Skip to content
be-human.ai
← Back to the journal

AI by industry

AI Agents for Propane Delivery: Tank Levels, Routes and Fill Requests

AI for propane delivery companies: how to organize tank level records, fill requests and route inputs so dispatchers see who needs delivery and why.

Be Human typographic cover with the title: AI Agents for Propane Delivery: Tank Levels, Routes and Fill Requests
Conceptual illustration created with AI · Be Human

An AI agent for a propane delivery company works best as a dispatch assistant. It reads tank level records, delivery history and customer fill requests, then shows dispatchers which accounts need a delivery, why they need it and how urgent each case is. It does not change routes or promise delivery windows on its own, and its value depends on organizing the inputs first so that every recommendation traces back to a reading, a date and a rule that a person approved.

Key takeaways

  • Start with tank level records that carry a source, a timestamp and a status, because a fill decision is only as reliable as the reading behind it.
  • Give every fill request a reason code so dispatchers can see whether it came from a low reading, a customer message, a scheduled fill or a projection.
  • Let the agent rank and explain candidates, while route changes, emergency reports and safety questions stay with people.
  • Limit the agent to read access plus draft actions, with written approval boundaries and a record of every recommendation.
  • Judge the first version by whether dispatchers accept the morning queue with fewer rewritten reasons, and whether any customer ran low because a stale record was quietly reused.

What inputs does a propane dispatch agent actually need?

A propane company usually has more data than it uses, and the data is spread across several places. Customer account records, tank sizes and locations, the last delivered quantity and its date, telemetry from connected tanks where available, manual gauge readings written by drivers, route assignments, truck capacities and the calls, portal messages and emails customers send asking for product all contribute to a fill decision. Each source is written by a different person or system, at a different time, and for a different purpose.

Before any agent is built, list these sources and mark which one is authoritative for which question. The billing system may be right about whether an account is active. Telemetry may be right about the current level until a sensor goes quiet. A driver's note may be the only evidence that a tank was accessible on a given day. An agent that treats all sources as equal will produce confident answers that nobody can check, and that is the failure to design against.

How should tank level records be organized?

Each reading should be a discrete record with a clear set of fields: the customer and tank identifier, the value and its unit, the time the reading was taken, the source (telemetry, driver gauge, customer report or estimate) and a status that says whether the reading is current, aging or stale under a threshold the operations team has set. Estimates must never look like readings. If the system projects a level from past consumption, the projection needs its own label, its own timestamp and a short note on the basis for it.

The staleness rule belongs in the operating procedure, approved by dispatch, not hidden inside a prompt. A reading from yesterday on a tank that usually drains slowly means something different from a reading three weeks old on a tank during a cold stretch. A short checklist keeps the records usable:

  • Each tank has one identifier that matches the account, the site address and the tank size on file.
  • Every level carries its unit, timestamp and source.
  • Estimates are stored apart from readings and labeled as estimates.
  • Stale readings are shown as stale, never silently reused.
  • Corrections keep the original value and record who changed it and why.

How do fill requests become a dispatch queue?

Fill requests arrive from several places: scheduled automatic fills, customers who call or message when they are low, a reading that crosses a threshold, a seasonal pre-fill request or a driver who notices a problem on site. These are different kinds of requests with different urgency rules. A customer who writes that the gauge is low during a cold week is not the same case as a scheduled fill due later in the week, even if both mention the same tank.

The useful output is a ranked list in which every line explains itself. Each line should name the account, the current reading and its source, the requested or projected date, the reason code and anything missing. If the reading is missing, the line says so instead of inventing a number. Dispatchers can then accept, defer or reject each line quickly. A similar queue logic, where flags are explained rather than acted on, is described in this fleet flags example for a car rental company.

A simple decision guide keeps the queue consistent:

  • Current reading below the threshold set by operations: rank by estimated days remaining and ask dispatch to confirm.
  • Customer request with no current reading: show it as a request, flag the missing reading and suggest a gauge check or a call.
  • Stale reading with a scheduled fill nearby: keep the scheduled fill and mark the record for a gauge check.
  • Estimate only: show the estimate and its basis, never as a confirmed level.
  • Any message mentioning a leak, an odor, frost or damage on a tank, or an outage: send it to a person immediately, outside the queue.

What can the agent suggest for routes, and what stays with dispatch?

