Petroleum AI Field Report

1. THE OPERATOR PROBLEM

Every Monday someone in accounting pulls the aging report and decides who to call. In a fuel business that list is longer and riskier than it looks. Commercial tank wagon and bulk accounts run on trade terms. Cardlock accounts keep fueling until someone lowers the limit. Keep-full propane and heating oil customers get product whether or not anyone checked their balance. The aging report shows what is owed. It does not show what is about to be delivered.

When prioritization goes wrong, the costs stack up: a driver pumps another load into an account that is 75 days out, a cardlock account runs up charges after the owner stopped paying, a collector spends the morning on small balances while a large contractor slides past 90 days, or a customer with a pricing dispute gets a past-due call and leaves. A second failure is the reverse: a hold placed in error on a residential heating customer in January.

Two federal rules shape the work, and both should be checked with counsel:

• The FDCPA defines a debt collector as one who collects debts "owed or due another," and it covers consumer debts "primarily for personal, family, or household purposes." A creditor collecting its own debts is pulled in if it "uses any name other than his own." The U.S. Code text is current through Pub. L. 119-120 (Oct 7, 2026); paragraph (6) was last amended in 1986. 15 U.S.C. 1692a: https://uscode.house.gov/us/usc/119-118/t15/s1692a/ Regulation F mirrors this at 12 CFR 1006.2(i); eCFR lists Part 1006 as last amended Apr 19, 2023. eCFR Part 1006: https://www.ecfr.gov/current/title-12/chapter-X/part-1006
• The FCC's Declaratory Ruling FCC 24-17 (released Feb 8, 2024) holds that the TCPA's "artificial or prerecorded voice" restrictions cover AI technologies that generate human voices, so AI voice calls need prior express consent. FCC 24-17: https://www.fcc.gov/node/261899

Interpretation: a distributor calling its own residential propane or heating oil customers in its own name is generally outside the FDCPA. A collection agency you hand those accounts to is not, and some state collection laws also cover creditors collecting their own accounts, so check each state you serve. An AI voice bot calling customers is a TCPA question before it is a technology question.

2. WHAT ACTUALLY DOES THE WORK

• Plain rules, not AI: aging buckets, credit limits, automatic delivery holds, cardlock limit blocks, and a worklist sorted by dollars at risk. This is where most of the value sits.
• Simple arithmetic, not AI: exposure = open balance + unbilled deliveries + scheduled keep-full stops + unbilled cardlock transactions. Sorting by that number is not optimization. It is a spreadsheet column.
• Machine learning and statistics: a payment-propensity score that estimates the chance each invoice is paid late, trained on your own invoice and payment history. Microsoft's documentation describes exactly this: three probabilities per invoice (on time, late, very late), with invoices under 50 percent on-time probability flagged for a collector. Microsoft Learn: https://learn.microsoft.com/en-us/dynamics365/finance/finance-insights/payment-insights-overview
• Generative AI: drafting reminder emails, summarizing a long dispute thread, and pulling the BOL or price question out of a customer reply. Useful, and every message needs a human send.

My view: fix the rules and the exposure number first. If your keep-full routes do not check account status before dispatch, no model will save you.

3. DATA REQUIRED AND WHERE IT LIVES

• Open invoices, payment history, terms, and credit limits: the ERP (PDI, Cargas, Dynamics, QuickBooks, or similar).
• Scheduled and unbilled deliveries: dispatch and the keep-full forecast, often fed by tank monitors.
• Cardlock activity and limits: the card network portal or your own cardlock system. Unbilled transactions are exposure.
• Disputes and promises to pay: the AR inbox, collector notes, and phone logs. This is the messiest source and the one generative AI helps with most.
• Customer master flags: commercial or residential, budget plan, seasonal payer (for example, ag accounts that pay after harvest), personal guarantee on file.
• Unapplied cash: lockbox and bank deposits not yet posted. Skip this and your first alerts will be false alarms.

4. BUILD, BUY, OR CONFIGURE WHAT YOU OWN

Configure (petroleum ERP): Cargas Energy's product page says its software can "automatically place customers on hold based on parameters you define" and "organize and prioritize delinquent accounts with the Collections Assistant." Those are vendor descriptions; test them on your own accounts. Cargas Energy: https://cargasenergy.com/software/customer-service/features/ Ask your ERP vendor the same two questions: does a hold stop dispatch, and does the worklist include unbilled deliveries?

