Superintelligence Prep for Propane Distributors: Delivery and Tank Alerts

An AI agent for a propane company is most useful when it reads delivery schedules, tank level data and safety-check records, then drafts recommendations for dispatchers and service staff. Before it gets more capable, such as rescheduling drops or triggering orders, the distributor needs those three records to be structured, current and owned by named people. The agent inherits the quality of its sources, so the work that matters happens in the records and the approval rules, not in the chat window.

Key takeaways

• Structure delivery, tank and safety records before giving any agent action rights.
• A tank alert needs a threshold, a named owner and a written response rule, not just a notification.
• Safety-check records must show what was checked, by whom, and what happened next.
• Start with an agent that reads and recommends. Dispatch and safety staff keep approval.
• Judge the agent against one observable success criterion, reviewed on a fixed schedule.

Why propane companies should prepare records before adding agents

Propane distribution depends on a few linked facts: which tank is at what level, when a truck can reach it, what the route allows, and whether the last inspection was completed and recorded. Much of that information lives in dispatch software, spreadsheets, driver notes and the memory of experienced staff. An AI agent does not repair gaps in those sources. It makes them faster to act on, so a disagreement between two records can turn into a wrong delivery promise within seconds.

Preparing for more capable automation means deciding in advance which system is the source of truth for each fact. It also means accepting that the first agent should have narrow read access and clear limits. The NIST AI Risk Management Framework frames this as an ongoing practice of mapping, measuring and managing risk across the life of a system. For a distributor, the practical version is simple: know what the agent touches, decide how you will check its output, and name the person who responds when it is wrong.

How should delivery scheduling be structured for an agent?

A delivery schedule an agent can use needs stable fields rather than free-form notes. Each delivery should link to a customer account, a site, a tank identifier, the product, the last delivery date and quantity, the route or zone, access notes, and the reason for the visit. Access notes matter more than most owners expect. Gate codes, animals on the property, seasonal access limits and hazards near the tank belong in the record, written by the person who visited, not in a driver's memory.

The minimum record for scheduling usually includes:

• One account identifier shared by billing, service and dispatch.
• One tank identifier per site, with its location described in plain words.
• A route or zone assignment that changes only through dispatch.
• Access and hazard notes that are dated and attributed to their author.
• A status for each drop, such as planned, confirmed, delivered, held or cancelled, with a reason recorded for every hold.

Scheduling rules also need to be written down. Decide which deliveries can move without a dispatcher, which need a call to the customer, and which cannot move at all during a cold spell. An agent given a vague rule will guess, and a guess about a delivery date is hard to notice until a customer runs low. A written rule gives the agent something concrete to check its suggestion against. Equipment yards face a similar problem with returns and fleet availability, and the approach in structuring equipment returns and damage records follows the same logic.

What should tank level alerts contain?

A tank alert is only useful if it tells someone what to do. A message that says a tank is low gives the dispatcher a question but no answer. A usable alert shows the tank identifier, the reading and its timestamp, the source of the reading, the threshold that triggered the alert, the account's current delivery status, and the named person who owns the response.

The source matters because not every number is a measurement. A reading from a telemetry feed, a manual gauge check and an estimate based on historical usage are different kinds of information. An estimate should be labeled as an estimate. The most common failure is an old gauge reading presented as if it were current. Every reading an agent can see should carry its age, and the alert should say so when the reading is stale.

Thresholds belong to the distributor's operations and safety policy, and a specific person should be accountable for them. An agent can point out that usage patterns have shifted and suggest a review. It should not change a threshold on its own.

A low-level alert without an owner is a notification, not a process.

Illustrative scenario: at the start of a cold week, a telemetry feed reports a falling level at a residential tank, while the last manual gauge check is several days old. A well-structured alert shows both readings with their dates, labels the telemetry value as the current source, flags the mismatch, and routes the item to the dispatcher on duty. It does not move the customer up the route without a person approving the change.

Which safety-check records must exist before automation?

Safety records are where an agent can cause the most harm by sounding confident. Each inspection record should show the inspection type, the equipment or tank identifier, the technician, the date and time, the observations, the result, any corrective action, and the sign-off. Corrective actions should stay linked to the original finding, so an open item cannot disappear when the next visit is booked.

Check your local fuel regulations, insurer requirements and internal procedures for the exact inspection intervals and form requirements. An agent should not decide those intervals or mark a check as complete. Its role at this stage is to read the records, summarize what is due, and point out open corrective actions to the person responsible. Keeping sign-off with a qualified person is a design decision, not a formality.

The OWASP Authorization Cheat Sheet is a useful reference for this stage. It describes checking access at each request and applying least privilege, which means an agent that drafts a safety summary should not also be able to edit the inspection history.

Where should the agent's authority stop?

Think of agent capability in three levels. Reading means retrieving and summarizing records. Recommending means drafting a reschedule, a customer message or a note for dispatch, with nothing sent or saved to a system of record. Acting means writing to a system of record or contacting a customer. Most distributors should begin with reading and recommending, then grant action rights one task at a time, with logs and a named approver.

A practical sequence looks like this:

• Name the source of truth for delivery dates, tank readings and safety sign-offs.
• Write the threshold and response rule for each alert type, with a named owner.
• Add source and age labels to every reading the agent can see.
• Give the agent read-only access to safety records, and keep sign-off with qualified staff.
• Require a recorded approval for any change to a customer delivery.
• Log every suggestion and its outcome, including the ones staff ignored.
• Review wrong and ignored suggestions weekly before expanding any permission.

The agent will also read customer messages and driver notes, and some of that text may try to change its behavior, for example by asking it to ignore a hold. The OWASP guidance on prompt injection recommends treating untrusted text as data to be analyzed, not as instructions. The Top 10 for LLM Applications from OWASP covers the wider risk categories to review before connecting an agent to live systems.

How Be Human approaches this and where to start

Be Human designs, implements and operates personalized AI infrastructure for companies, with agents that keep their identity, memory, channels, tools, permissions and operating context under human control. For a propane distributor, a first engagement defines one useful role, such as a dispatch assistant that reads delivery and tank records and drafts alerts for staff. It also specifies the memory and tools that role needs, the approval boundaries, and an observable success criterion, such as alerts reviewed within a set window each day. Managed operations can then cover monitoring, recurring evaluations, integration maintenance, permission review and incident response, so access changes are controlled rather than accumulated over time. The service is remote and available in English and Portuguese.

If you are unsure which delivery, tank or safety task is the right first one to hand to an agent, take the free Be Human AI diagnostic to find it.

Frequently asked questions

Can an AI agent for a propane company schedule deliveries on its own? It can draft schedules and flag conflicts, but a distributor should not let it confirm or move deliveries without an approval step at first. Dispatch rules, access notes and hold reasons need to be structured before the agent can make a reliable suggestion. Action rights can be added later, one task at a time, once the logs show consistent results.

Should an agent be allowed to change a tank alert threshold? No, not as a standing permission. Thresholds reflect operations and safety policy, so a named person should own them and approve any change. The agent can present usage evidence and recommend a review, which keeps the decision with the accountable owner.

What is the first task to hand to an agent at a propane distributor? A good first task is usually read-only and repetitive, such as producing a morning list of tanks with stale readings, open alerts and unresolved corrective actions. It has clear inputs, a visible output and a simple success criterion. Staff can check it quickly, which makes errors easy to catch.

Does this preparation matter if we only run a few trucks? Yes, because small teams often depend more on individual memory, which is harder to hand off or audit. A small fleet still needs a clear source of truth for each tank, a written response rule for alerts and a safety record that survives staff changes. The amount of structure can be modest, but the decisions still need to be written down.

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