Route planning depends on things an agent cannot see completely: road closures, driver hours, compartment layouts, gate codes, access notes and the judgment a dispatcher builds over seasons. The agent should therefore offer route inputs, not route decisions. It can group candidate stops by area, mark which stops are time sensitive and total the product the queue would require against the truck capacities on file. It should not reorder a live route, add a stop to a driver's sheet or confirm a delivery window to a customer without a person approving the change.

The same separation of suggestion and approval applies to any operation where an agent touches physical work. A closely related pattern appears in equipment rental returns and damage claims. For a propane company, a practical rollout usually follows these steps:

  • Step one: Agree on the definition of a fill candidate, including the threshold, the staleness limit and what counts as an estimate.
  • Step two: Give the agent read access to account, tank and route data, and keep write access out of the first version.
  • Step three: Write the approval boundary: the agent drafts, dispatch approves, drivers execute.
  • Step four: Run the agent's queue beside the current process for several weeks, logging each place where dispatch disagreed and why.
  • Step five: Review those disagreements with dispatch, fix the data or the rule behind each one, and only then decide whether any narrow action can be delegated.
  • Step six: Set a regular permission review and a clear path for handling a recommendation that turned out to be wrong.

What does a dispatch morning look like in practice?

Illustrative scenario: during a cold spell, a dispatcher at a regional propane company opens the morning queue. The agent lists several candidates, but it marks a few as stale readings, one as an estimate only and two as customer messages with no current gauge data. Its top item is a home whose telemetry reading is below the operations threshold and whose scheduled fill is later in the week. The dispatcher moves that stop up, asks the driver to check a second account along the way and sends a customer message that mentions an odor to the on-call technician instead of the queue. Nothing was routed automatically. The day's record shows what the agent recommended, what dispatch approved and the reason for each decision.

This scenario is an example of how the work could be divided, not a result from any company. The useful test is whether your dispatchers can read the queue in a few minutes and tell, for each line, what the agent knew and what it did not.

How do you keep access and inputs safe?

An agent that reads customer messages carries a specific risk. Text written by customers, or pasted from web forms, can contain instructions meant to change what the agent does. Treat all customer text as data to be summarized, never as a command. Keep the tool set narrow: read-only access to tank records and permission to create drafts in the dispatch queue, with no direct write access to billing or routing. OWASP guidance on prompt injection and on authorization both point toward this pattern, and the NIST AI Risk Management Framework is a useful structure for deciding who reviews the risks and how often.

Authorization decisions should live in the systems of record, not in the agent's instructions. Log every recommendation together with the inputs it used, including the reading, its source and its timestamp. When a dispatcher asks why a stop was ranked high, the answer should come from that record, not from a reconstruction after the fact. Anthropic's guidance on building effective agents makes a similar point: simple, well-bounded patterns with clear checkpoints are easier to trust than broad autonomy.

How does Be Human run this in managed operations?

Be Human starts with one useful role, usually the morning fill queue. The first engagement defines the memory the agent needs, which includes tank records, thresholds, reason codes and escalation rules, along with the tools it may use, the approval boundary and one observable success criterion, such as dispatch accepting the queue without rewriting reasons. Managed operations then cover monitoring of the queue, recurring evaluations against recorded dispatcher decisions, maintenance when telemetry or billing integrations change, permission reviews and incident response when a recommendation is wrong. Rule changes are made in controlled steps, with the operations lead approving each one. The service is delivered remotely, in English or Portuguese.

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

Frequently asked questions

Can an AI agent decide which propane customer gets a delivery today? No, not on its own. A well-designed agent ranks candidate accounts, explains each ranking with its reading, source and reason code, and leaves the decision to dispatch. Route changes, delivery promises and any safety-related message remain with people, and the record shows who approved each action.

Do we need connected tank sensors before starting? No. Manual gauge readings taken by drivers and customer reports can support a useful queue, provided each one is labeled with its source and timestamp. Telemetry improves currency for the tanks that have it, but the agent should show clearly which readings are current, which are aging and which are estimates.

What data should we clean up first? Start with the link between tank identifiers, accounts and site addresses, then confirm units and timestamps on current readings. Next, separate estimates from readings and agree on the staleness threshold with dispatch. These steps are usually more valuable than adding new data sources, because the agent can only be as clear as the records it reads.

How do we know the agent is working? Compare the morning queue with the decisions dispatchers actually make, and record each disagreement with its reason. Watch for two warning signs: stale readings used without a label, and reason codes that dispatchers keep rewriting. If the disagreements shrink after data and rules are fixed, the queue is becoming a dependable input to the dispatch process.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

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

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →