You did the work. You sent the invoice. Now you are the one who has to remember that it is thirty days old, write a reminder that does not sound desperate or rude, and send it, while also running the business. Some months you chase properly and cash comes in. Some months you forget, and the late payers learn that you forget.
This article is about how to automate invoicing and collections properly: from the accepted quote through the invoice, the reminders, the escalation, and the moment the payment lands, with the accounting system as the single source of truth. It is not about nagging harder. It is about a process that runs the same way every time, is polite at every step, and stops instantly when the money arrives.
We will be specific about what the automation must never do, because collections is where a badly built automation embarrasses you in front of your best customers.
What this actually is
Invoicing and collections automation is a chain of steps triggered by events in your accounting system. A quote is accepted, so an invoice is created and sent. A due date passes unpaid, so a reminder goes out. More days pass, so a firmer one goes out, and eventually a person is told to pick up the phone. A payment is recorded, so everything stops and a thank-you goes out. Each step is ordinary plumbing between your accounting software (QuickBooks, Xero, or similar), your email or texting tool, and a small automation platform.
The ordinary-business analogy is a good credit controller with a tickler file. Every morning they pull the cards for invoices that hit a milestone today, send the standard letter for that milestone, and move the card forward. They never send a letter for an invoice paid yesterday, because they check the ledger first. They never send the day-thirty letter to a customer the owner has an arrangement with, because the card says so. Automation is that person, never forgetful and, if built well, never tactless.
AI fits in sparingly. A language model (software that reads text and writes text back) is useful for reading a customer’s reply (“we sent it Tuesday”, “the amount is wrong”) and routing it, and occasionally for drafting a reminder that references the specific job. It is not useful for deciding amounts, dates, or whether a customer is in arrears. Those come from the ledger, always.
1. Make the accounting system the only source of truth
Every reminder is generated from the live state of the invoice in your accounting software at the moment it is sent, not from a spreadsheet copy or a list exported last week. The automation reads the status, amount outstanding, due date, and contact from the accounting system, builds the message, sends it, and writes a note back. If the invoice shows as paid, partially paid, disputed, or voided, the reminder does not go out.
This one rule prevents the most damaging failure in collections automation: reminding a customer for money they already paid, which happens whenever a copy of the data lives somewhere the payment does not update. QuickBooks, Xero, and most modern accounting tools have connectors on n8n, Make, and Zapier, and an API for anything the connector misses. What AI can and cannot safely touch in the books is in AI QuickBooks Automation: What AI Can Safely Do With Your Books.
2. Start at the quote, not the invoice
Collections problems often begin before the invoice exists: a quote is accepted verbally and the invoice goes out two weeks later, when the customer has forgotten the number. Automate the front of the process: when a quote is accepted (a signature, a click, an email saying yes), the invoice is created from the quote and sent the same day, with the payment link in it and the terms stated plainly.
If you take deposits, the deposit invoice goes out on acceptance and the balance invoice on completion, both automatically, both referencing the quote number. A customer who sees the invoice within hours of saying yes pays faster than one who sees it in a month.
3. Build the escalation ladder and write every rung down
The ladder is the schedule of what happens at each age of an unpaid invoice. A reasonable one for net-30 terms, which you should adjust to your industry:
- Day 0: invoice sent with payment link.
- Three days before due: a courtesy note that it is coming due. Automatic.
- One day after due: a friendly reminder. Automatic.
- Seven days after: a second reminder, still warm, with the amount and date explicit. Automatic.
- Fourteen days after: a firmer note, plus a task for a named person to call. Sent after a person approves.
- Thirty days after: a formal notice referencing the terms. Sent only after approval, with the account flagged for review.
- Beyond: a person, not a system.
Write the ladder as a table with the day, the channel, the tone, and whether it is automatic or approved. That document is the specification for the automation and what you review every quarter. In our experience the day-before and day-after messages resolve most late payments from customers who simply had not got to it.
