A Missed Cart and a $1,400 Contamination Dispute Are Not the Same Call. Model Routing Finally Splits Them.

Two calls, ninety seconds apart, worth $8 and $1,400

The Monday after Christmas, a CSR at a hauler with 14,000 residential accounts and 900 commercial stops takes these two calls back to back. The first: a homeowner on a Thursday route whose 96-gallon cart was still at the curb at 6pm. The second: a property manager disputing a $1,400 contamination charge on a 30-yard open top that the MRF rejected and sent to the landfill as mixed waste.

The first call has one right answer and it is in the system already — the lift confirmation off the truck's onboard, the route status, whether the holiday slide moved Thursday to Friday. Cost of getting it wrong: a return trip, maybe $90. The second call has a contract, a photo, a scale ticket, a rejection notice from the MRF, and a franchise agreement clause about who eats a rejected load. Cost of getting it wrong: $1,400 plus a commercial account that leaves at renewal.

Both calls hit the same queue. For most of 2024 and 2025, the shops that put AI on that queue had one setting for both: either a cheap model that was fine on the cart call and dangerous on the contamination call, or an expensive one that handled both and cost enough that owners only ran it on a slice of the volume. That is the thing that changed this year.

Model routing stopped being a science project

Routing means running more than one model behind the same front door and sending each request to the right one — a fast cheap model for the routine, a stronger and more expensive one for the genuinely hard case. It became standard practice in 2026 rather than a clever trick; Cisco built routing into the personal AI agent it is rolling out to roughly 90,000 employees, specifically to keep cost in line with capability. The economics helped: frontier AI costs about a tenth of what it did in 2025, which means the cheap tier is now cheap enough to run on every single contact rather than a sample.

For a hauler, the practical translation is this: routing lets you answer the "was my cart emptied" question for a fraction of a cent while still sending the contamination dispute, the driveway damage claim, and the municipal contract question to a model good enough to be trusted with them — and, more importantly, to a human. The routine tier is not a worse assistant. It is the right assistant for a question whose answer is a lookup.

flowchart TD
  A["Call, email or web form hits the service queue"] --> B{"What is the customer actually asking?"}
  B -->|"Cart not emptied"| C["Fast model checks lift confirmation and route status"]
  B -->|"Roll-off swap or delivery"| D["Fast model books against the dispatch board"]
  B -->|"Fee dispute, damage claim, contract change"| E["Strong model assembles the file, sends no answer"]
  C --> F["Answer sent, note written to the account"]
  D --> F
  E --> G["CSR supervisor reviews and decides the credit"]
  G --> F

Where the line actually sits in a hauling company

Here is the routine list, and it is longer than most owners expect. Was my cart serviced. When is my pickup after the holiday. My cart is cracked, send a new one. I need an extra pickup Friday. Deliver a 20-yard for Thursday. Swap the 30 at the Elm Street job. What is my balance. Change my billing address. Cancel service at the end of the month. Every one of those is a lookup or a straightforward write into Soft-Pak, Trux or AMCS, and every one of them arrives dozens of times a day in the week after Christmas when cardboard is piled next to every cart in the city.

Here is the hard list. Any contamination or overweight charge. Any damage claim — driveway, mailbox, gate, garage door. Anything on a municipal contract, because franchise agreements carry liquidated damages for missed collections and service level credits, and a casual answer from a service agent can become an admission. Anything involving a commercial customer inside a term agreement, especially the ones that call to cancel and are actually calling about price. Anything with a lawyer, an insurance adjuster, or a code enforcement officer on the line. Anything where the words "hazardous," "battery," "propane tank," or "medical" appear.

Notice what makes the second list hard. It is not linguistic difficulty. It is that the answer commits money or admits fault. That is the boundary an owner should draw, and it is drawn in dollars and in document types, not in how complicated the sentence was.

