because the holiday moved their day, because they are moving out, or because they do not understand a fuel surcharge. Almost every one of those questions has an answer that already exists in a system you own — the route completion record, the billing system, the service calendar, the container inventory. The agent's job is not to be clever. It is to look up the truth and say it in a sentence, at eleven at night, without a supervisor.

Collection customer-service call taxonomy mapped to systems of record and permitted agent actions — Frenchy Digital, 2026.

The half that is not addressable is the half worth naming out loud: damage claims, injuries, contract disputes, and municipal complaints with political weight behind them. Those calls need a human within one turn, and the design decision is a hard routing rule rather than a model judgment. If an agent has to decide whether a caller is describing an injury, you have already built the wrong thing — the classifier should fail toward a human every time it is unsure.

The measurement to set up before you build: pull twelve months of call volume broken out by reason code, along with average handle time and after-hours abandonment. If your system does not carry reason codes, that is the first project, not the second. You cannot show a return on call deflection without a baseline of what you were deflecting, and reconstructing it retroactively is not possible.

Recycling contamination is the most discussed and worst-sourced topic in this industry. There is one excellent dataset, one useful filed figure, and a great deal of confidently repeated arithmetic that does not survive being traced.

Start with the good data. The Recycling Partnership maintains a dataset covering more than 9,000 programs serving 97% of the US population. It finds that only 21% of residential recyclable material is actually recycled, and — the more important finding — that 76% of that loss occurs at the household level.

Now the economics, which is where the sourcing usually collapses. You will see a multi-billion-dollar annual figure quoted for the cost of US recycling contamination. We could not trace it to any primary source — it appears in dozens of articles and originates in none of them — so it does not appear here, and it should not appear in your business case either.

The defensible substitute is filed. Republic Services reported an average recycled commodity price of $135 per ton in FY2025, down $29 per ton year over year. That is the revenue side of the equation for material you sorted correctly. It is what makes contamination expensive: every ton of the wrong thing was collected, hauled, run across a line, pulled off by hand or machine, and disposed of — at a cost, against a commodity price that fell. When the price of the good outcome drops, the cost of the bad outcome rises even if your error rate never moves.

Contamination statistics in circulation, traced to source — Frenchy Digital, August 2026.

One correction worth making inside your own organization, because it is nearly universal. The United States does not have a current published recycling rate. The 32.1% figure that appears in almost every deck comes from EPA's Advancing Sustainable Materials Managementreport, which was last published in 2020 using 2018 data. It is a seven-year-old snapshot. Cite it if you need to, but cite it with its vintage attached, and treat any 2026 document presenting it as current as evidence about that document's sourcing rather than about American recycling performance.

On-truck detection is the clearest, most operationally sensible AI application in collection, and it is one where the mechanism is easy to explain to a driver, a supervisor and a city council in the same meeting.

A camera watches the hopper. A model classifies what comes over the sill at the moment of tip. The event is stamped with GPS position and route sequence, which attributes it to a service address. That attribution is the whole product — not the classification, the attribution. Once you can say this cart, at this address, on this day, contained this, you have an outreach workflow: a cart tag, a letter, a portal notification, a call, and in a program with teeth, a rejection or a fee.

On-truck contamination detection deployments and the evidentiary weight of each reported result.

WM's Smart Truck runs on thousands of vehicles, including deployments in California and Texas, and the company claims up to a 20% contamination reduction. Read the whole claim, not the number: it is explicitly conditioned on pairing detection with targeted customer outreach. That is an unusually honest framing from a vendor-adjacent source, and it is the most useful sentence in the category. The camera does not reduce contamination. The letter reduces contamination; the camera decides who gets the letter.

Prairie Robotics in East Lansing, Michigan is credited with roughly a 25% reduction, and the figure is attributed to the Michigan Department of Environment, Great Lakes, and Energy rather than to the vendor. That third-party attribution is why it is the strongest number in this section. It is still a single municipal program, and you should ask what else changed in that program during the measurement window — but a state agency attributing a result is a different class of evidence than a case study.

Two other deployments are worth knowing as reference points rather than results: AMCS Vision AI with Peninsula Sanitary Service at Stanford, and Zabble, which has done EPA Phase II grant work and shows up frequently in California SB 1383 organics-diversion contexts. Neither has a published reduction figure we could locate. A logo on a slide is a reference site, not an outcome.

Sorting robotics is the most visible AI in this industry and the most heavily marketed. The technology is real and improving. The published performance numbers are, without exception, vendor claims — measured on someone else's stream, at someone else's line speed, with someone else's inbound contamination. Every figure in the table below carries that qualifier.

