Waste collection routing is a solved problem in the way that a map is a solved journey. Most fleets already run optimised rounds — the sequence is efficient, the densification work has been done, and the planning software earned its keep years ago. What still costs money, every single day, is the gap between that plan and the round as it actually runs: the blocked access, the over-weight bins, the missed pickup reported at 4pm, the vehicle that fills two streets early.
That gap is where collection operations lose margin and service quality, and it is a decision problem, not a planning problem. Someone has to decide, dozens of times a day, what to do about the exception in front of them right now. Finmile puts AI workers in that seat: they take the plan as their starting point and run execution — dispatch, in-day adjustment, driver communication, follow-through and reporting — against the rules you set.
This page is about what that looks like on a real collection day, and what it changes for the people running it.
The round, as it actually runs
Picture a typical commercial round: 600-odd collections across a city, eight vehicles, a mix of scheduled bins, ad-hoc lifts and contractual service-level commitments. The night-before plan sequences it all cleanly.
By 9:30am the plan is already fiction in three places. A delivery van is parked across a bin store. One round is running hot because trade waste came in heavier than usual. A customer calls: their lift this morning was missed and they need it today. Multiply that by every day of the year and you have the real operating model of most collection fleets: a thin plan, propped up by dispatchers and supervisors doing continuous manual repair.
The cost is rarely one dramatic failure. It accumulates: second passes burning fuel and time, missed collections turning into rebills and complaints, drivers waiting on instructions, supervisors absorbing the day instead of improving it. And the reporting at the end — what was actually collected, what failed, why — is assembled by hand from in-cab notes and phone calls, which is why it is often incomplete.
What changes when a worker owns execution
An AI worker is given a job, not a script. It has an objective (complete today's collections to standard), access to your live data (vehicle positions via telematics, job status from the in-cab system, inbound service requests), and a defined level of authority. Then it works the round the way your best controller would: continuously, against the rules, closing each case to "confirmed done".
A blocked bin store illustrates the pattern. The driver marks access failed. The worker registers it immediately, checks which vehicles are near the failed streets later in the day and have capacity, decides between a second pass by the same vehicle and a reallocation to a neighbouring round, pushes the new instruction to the right in-cab screen, and — the part that usually falls through — makes sure the round record and the customer's account both reflect what actually happened. Nobody had to stop what they were doing to manage it.
The same loop handles the missed-lift call: the worker takes the report, finds today's feasible insertion point, confirms the recovery slot back to the customer, and follows through until the collection is confirmed, not just scheduled.
Where a decision genuinely needs a person — a contract question, a complaint with history, anything outside the worker's defined limits — it escalates with the full context attached. The worker cannot grant itself authority; the boundaries you set are the boundaries it keeps.
The evidence trail, and why it matters commercially
Every action the worker takes lands in an audit trail: what triggered it, what it checked, what it sent, what it decided, what the outcome was. In waste operations this has two practical payoffs.
First, the end-of-day position builds itself. Missed collections, failed accesses, second passes and recoveries are recorded as they happen, so your reporting reflects the day rather than a reconstruction of it. For contract management — especially with municipal and commercial service credits in play — that record is the difference between arguing about what happened and showing it.
Second, recurring problems become visible. When the same street fails access every Tuesday, or one round is structurally over capacity, the pattern is in the data with the evidence behind it. Your supervisors move from firefighting the day to fixing the things that keep causing the fires.
Working with what you already run
Most collection fleets are not short of software. There is routing, telematics, an in-cab app, a customer service desk, maybe a weighing system. Finmile Autonomy does not ask you to swap any of it out. Workers connect to the existing stack — by API and webhook where available, otherwise through approved databases, files and controlled interfaces — and operate in those systems like a trained controller would: reading job state, updating records, sending instructions and confirmations.
Deployment follows one principle: One AI worker. One workflow. Live in 30 days. A typical first workflow for a collection fleet is either missed-collection recovery or in-day exception handling — the two places the plan-to-reality gap hurts most. Prove it on one workflow, then extend: driver communication, service reporting, customer notifications.
For operators who want to go further and consolidate, Finmile OS is the full operating system for the physical operation, with the same autonomy built in. It is an option, not a prerequisite; most fleets start with an execution layer over the tools they already pay for.
What your team's day becomes
Dispatchers stop being the exception-processing bottleneck. Drivers get clear, current instructions on their in-cab screens instead of radio calls and ad-hoc texts, and they stop being sent to places they cannot serve. Supervisors spend the day on the work that actually needs judgement: contract issues, persistent access problems, capacity planning. The customer service team answers fewer "where is my collection" calls, because the updates went out before anyone had to ask.
None of this removes people from the operation. It removes the parts of their day that a machine following your rules does better, and it leaves them the parts that don't.
Frequently Asked Questions
We already have routing software. What does this add? Routing software produces the plan. A Finmile AI worker executes the day: handling exceptions as they happen, adjusting assignments, communicating with drivers, and keeping records current. The two work together — the worker treats your existing routing output as its input.
Can it handle missed collections reported by customers? Yes. End-to-end is the point: the worker takes the report, finds a feasible recovery slot, confirms with the customer, dispatches the instruction, and follows the case until the collection is confirmed complete and the record shows it.
Does the AI decide on its own when to reroute a vehicle? Only within the authority you give it. Actions can run Assisted (it recommends, a person approves), Supervised (it acts inside defined boundaries), or Autonomous (it owns the outcome, with defined exceptions escalated). Most deployments start supervised and widen autonomy as the rules prove out.
What does it need from us to go live? System access, a named owner, the relevant data, and your decision rules. With those in place, a first workflow is live in 30 days.