Accounts receivable is where cash flow gets won or lost, yet most finance teams still run it on manual effort. A controller exports an aging report from the ERP, pastes it into Excel, sorts by days past due, then writes the same follow-up emails that went out last month. The work is repetitive, it scales linearly with invoice volume, and it competes with closing the books, forecasting, and everything else that actually requires judgment. The result is predictable: DSO creeps up, high-risk accounts sit untouched for weeks, and the collections process depends on whoever happens to have time that day.
The deeper problem is not laziness or lack of skill. It is that AR work is a high-volume mix of structured data and nuanced communication. Deciding which accounts to chase first requires reading an aging report. Writing a follow-up that is firm without damaging a long-standing customer relationship requires tone control. Escalating a delinquent account requires judgment about history, payment patterns, and dispute status. Each of these steps is small, but together they consume hours every week and still leave gaps.
ChatGPT closes much of that gap. It can read an exported aging file, segment accounts by risk and value, draft tiered collection emails matched to each segment, and produce a summary of collection priorities for the week. It does not replace the ERP or the collections team. It removes the manual sorting, drafting, and formatting that sit between the data and the decision. For CFOs and controllers, that means faster follow-up, more consistent messaging, and an aging report that actually drives action instead of just documenting the problem.
Where ChatGPT Fits in the AR Workflow
The most reliable way to use ChatGPT in AR is to treat it as a drafting and analysis layer, not a system of record. The ERP or accounting platform remains the source of truth. You export the aging report as a CSV or markdown table, feed it to ChatGPT with clear instructions, and get back structured output: prioritized account lists, segment definitions, and ready-to-send email drafts. Nothing is sent automatically. A human reviews and approves every message before it goes out.
This matters for compliance and relationship management. Collections communication is sensitive. A poorly worded email to a strategic customer can cost more than the overdue invoice. ChatGPT is strongest when it is constrained by your standards, your tone rules, and your escalation thresholds, which is exactly what the prompts below are designed to enforce.
First, read these files completely before responding:
[aging_report_export.csv] — the raw AR aging export with columns for customer name, invoice number, invoice date, due date, days past due, amount, and account owner
[collections_policy.md] — our internal escalation thresholds, grace periods, and dispute handling rules
[tone_guide.md] — approved language for first notice, second notice, and final demand, plus phrases we never use
Here is a reference for what I want to achieve:
[Upload a prior week’s worklist as markdown, or describe the format: a table sorted by priority score with columns for customer, total overdue, oldest invoice age, risk tier, recommended action, and owner]
Here’s what makes this reference work:
It ranks by a blended score of amount and age rather than age alone, it separates disputed accounts from clean overdue accounts, it assigns a single recommended action per account, and it keeps the output to one screen so the team can act on it in a morning standup.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: one prioritized table of all accounts over [threshold] days past due, plus a short summary of the top 10 accounts and why they rank highest
Recipient’s reaction: the collections lead should be able to assign the entire list in under 15 minutes without re-sorting anything
Does NOT sound like: a raw data dump, a generic “please pay” list, or anything that ignores dispute status
Success means: every account over threshold has a risk tier, a recommended action, and an owner, and no account is contacted twice in the same week
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
Notice what the prompt does. It does not ask ChatGPT to “analyze my aging report.” It defines the inputs, the reference format, the success criteria, and the constraints. The clarifying-questions step is deliberate: it forces the model to surface ambiguities before producing output, which is where most bad AR automation starts. If the aging export has a column the model does not understand, or the escalation thresholds conflict with the dispute rules, you want that flagged before a single email draft is written.
Drafting Tiered Collection Emails That Protect the Relationship
The second high-value use case is communication. Most collections teams do not struggle to know who owes money. They struggle to write the right message for each account. A first reminder to a reliable customer who is five days late should read nothing like a final notice to an account that has ignored three prior emails. ChatGPT handles this well when you give it the tier definitions and the tone rules explicitly, and when you require it to justify each draft against those rules.
This is also where the reference-file approach pays off. If you have a set of collection emails that have worked historically, upload them as markdown and let the model extract the patterns. The output will match your voice instead of sounding like a generic template, which matters when the recipient is a customer you want to keep.
First, read these files completely before responding:
[prioritized_worklist.md] — the ranked account list from the previous step with risk tiers and recommended actions
[tone_guide.md] — approved language, prohibited phrases, and sign-off conventions for each notice level
[customer_notes.md] — account history, relationship context, and any open disputes or payment plans
Here is a reference for what I want to achieve:
[Upload 3 to 5 past collection emails as markdown, one for each tier, or describe the structure: subject line, one-line context, clear ask with amount and due date, payment link placeholder, and a professional close]
Here’s what makes this reference work:
Each email states the exact amount and invoice numbers in the first two lines, the ask is unambiguous, the tone escalates by tier without becoming hostile, and the sign-off is consistent across all tiers.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: one email per account, maximum 150 words each, grouped by tier, with a subject line for every message
Recipient’s reaction: the customer should understand exactly what is owed and by when, and should not feel accused if the account is in good standing
Does NOT sound like: a legal threat at tier one, an apology at tier three, or a copy-paste template with only the name changed
Success means: every draft passes the tone guide, references the correct invoice numbers and amounts, and requires no more than one edit before sending
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
The two prompts work as a sequence. The first turns raw aging data into a decision-ready worklist. The second turns that worklist into communication. Together they compress a process that typically takes a collections analyst most of a day into a review-and-approve cycle of under an hour, and they do it without removing human judgment from the loop.
A practical tip before you start: run the first prompt on a historical aging report where you already know the outcome. Compare ChatGPT’s priority ranking against what your team actually did. If the model consistently over-weights small, very old invoices or under-weights large accounts nearing a dispute, adjust the scoring rules in your context file rather than rewriting the prompt. The context file is where your standards live, and it should evolve as you learn what works. Once the ranking matches your judgment, move to the email drafts, and keep a human approver on every send for at least the first month. The goal is not to automate collections decisions. It is to automate the sorting and drafting so your team spends its time on the accounts that actually need a conversation.
Published on 24 September 2026 on growwithgpt.com