MRF sorting robotics vendors and their published specifications. All performance figures are vendor claims and none are independently audited.

AMP Robotics is the scale player, with roughly $266.1 million raised. Its published specifications state up to 99% accuracy at 80 to 120 picks per minute, and its AMP ONE facility in Portsmouth, Virginia is described as processing up to 150 tons of MSW per day at greater than 90% uptime, with claimed diversion above 50% when organics are included. It signed an agreement with Waste Connections in 2020 for 24 systems, expanded in 2022 to more than 50. The deployment history is the most persuasive part of that paragraph — a large hauler expanding an order is a purchasing decision, which is worth more than a spec sheet.

Glacier has raised $28.2 million in total, including a $16 million Series A in April 2025 after $7.7 million in 2024 from Amazon's Climate Pledge Fund. Its Recology Seattle installation runs four robots and four vision systems, handling more than 30 material types at up to 45 picks per minute. Note the ratio: four vision systems alongside four robots. Recycleye in the UK has raised roughly $26 million and publishes no verified throughput figures we could locate.

How to evaluate, in one paragraph. Ignore the headline pick rate. Insist on a trial on your material, at your line speed, over a period long enough to include a bad week. Define uptime to include changeovers, jams, cleaning and maintenance windows, not just powered-on hours. Run an independent composition audit before and after rather than accepting the vendor's own measurement of its own effect. And price the maintenance relationship as carefully as the capital, because a robot that is down is worse than a sort position that was never automated — you removed the person who used to stand there.

Everything above this section is an economic argument built partly on company statements. This section is built on a federal dataset, and it is the argument that should be made first — to a board, to a union, to a city council, and to the drivers themselves.

The Bureau of Labor Statistics Census of Fatal Occupational Injuries for 2024, published in February 2026, puts refuse and recyclable material collectors at a fatality rate of 37.4 per 100,000 full-time-equivalent workers. That makes collection the fifth-deadliest occupation in the United States, behind only logging, fishing and hunting, roofing, and structural iron and steel work.

BLS Census of Fatal Occupational Injuries data for refuse and recyclable material collectors, 2024 reference year, published February 2026.

The year-over-year direction is good and the multi-year direction is not. The rate fell from 41.4 in 2023 to 37.4 in 2024, and total fatalities fell from 41 to 36. But the 2022 rate was 22.6. Whatever changed between 2022 and 2023 has not been undone, and the last three years are materially worse than 2022 on the same measure. Nobody in this industry should be reading a one-year improvement as a trend.

MRF safety is a separate problem with a separate profile: 8 deaths in 2024, 9 in 2023, driven by machinery, conveyors, and lockout-tagout rather than road exposure. Automation changes that exposure directly by removing people from sort positions — and introduces new exposure during maintenance and clearing, which is when a robot cell is most dangerous. If you automate a line, the lockout-tagout procedure is part of the project, not a follow-up.

A word on in-cab systems, because this is where good intentions produce bad outcomes. Driver-facing cameras and behaviour analytics are labour-relations decisions before they are technology decisions. Deployed as a coaching tool with the policy agreed in advance — who sees footage, how long it is retained, what triggers a review, what cannot trigger discipline — they are workable. Deployed as a discipline instrument discovered after installation, they produce grievances and turnover, and the safety benefit never arrives because the drivers you were trying to protect have left. We have no verified figures on the safety effect of in-cab AI in refuse collection specifically, so treat any vendor number in that category as unaudited.

This industry runs on a labour-shortage narrative that is much less well-evidenced than its confidence suggests, and getting it wrong in a board paper is an unforced error.

What can be said: NWRA maintains an active driver-shortage programme, which tells you the industry itself treats the problem as real and current. That is meaningful evidence about industry perception and priority.

What should be said carefully: the figure that gets quoted — the American Trucking Associations' 80,000-driver shortfall — comes from ATA's 2024 report, is industry-wide rather than waste-specific, and ATA has since revised its methodology. It is a trucking number, not a refuse number, and it is not current. Economists have contested the shortage framing itself for years, arguing that what looks like a shortage in the data is better described as a retention and compensation problem in specific segments. You do not have to take a side in that argument. You do have to stop citing a 2024 industry-wide figure as though it described your yard in 2026.

What we will not print: the waste-specific turnover percentage and average-tenure figure that circulate widely, and a specific projection of new collection driver jobs by 2026. Both trace to a business-valuation marketing page rather than to BLS, NWRA, or any operator. They may well be directionally right. They are not sourced, and an unsourced number in a labour business case is the one a sceptical CFO will pull on first.

