In the Federal Reserve's latest Small Business Credit Survey, fielded from September to November 2024 across more than 7,600 employer firms, 51% of firms named uneven cash flow as a financial challenge, and the Fed counts collecting on receivables inside that bucket. Another 56% said paying operating expenses was a challenge. Those two numbers are the same problem seen from two ends: money you earned is sitting in someone else's bank account while your payroll date comes closer.
Collections is the part of finance that small businesses most often run on memory. Someone remembers that a customer is slow, someone else remembers to send a nudge, and nobody remembers to log that the customer disputed line 4 three weeks ago. That is exactly the kind of work software should take. It is also the part of finance where a badly tuned robot can cost you a customer you spent two years winning. When I ran finance at a cultivator, the lesson that stuck from the AR file was which accounts needed a person, not a template.
What AI and automation should own
1. The reminder cadence
The single biggest gain is boring: send the reminder every time, on schedule, with the invoice attached and a way to pay. Humans skip reminders when they are busy, and they are always busy at month-end, which is exactly when the reminders matter.
You likely already pay for this. QuickBooks Online lets you set up to three automatic reminders and schedule them up to 90 days before or after the due date; on QuickBooks Online Advanced you build them as workflows instead. Stripe lets you send automatic reminders for unpaid recurring invoices before, on, or after the due date, and its invoice emails can link to a Stripe-hosted page, so the customer can pay from the email. Intuit says its Payments Agent predicts late payments and drafts reminders, and it claims customers got paid an average of five days faster. Read the footnote before you quote that: it compares Intuit Assist beta users with customers sending standard reminders, from January to August 2024. Treat it as Intuit's own figure.
A cadence that works for most B2B sellers on net-30 terms looks like this:
| When | Channel | Who sends | Tone |
|---|---|---|---|
| 7 days before due | Email with payment link | Automated | Friendly heads-up, invoice attached |
| Due date | Email with payment link | Automated | Plain reminder |
| 7 days past due | Automated, AI-drafted | Asks if anything is blocking payment | |
| 15 days past due | Email to AP contact and your buyer | Automated, reviewed | Firm, cites terms |
| 30 days past due | Phone call | Human | Find the real reason |
| 45+ days past due | Call plus written notice | Owner or controller | Credit hold, payment plan or next steps |
The line between automated and human in that table is the most important design choice you will make. Past 30 days, the reason is rarely that they forgot.
2. Aging triage
An aging report tells you how old each balance is. It does not tell you which balances to work first. A model can rank them: weight by dollar amount, days past due, the customer's own payment history, and whether the last email got a reply. A $40,000 invoice at 20 days from a customer who always pays at 25 is less urgent than a $6,000 invoice at 20 days from a customer who has gone quiet.
This is a good AI job because it is a sorting problem with clear inputs, and a wrong ranking costs you a day at most. Give the person working the list a ranked queue each morning with one line of reasoning per account, and let them reorder it.
3. Payment-link reminders
Every reminder should carry a way to pay in one click. Stripe's hosted invoice page does this, and most ledgers with built-in payments offer something similar. If you send invoices as PDFs from a word processor, fix that before you add any AI; the fastest collection gain is removing the steps between "I should pay this" and "paid".
4. Dispute logging
Disputes hide in email threads. A customer replies to a reminder with "we never got the second pallet" and that sentence lives in one person's inbox. An LLM is good at reading replies and pulling out the claim, the invoice number, the line, and the amount at issue, then writing it to a dispute log and pausing the reminder sequence for that invoice. Pausing matters: nothing annoys a customer more than a "friendly reminder" two days after they told you the invoice is wrong.
What stays human
- Credit holds. Stopping shipments to a customer is a sales decision as much as a finance one. AI can flag that a customer crossed your limit. A person decides whether to stop the truck.
- Payment plans and settlements. Any change to what is owed or when it is due needs a person with authority, and it needs to be in writing.
- Write-offs and bad debt. These hit your P&L and, depending on how you report, your taxes. They need a documented approval from a person.
- Dispute resolution. Logging the dispute is automatable. Deciding who is right, and issuing a credit memo, is not.
- The 30-day phone call. A call finds out that the AP clerk left, that the PO number was wrong, or that the customer is in trouble. Email rarely surfaces any of those.
- Anything that sounds like a legal threat. Do not let a model write the word "collections agency" or "legal action" on its own. Debt collection rules differ by state and by whether the debt is consumer or commercial; that is a conversation for your attorney.
Where AI collections breaks
Three failure modes show up again and again.
Stale data. If a payment lands Tuesday and the ledger syncs Thursday, your bot sends a past-due notice to a customer who already paid. Run the reminder job only after the bank feed and cash application are current, and suppress reminders on any invoice with a payment in transit.
Wrong contact. Reminders that go to the buyer who signed the deal get ignored, because the AP inbox that pays it never sees them. Keep a billing contact on every customer record and check it when the first reminder bounces or goes unanswered.
Tone drift. A model asked to "be firmer" with each round will, by round four, write something you would never send. Lock the templates for each stage and let the AI fill in facts: invoice number, amount, days late, the open dispute. Do not let it pick the tone.
A minimal setup for a small team
- Turn on the native reminders in the ledger or billing tool you already use. Set the first three rows of the cadence table.
- Make sure every invoice has a payment link.
- Keep one AR register: invoice, customer, amount, due date, aging bucket, last contact, next action, dispute flag.
- Each Monday, sort that register by risk and work the top ten by phone.
- Track days sales outstanding (DSO) monthly. If it is not moving after a quarter, the problem is likely terms or customer mix, and no reminder will fix it.
Add an AI layer only after those five steps run on their own for a month. The model needs a clean register to rank and a clear log to write to. Put it on a messy file and it will just send wrong emails faster.
Get the tracker
If you want the register without building it, the AR Aging & Reminder Tracker ($29) is a working Excel workbook with an invoice register, automatic aging buckets, a DSO calculation, and a reminder schedule with email templates for each stage.
If you would rather have this built than build it, Agor AI sets up AR, AP and close automations for small finance teams. See Agor AI. Start with the $250 AI Automation Audit, credited toward a build.
This is part of a weekly series on AI for finance and back-office work. Get it by email.
This article is educational and is not accounting, legal or tax advice. Talk to your CPA before changing how you close your books or file.
Sources
- Federal Reserve Banks: Small Business Credit Survey, 2025 Report on Employer Firms (2024 survey)
- Intuit: Send invoice reminders automatically or manually in QuickBooks Online
- Stripe Docs: Send customer emails (invoice reminders)
- Intuit: Intuit Introduces Virtual Team of AI Agents (July 1, 2025)
- Xero: New AI innovations at Xerocon London (July 2026)