AI for Propane Delivery: Tank Levels, Routes and Dispatch Approvals

An AI agent can help a propane or heating oil distributor sort tank-level signals, draft delivery routes and answer routine customer calls, but only after the tank records behind those tasks are reliable. The practical version is narrower than most vendor demos suggest: the agent reads data it is allowed to see, proposes routes and call replies, and sends safety questions and approvals to people. Anything touching safety, credits or a change to a truck schedule stays with a dispatcher.

Key takeaways

• Tank-level routing is only as good as the tank records behind it, so start with a clean register of tanks, capacities, access notes and reading timestamps.
• A capable agent can rank which tanks need attention first, but a dispatcher should approve every route change before a truck moves.
• Delivery windows and routine calls are where an agent saves the most office time, provided urgent cases such as no heat or a suspected gas smell follow a fixed escalation path.
• Safety reports must go straight to a trained person, never to an automated reply.
• Define the approval boundaries, the audit trail and one success measure before connecting more tools.

Why tank levels make routing harder than it looks

A route map assumes you know where each truck needs to be. A fuel distributor works from estimates. Many residential and farm tanks are not read on a fixed schedule, so the level a dispatcher sees is often built from the last reading, the last fill, the season and the customer's usual use. Each of those inputs ages at a different rate. A telemetry gauge may be current, while a driver's note from three visits ago may describe a tank whose access has since changed.

Routing has to balance several constraints at once: truck capacity, driver hours, delivery windows agreed with customers, tanks that are close to empty, partial fills that leave part of a route unused, and access problems such as gates, long driveways or snow. An experienced dispatcher handles this well. The difficulty is that the knowledge sits in several places, and the people who hold it are often on the phone or in a truck.

Illustrative scenario: a distributor runs two trucks across a rural area. One customer's last reading is several weeks old, the previous fill was partial, and the customer has called twice asking whether a delivery can happen sooner. An agent with access to the reading history, the fill log and the access note can flag this tank as a probable early visit. It should not move a truck on its own, because the reading may be stale and the dispatcher may know about a gate code change that never reached the system.

What a more capable agent can change in tank-level routing

A more capable agent does three things well here. It reads the full history of a tank instead of only the latest figure, so it can notice that consumption rises after a cold spell or that a particular customer always orders early. It can estimate the days of supply that remain and show the basis for that estimate, including how old the underlying reading is. And it can draft a route sequence that groups nearby tanks, respects agreed windows and stays within the truck capacity the dispatcher has entered.

What it should not do is treat its estimate as a measurement. A level estimate tells the team where to look first, not whether a tank is safe to leave alone. Every recommendation should carry its source and age, so a dispatcher can tell a telemetry reading from an inference. Seasonal route businesses face a similar pattern of recurring visits and reschedules, and AI for Lawn Care Companies: Quotes, Routes and Seasonal Rebooking is a useful comparison for how recurring routes and rebooking connect.

Delivery windows and customer calls

Most inbound calls to a distributor are short and repetitive: when is my delivery, can you come sooner, how much is left, can I change my window, is my account current. An agent can answer the first three from the schedule and the tank record, offer an open window within a set policy, and log anything else for dispatch. That removes repetitive work from the office during the busiest weeks.

The call flow needs a hard boundary for urgent cases. A caller who reports no heat in cold weather, a smell of gas, a carbon monoxide alarm, a leak, or a tank that appears empty while people depend on it is not making a scheduling request. The agent should stop the normal script, give the safety instruction your company has approved in writing, and transfer the call or page the on-call person. It should never diagnose an appliance or tell a caller that a problem is probably minor.

Callbacks need the same discipline. When a question cannot be closed in one call, the record should hold the reason, the promised time and the named owner. The handoff pattern in Superintelligent Agents for Plumbing Companies: Dispatch and Callbacks applies here: a missed callback is often a handoff failure, not a scheduling failure.