That last point is the one operators underrate. In most yards, a route supervisor spends a large share of the shift on work that is entirely clerical: reconciling completion data, chasing missed pickups, retyping information between two systems that do not talk, assembling a report for the municipality. None of that requires judgment and all of it consumes the person whose judgment you actually need. It is also, conveniently, the safest place to start, because a mistake produces a wrong document rather than a wrong truck movement.

If you are building a three-year technology plan for a collection or recycling business, the forcing function is not AI capability. It is extended producer responsibility, because EPR is the only thing on the horizon with hard dates and a compliance mechanism attached.

Seven states have enacted EPR packaging laws: Maine, Oregon, California, Colorado, Minnesota, Maryland, and Washington. They are not synchronized, they do not share definitions, and the implementation dates are staggered across the rest of the decade.

EPR packaging law implementation milestones across the seven enacted states, per Mayer Brown's February 2026 compliance survey.

The date that already mattered this year is May 31, 2026, a multi-state packaging data reporting deadline that landed on consumer packaged goods companies. That deadline is why this section belongs in an AI article. EPR reporting runs on material-composition data — tonnages broken out by material category, moving through a defined system. Producers need it to report. States need it to set fees. And the parties best positioned to generate it are the ones with material physically moving across a line: the PRO needs the number, and you are standing on the scale.

Two corrections about federal rules, because both are being mis-sold right now.

Landfill methane. There is no new final rule. Landfills remain governed by the 2016 New Source Performance Standards Subpart XXX and Emission Guidelines Subpart Cf, finalized August 29, 2016, with the federal plan issued May 21, 2021. EPA opened a non-regulatory docket on October 25, 2024 to gather input on new monitoring technologies, with comments running through May 23, 2025 — a signal of interest, not a rulemaking. If a vendor is selling you a compliance product against a 2026 methane rule, ask them to cite it.

PFAS. EPA's September 2025 Unified Agenda anticipates a landfill leachate effluent guideline in 2026 as a proposal, not a final rule. EPA has issued 2026 interim guidance on PFAS destruction and disposal and draft guidance on reducing risk from PFOA and PFOS in biosolids dated June 29, 2026. Guidance is not an enforceable limit. That said, the data posture the guidance implies — leachate sampling records, disposal chain of custody, documented decision-making — is cheap to build now and expensive to reconstruct later, which makes it a reasonable thing to start on ahead of a rule that does not yet exist.

In four years of building operational AI for asset-heavy businesses, the constraint has almost never been model capability. It is the write path into systems that were not designed to be written to.

A hauler's stack is typically a route and dispatch platform, an on-board computer with its own telematics backend, a billing system, scale-house software at the facility, a customer portal, and — in a municipal contract — a reporting obligation that lives in a spreadsheet somebody maintains by hand. Most of these are closed or semi-closed commercial products. Getting data out is usually solvable. Putting a change back in is the project.

Integration patterns for waste operations systems, in descending order of preference — Frenchy Digital, 2026.

Three practical rules follow from this. First, scope the integration before you scope the agent. A two-week feasibility check across your route, dispatch and billing platforms will tell you more about project cost than any amount of workflow design, and it occasionally kills a project cheaply, which is a good outcome. Second, a human commit step is a legitimate architecture, not a compromise. Where no supported write path exists, an agent that prepares the change and a dispatcher who approves it delivers most of the labour saving with none of the operational risk. Third, UI automation needs an owner and a canary. It works, and it breaks silently on a vendor release, usually on a Monday.

One more thing that surprises people: the data quality problem is usually worse than the integration problem. Service addresses that do not match between the route platform and the billing system. Container inventories that drifted from reality three acquisitions ago. Reason codes that three different CSRs use three different ways. An agent inherits all of it, and confidently. Budget for reconciliation in the discovery phase rather than discovering it in week seven.

The sequencing that works is boring on purpose: start where the volume is documented, the answers are machine-verifiable, and the worst outcome is a wrong sentence rather than a wrong truck movement.

Waste operations AI workflows ranked by payback confidence and implementation risk — Frenchy Digital scoping framework, 2026.

Customer service triage is first for three reasons. The volume is documented at industry scale — Republic's roughly 11 million calls with about half considered AI-addressable is the best public benchmark anyone has. The answers resolve against systems you already own, so quality is measurable rather than subjective. And the failure mode is a bad answer with a human appeal path behind it, which is recoverable in a way that a misrouted roll-off is not.

Every one of these has shown up in a real evaluation. Most of them are sourcing failures rather than technology failures, which is why they are easy to catch if you know to look.

