Sorting a “Tank Reads 12%” Ticket From an Odor Call: Who Draws the Line in Your Dispatch Queue?

Of the 340 messages that hit the counter yesterday, how many actually needed a person?

Ask your customer service lead that question on a January morning and you will get a real answer, because she has been counting in her head for years. At most propane marketers the honest split is something like: two out of three are a delivery date, a balance, a price per gallon, a budget-plan question, or a will-call order for 150 gallons. One in six is a genuine problem — a regulator iced over, a customer at 8 percent on the gauge who is also on credit hold, a tank monitor screaming low on an address with a driveway nobody can get a bobtail up after an ice storm. And a small handful are the ones where somebody says the word "smell."

The trouble is that all of them arrive in the same order, on the same line, into the same inbox. The easy ones are what make you late on the hard ones.

The plain-English definition: model routing is a traffic cop for your own incoming work — the cheap, fast AI handles the requests that look like the three thousand before them, and only the genuinely odd ones get handed up to the slower, more expensive one that thinks harder.

What counts as "routine" at a propane counter in January?

Routine does not mean unimportant. It means the answer is already sitting in your own system and the shape of the question repeats. In Cargas Energy, ADD Systems Energy Force or Blue Cow, that is: last delivery date and gallons, tank size and current percentage from the Otodata or Tank Utility monitor, degree-day burn rate, the account's price tier, budget-plan balance, whether the account is keep-full or will-call, and the delivery notes about the gate code and the dog.

A hard case is one where two of those things disagree, or where safety enters the picture. A customer says he is at 20 percent; the monitor says 9 percent; the last delivery was 31 days ago and the K-factor says he should be at 25 percent. Something is wrong — a new fireplace insert, a tenant who moved in, a leak, or a gauge that is lying. That case needs the thing that reasons, and probably needs a human after that.

The workaround everyone runs today is triage-by-interruption. The CSR handles them in order, and the hard ones get shoved to a callback list on a sticky note that gets worked after 4 p.m. In the second week of a cold snap that list is 40 deep and some of those people have already called the marketer on the other side of town.

Why 2026 is when routing stopped being a science project

Routing between a cheap model and a strong one used to be something a software team built for itself. In 2026 it became normal practice. Cisco built model routing directly into the personal AI agent it is rolling out to roughly 90,000 employees, specifically to balance what each request costs against how much capability it actually needs. When a company that size makes routing the default, the tooling stops being exotic and starts showing up inside the products smaller businesses already buy.

The practical difference from the 2024 version: back then you picked one model and lived with it. Pick the cheap one and it fumbled the odd case. Pick the strong one and every "what's my balance" took fifteen seconds and cost real money. Now the same request stream can go two ways, and the rule for which way is a list you can read out loud.

flowchart TD
  A["Request arrives: call note, text, web will-call form, tank monitor alert"] --> B{"Any word on the safety list: smell, odor, hiss, no heat, ran out?"}
  B -->|Yes| C["Straight to the on-call tech, nothing automated"]
  B -->|No| D["Fast model drafts from the account record in Cargas"]
  D --> E{"Confident, and inside the rules the owner signed?"}
  E -->|Yes| F["Answer sent, will-call order written on the route"]
  E -->|No| G["Strong model reviews the full account and delivery history"]
  G --> H{"Still unclear?"}
  H -->|No| F
  H -->|Yes| I["Queued for the CSR with a one-paragraph summary"]

Who draws the line — and what does the line look like on paper?

Not the vendor. The line is drawn by three people in your building: the owner, the service manager, and whoever carries the safety title. It is a written list, one page, signed and dated, and it belongs in the same binder as the driver qualification files.

What I have seen work is three columns. Column one, answer automatically: delivery date, gallons and price on the last ticket, current balance, budget-plan amount, tank percentage from the monitor, adding a will-call order under a gallon ceiling you set, confirming the route day for a ZIP. Column two, think harder, then send to a person: percentage that disagrees with the burn rate, first delivery to a new address, an account on credit hold, anything involving a tank set, anything where the customer is arguing about a price they were quoted. Column three, never automated: odor, hissing, no heat, out of gas, carbon monoxide, a tank hit by a vehicle, a delivery driver reporting something at a site.

One more rule that matters more than the columns: when the fast model is not sure, it does not guess — it hands up. You want the system biased toward escalating, especially in the first season. And you want a weekly report of what escalated and why, because that report is how column one grows and how you find out that thirty customers this month asked something you have no good answer for.

What the money looks like, and why it is not mostly the model bill

Illustration only. Assume 6,400 residential and small-commercial accounts, a 22-week heating season, and about 38,000 inbound items across phone notes, texts, web will-call forms and monitor alerts — roughly 1,730 a week. Assume 78 percent land in column one.

The model bill difference is $549 across a whole season. That is not the reason to do this. The reason is the third and fifth rows: routine answers come back before the customer puts the phone down, the hard ones arrive at your CSR already summarised, and the seasonal desk hire you have penciled in every October for the last four years stops being automatic. Prove it by counting one thing you can count — the callback list at 4 p.m. If it is 40 deep in December and 9 deep the following December, you have your answer.

Where nothing should be routed anywhere but to a human

Odor calls. Full stop. Not to the cheap model, not to the strong one, not to a queue. If a caller says they smell gas, they get a person and an on-call tech, and the only thing software should do is get them there faster. The same goes for anyone reporting a tank struck by a vehicle, an appliance that will not stay lit, or a headache-and-nausea call that could be carbon monoxide.

Out-of-gas is the second category. It sounds routine — the customer wants gas — but it triggers a required leak check before the system goes back into service, it needs a CETP-trained tech, it may need an appointment where somebody is home, and it usually needs a conversation about why they are on will-call in the first place. Let the system take the details and put it in front of a person immediately.

Third: anything about price that could become a promise. A customer who thinks he locked in at $2.19 in September and is looking at $2.74 on a ticket is a retention conversation, not a lookup. Automated confidence in that moment costs you the account.

Frequently asked questions

How do I keep the fast model from confidently making something up about an account?

Two guardrails. It answers only from your own system's records and shows which record it read, and it is set to hand up rather than guess. Then you read the weekly escalation report yourself for the first two months. If you are not willing to read that report, do not start.

My delivery notes are a mess — "call before, back gate, big dog." Does that break it?

It reads them fine, which is the surprise. What it cannot do is know that "back gate" means the one on Miller Road that has been chained since the sale of the neighbor's field. Free-text notes get better once you see them quoted back to you in answers; most marketers end up cleaning fifty or sixty of them in the first month and leaving the rest.

Does this work for the natural gas side of my business too?

The routine column looks similar — billing, meter reads, service appointments. The safety column is stricter, because an odor report on a distribution system carries response-time obligations that are not yours to negotiate. Route those to a person and to your on-call, and keep the automated part on the billing and scheduling side.

What happens in a July when volume drops by two-thirds?

Costs fall with volume, which is the point of paying per use rather than per desk. The rules you wrote in January still hold in July; you simply see fewer of them fire.

Where CallSphere sits in this

Everything above assumes the request actually made it into your system. On the phone side, that is the hard part in a cold snap. CallSphere builds AI voice and chat agents that answer the line and the website chat, take a will-call order or a delivery question, and pass anything on your never-automate list straight to a live person. The routing rules stay yours; the agent is what makes sure the call is picked up before the caller hangs up and dials the marketer across town.