Configure (general ERP): Dynamics 365 Finance offers customer payment predictions built on AI Builder; Microsoft says a power user can choose the fields used and retrain the model. Microsoft does not publish a minimum history requirement on that page. Microsoft Learn: https://learn.microsoft.com/en-us/dynamics365/finance/finance-insights/payment-insights-overview Intuit says the QuickBooks Payments Agent predicts late payments and sends invoices and reminders. Intuit's claim, not verified: customers get paid about five days faster on average. Accounting Today: https://www.accountingtoday.com/news/intuit-debuts-ai-agents-for-quickbooks

Buy: AR automation suites add scoring, dispute workflows, and cash application. Most are built for larger finance teams. Treat every percentage on their sites as a claim until a pilot on your data supports it.

Build: an exposure-sorted worklist in SQL or a spreadsheet, plus a basic logistic regression if you have the history and someone to own it. For most small distributors, configure first.

5. A 30-DAY PILOT

• Scope: commercial tank wagon, bulk, and cardlock accounts at one branch. Leave residential heating accounts out of the pilot.
• Week 1: post all unapplied cash, define exposure, and write down the current worklist method. The credit manager sets the score threshold and the stop condition before day 1.
• Weeks 2 to 4: run the ranked list beside the old list. Collectors work the ranked list; the old list stays as a check.
• Human review point: the credit manager approves every hold, limit change, and outbound message before it happens.
• Measure: collector time to build the daily list; accounts that went past 60 days and were ranked high beforehand (caught) or low (missed); false alarms, meaning flagged accounts that paid on time or were in dispute; holds placed and released within 48 hours.
• Stop condition: any hold placed in error that stops a delivery, any message sent without approval, or the ranked list catching fewer slipping accounts than plain aging did.

6. CONTROLS AND FAILURE MODES

• Audit trail: log the score, the top reasons, the recommended action, who approved it, and the outcome. Keep it with your credit files.
• Permissions: the tool reads AR and dispatch data. Only the credit manager sets holds and cardlock limits. Only named staff send messages.
• Sign-off: the credit manager owns thresholds; the controller or general manager signs off before scores touch dispatch.
• Failure modes: unposted payments, price or BOL disputes scored as delinquency, budget-plan and seasonal payers flagged every month, and a model that learns only from the accounts someone already chose to call.
• Never automated: residential delivery cutoffs, placing accounts with an agency, legal action, write-offs, settlement offers, dispute rulings, and AI voice calls without documented consent.

7. CURRENT DEVELOPMENT

On Oct 14, 2026, Genpact announced general availability of its Accounts Receivable Suite, including a collections module that ranks accounts and suggests next-best actions from risk and payment signals. The release targets industrial clients in manufacturing and mining. Genpact's claims, which it calls "indicative estimates only": up to 20 percent fewer past dues and up to 40 percent higher collector productivity. Genpact release: https://media.genpact.com/2026-10-14-Genpact-Launches-Agentic-Accounts-Receivable-Suite-to-Optimize-Cash-Flow-and-Working-Capital

Interpretation: this is enterprise outsourcing, not a small-distributor product. It does show where AR tools are headed: ranking plus suggested actions, with humans on exceptions. The release does not describe approval thresholds or audit controls, so the controls above still apply.

DEEPER READS

1. Microsoft Learn: Customer payment predictions overview: https://learn.microsoft.com/en-us/dynamics365/finance/finance-insights/payment-insights-overview
2. FCC Declaratory Ruling FCC 24-17: AI-generated voices under the TCPA: https://www.fcc.gov/node/261899
3. 15 U.S.C. 1692a, FDCPA definitions (Office of the Law Revision Counsel): https://uscode.house.gov/us/usc/119-118/t15/s1692a/

Not legal advice; confirm applicability with counsel or the issuing agency.

Also from Pipeline Punks: the Vendor and Tech Watch (claims vs. results). Sign up or pick your briefs at https://briefs.pipelinepunks.com/signup

Pipeline Punks - Question the hype. Keep it flowing.