Who decides, and how the line moves

The owner decides, once, in writing, and then the CSR supervisor maintains it. In practice that is a one-page list with three columns: handled automatically, drafted for a CSR to send, never answered without a supervisor. Attach a dollar threshold to the middle column — most shops land somewhere between $50 and $100 of credit authority — and a flat rule that every municipal account and every account inside a term contract escalates regardless of the amount.

Then review the escalation log weekly for the first two months. You are looking for two failures in opposite directions. Things escalating that should not — the same simple question landing on the supervisor forty times because it mentioned the word "charge." And things not escalating that should have, which you will find by reading the credits your CSRs issued and checking whether the agent handled the call that preceded them. Move the line based on that log, not on a vendor's default settings.

One more decision belongs to the owner: what the routine tier is allowed to say about price. Recommended answer for a hauler with an environmental recovery fee and a fuel surcharge on every invoice — nothing. It reads the balance, it does not explain the surcharge. Fee explanations are where a cheerful automatic answer turns into a chargeback.

Running the numbers on a 14,000-account book

Illustrative assumptions: 3,000 inbound contacts a month across phone, email and the web form. 74% of them fall on the routine list. Average handle time on a routine contact for a human CSR is 3.5 minutes. CSR fully loaded cost $27/hour. Assume the routine tier costs about a cent per contact to run and the strong tier about fourteen cents, all in.

The $289 monthly difference in model cost is not the story and no owner should buy on it. The story is that at a penny a contact you can afford to run this on all 3,000 contacts instead of the 400 you piloted, and the 129 hours it frees is roughly three-quarters of a CSR seat during the exact weeks — the post-holiday cardboard surge, the first spring cleanup week, the Monday after every federal holiday when every route slides a day — when you cannot hire one anyway.

What routing does not fix

It does not fix a bad answer on a routine call, and routine calls can still hurt you. If the onboard lift confirmation says the cart was serviced and the customer says it was not, an automatic "our records show it was collected" is the single fastest way to generate a call to the city. Write the routine tier so that a disagreement about facts is itself an escalation.

It does not fix the driver conversation. Half of what a route supervisor does is decide whether the container was genuinely blocked or the driver was behind and skipped it, and no routing decision touches that. It also does not fix the accounts that should never have been sold at that price — an agent answering the cancel call politely and quickly just makes the churn tidier.

And it does not help if your account records are wrong. Routing sends the routine question to a fast model that answers out of your system. If the service day in the system does not match the day the truck actually runs, the fast tier will confidently tell 14,000 people the wrong thing at machine speed. Clean the route master first.

Frequently asked questions

How do I know it is picking the right tier?

Ask for the routing decision to be written on the account note, in plain words, next to the answer: routine, drafted, or escalated, and why. Then spot check thirty notes a week. If you cannot see the decision, you cannot audit it, and you will find out about the bad ones from your city contract administrator.

Does the cheap model sound cheap to the customer?

Not on the routine list. On "when is my pickup this week," a fast model answering in under a second reads better than a hold queue. Where cheapness shows is on anything requiring judgment or sympathy, which is precisely what the hard list keeps away from it.

What about the calls that come in at 11pm?

After hours is where routing pays first, because the alternative is voicemail. Overnight, the routine tier can confirm a service day, book a roll-off delivery, and take a container-blocked report. Anything on the hard list gets captured with a callback promise for the morning — which is a better outcome than a message nobody returns.

Can I start with just the missed-pickup calls?

That is the right start. It is the highest volume, the answer already exists in your system, and the failure mode is small and visible. Get that one right through a full holiday week before you let anything near a fee.

If the front door of this is your phone line, that is the part CallSphere builds: AI voice and chat agents that answer the line and the website chat 24/7, confirm a service day, book a roll-off delivery or swap, and capture the account details on anything that needs a person — then hand the hard ones to your CSR supervisor with the whole conversation attached. The routing decisions above are yours to set. CallSphere just makes sure the 6:40pm call gets answered instead of counted.