The Frenchy Digital red-flag list for waste and recycling technology procurement, 2026.

Ask one question of every vendor number: who measured this, on whose material, over what period, and would they put it in a filing? The good vendors answer immediately and specifically. The answer you get is worth more than the number you were given.

— Frenchy Digital buyer's principle

These are the bands Frenchy Digital uses to scope operational AI work in waste and recycling in 2026. They assume the integration assessment is inside the project rather than discovered inside it, because that discovery is what turns a fixed price into a change order.

Frenchy Digital cost bands for waste and recycling operations AI engagements, 2026.

Senior-led delivery runs $150 to $225 per hour, and ongoing retainers run $2,500 to $9,500 per month covering model and dependency upgrades, integration monitoring as your vendors ship releases, evaluation expansion, incident response, and a quarterly technical review. Every engagement carries a 30-day post-launch warranty, and you receive a written scope with a fixed-price phased proposal within 5 business days of the discovery call. Book that call at calendly.com/frenchydigital/discovery-call or call +1 (424) 272-5601, and bring your route counts, your system list, and twelve months of call volume by reason code.

One budgeting note. The integration substrate is largely a fixed cost paid once. The first workflow carries the connections to the route platform, the billing system and the on-board backend, plus the identity model and the audit pipeline. The fourth workflow inherits all of it and costs a fraction of the first. Operators who sequence their automation get considerably better economics than operators who run four disconnected vendor pilots in parallel — and the second group also ends up with four separate data models describing the same trucks.

A realistic account of where this work disappoints, because the disappointments are predictable and mostly avoidable.

• The public evidence is company statements, not audited results: Every headline figure in this article from Waste Connections, Republic Services, GFL and WM is a company statement, and several are forward projections rather than realized outcomes. That is still the best evidence in the sector, and it is not the same as an independently verified result. Weight it accordingly in a capital plan.
• The margin effect is real and small: Fifty to sixty basis points against a low-thirties EBITDA margin is a permanent, compounding improvement. It is not a business-model change, and a business case that requires it to be one will not survive its first review.
• GFL's three points came from one yard over three months: Single site, short window, no published methodology. It is an existence proof that a yard-level result is possible, not a planning number you can apply to your own book.
• No robot fixes a household problem: With 76% of residential recyclable-material loss occurring at the household level, sorting automation addresses the smaller share of the failure. Capital spent at the MRF to solve a participation problem will underperform, and the post-mortem will blame the technology.
• Contamination detection only works with the outreach attached: WM conditions its own claim on targeted customer outreach. If you buy cameras without funding the letters, the tags, the calls and the appeal process, you have bought a dataset rather than an intervention.
• Vendor specs do not transfer to your line: Pick rates and accuracy figures were measured on a different stream at a different line speed with different inbound contamination. Insist on a trial on your material and an independent composition audit before and after.
• Integration, not intelligence, is where projects die: Route, dispatch, scale-house and billing platforms are mostly closed. Where there is no supported write path, the honest answer is a human commit step — and a project scoped without checking the write path will be repriced.
• Data quality is inherited, and inherited confidently: Mismatched service addresses, stale container inventories, and inconsistently applied reason codes do not announce themselves. An agent will reason fluently over bad data and produce a wrong answer in a confident sentence.
• Prompt injection has no reliable fix: Any agent reading customer email, tenders or resident submissions is exposed. The control is blast radius — read-only tools, no commit capability, separate credentials — not a better system prompt.
• Some of the most-quoted numbers in this industry are unsourced: The multi-billion-dollar contamination cost, the cost-per-stop benchmarks, the density-uplift percentages, and the waste-specific driver turnover figures all trace to marketing rather than to filings or agency data. If one of them is load-bearing in your model, the model is not.
• Safety improvement is inferred, not measured: The BLS fatality data is solid and the mechanism — fewer miles, fewer backing manoeuvres, less road exposure — is sound. But we have no verified figures on the safety effect of AI routing or in-cab systems in refuse collection specifically. Present the mechanism honestly and do not invent an outcome for it.

None of this argues against building. It argues for building one workflow at a yard, with the before-state instrumented, on evidence you can trace, with a human on every action that touches money or moves a truck. The operators who get value from this work are the ones who measured first and expanded second.

And the boundary holds throughout: these are administrative and operational support systems working under human review. An agent drafts a credit, prepares a dispatch, answers a schedule question, and assembles a report. A person commits it. In a business where 23 of last year's 32 collection deaths happened on the road, that is not caution for its own sake — it is the only defensible design.