Records and safety checks to build first

Before an agent touches a route, the records must be readable without guesswork. Build them in this order:

• Tank register: site address, tank location, capacity, fuel type, gauge type, access notes and preferred contact method.
• Reading log: each level with its timestamp, its source (telemetry, driver or customer report) and a confidence note.
• Fill and delivery history: volume delivered, partial fills, missed or rescheduled visits, and the reason for each.
• Safety record: last safety inspection date, known hazards, prior gas smell or leak reports, and equipment notes the company is permitted to keep.
• Policy file: approved safety wording, escalation triggers, the on-call roster, and the limits on what the agent may promise about windows or credits.
• Privacy notes: which personal data is held, why it is held, who can see it, and how it is reviewed against the rules where you operate. In Brazil, that means the LGPD and the guidance of the national data protection authority, ANPD.

A simple test helps: if a dispatcher cannot tell from the register alone when a tank was last read and by whom, the agent will not be able to either.

Dispatch approvals and permissions

Permissions should follow the decision, not the software. Agree on the split with the dispatch lead before connecting any system:

• The agent may read the tank register, the reading log and the delivery history.
• The agent may draft a route sequence and mark it as a proposal in the dispatch tool.
• The agent may tell a customer about windows that already exist in the approved schedule.
• The agent may log a callback request with a reason and an owner.
• A dispatcher must approve any change to a route, any move of a tank to an earlier window, and any change to fill size.
• A person must handle every safety report, every credit or refund, and every dispute about a delivery.
• The agent must not edit tank records, safety records or policy text.

Scope each connection to the actions it needs and nothing more. This follows the least-privilege principle in the OWASP authorization guidance. Customer messages and driver notes add a second risk. Free text can contain instructions meant to change the agent's behavior, such as a message telling it to ignore its policy. Treat all customer and driver text as data to be read, never as instructions, and keep the approval step in the system and in a dispatcher's hands rather than in the prompt alone. The OWASP guidance on prompt injection covers this pattern in more detail.

How to measure whether it is working

A tank-routing agent needs an observable success criterion before it goes live, and the dispatch lead should be able to check it without reading model output. Useful examples include the number of urgent-tank visits completed inside the agreed window, the number of callbacks closed without a repeat call, and the number of route proposals accepted without edits.

Log every recommendation, the data it used, the approval or override, and the outcome. Traces structured the way the OpenTelemetry observability primer describes make it possible to see where a recommendation failed: a stale reading, a missing access note, or a misread window. Each time instructions or tools change, run a fixed set of past calls and past routes through the agent and review the failures with the dispatch lead before the change reaches customers.

Review the approval boundary every month. Staff change, customers move, and new tools get added. A permission that made sense in spring can be too broad by winter.

How Be Human approaches this in managed operations

Be Human designs, implements and operates personalized AI infrastructure for companies, and a fuel distributor is a natural fit for a narrow first engagement. The first step defines one useful role, such as a dispatch assistant for tank-level review and delivery windows, along with the memory and tools it needs, the approval boundaries, and one observable success criterion. The agent keeps its identity, memory, channels, tools and permissions under the company's control, so the records and policies stay where the dispatch team can review them. After launch, managed operations can include monitoring, recurring evaluations of routes and call replies, integration maintenance when the dispatch or telemetry system changes, permission review, incident response, and controlled improvements. The service is remote and available in English and Portuguese.

If you want to find the first task worth handing to an agent in your delivery operation, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent decide which tank to fill first? It can rank tanks by estimated days of supply and reading age, and it can show the basis for each rank. A dispatcher should still approve the final order, because the estimate may miss something the team knows, such as an access problem or a customer who has just moved.

Should heating emergencies go to an AI agent? Not as a decision point. The agent can recognize trigger words and transfer the call or page the on-call person using wording your company has approved in advance. Any mention of gas, carbon monoxide or a lack of heat in cold weather is a safety call for a person.

What data is needed before starting? The core set is a tank register, a reading log, a delivery history, a safety record and a policy file. Data quality matters more than data volume, so a smaller set of accurate, dated records will serve the agent better than a large export with unclear sources.

How is this different from routing software? Routing software optimizes over the inputs it already holds. An agent can also read call notes and customer messages, explain its reasoning and draft replies, but it does not replace the optimizer or the dispatcher, and it works best alongside both.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.