4. Set the tone rules and never let the system get angrier than you would
Each rung has a tone. Early: warm and brief, assume good faith, make paying easy. Middle: clear and specific, still respectful, name the amount and the date. Late: formal, factual, references the agreement, no emotion. At no point does the automation threaten, guess at reasons, or use language you would be uncomfortable reading aloud to that customer.
Write the templates yourself and read every one aloud. The name, invoice number, amount, due date, and payment link are filled from the accounting record; everything else is fixed text. If a model drafts a more specific version, the fixed facts are still inserted by the workflow, never written by the model, because a model allowed to write the amount will one day write the wrong amount (AI Hallucination Guardrails for Business Applications). One message per invoice per rung, never two in a day, and one consolidated statement when a customer has several overdue invoices.
5. Stop the moment the payment lands
This is the rule that separates a professional system from an annoying one. Payments are recorded in the accounting system by bank feed, card processor, or bookkeeper. The automation checks the live status immediately before each send, so a payment recorded at 9:02am cancels the reminder scheduled for 9:15am. If your accounting system supports webhooks (a message it sends the instant an invoice is paid), use them so the workflow is told rather than having to ask.
When payment lands, one more automatic message goes out: a thank-you with the receipt. If the payment is partial, the ladder resets to the balance. Handle the edge cases explicitly too: a disputed invoice pauses the ladder and creates a task; a customer on a payment plan is excluded; a customer the owner has tagged “handle personally” is excluded entirely. These are fields or tags in the accounting system, checked before every send.
6. Route replies to a person, with AI reading them first
Reminders get replies. “Paid this yesterday.” “This amount is wrong.” “Who are you?” Each needs a different response and none should be answered by a machine. The automation’s job is to get each reply to the right person quickly, with the context attached.
A language model is genuinely useful here. It classifies the reply into a small fixed set (claims paid, requests document, disputes, requests plan, hostile, unrelated) and the workflow routes accordingly: claims-paid to the bookkeeper to verify, disputes to the account manager with the ladder paused, hostile to the owner. The classification must be one of the allowed values or the reply goes to a default human queue. Nothing sends back automatically except, at most, an acknowledgement that a person will respond.
7. Use more than one channel, carefully
Email is the default. Text messages get read faster and work well for the day-before note and the day-one reminder, especially for consumer-facing businesses. They are wrong for formal notices and they require consent: only text customers who agreed to be texted, always include a way to opt out, and keep it short and factual. Twilio and the texting features in many CRMs handle the sending; the automation handles the rules.
Never use a channel the customer has not used with you. A business customer who has only ever emailed you should not get a text about an overdue invoice. Store the preferred channel on the customer record and let the workflow read it.
8. Log every send, review quarterly, and keep the judgement human
Every message the system sends is recorded (invoice, rung, channel, time, recipient), written back to the accounting record as a note. When a customer says “you never told me”, you have the trail. Every quarter, look at days to pay, how many invoices reach each rung, and how many replies each rung generates, then adjust the ladder document, the templates, and the workflow, in that order. The same discipline applies to any automation, as laid out in Automation Error Handling for Businesses Tired of Silent Failures.
Some steps are never the automation’s: the day-fourteen phone call, the payment plan conversation, the decision to stop work, anything going to a collections agency or a lawyer, a customer going through something difficult. For these, the automation creates a task for a named person with the full history attached, and waits. The point of automating collections is to make the routine part invisible so a person’s attention goes to the accounts that need judgement.
Picture a business like this one
The business below is a composite of the kind of company that writes to us, not a client. The numbers describe the shape of the problem, not a case study.
Picture a business like this one: a commercial landscaping company with about 40 staff, roughly 300 active commercial accounts, monthly maintenance invoices plus one-off project invoices. The office manager sends invoices from QuickBooks on the first of the month. Reminders happen when she has time, which is rarely, so days to pay have drifted well beyond terms. Twice in the last year a reminder went to a customer who had paid, because it came from an exported list.
What a company like this would build:
- Contract invoices generated and sent automatically on the first, and project invoices created from accepted quotes the same day, each with a payment link.
- The ladder above, with the day-before, day-one, and day-seven rungs automatic, and the day-fourteen and day-thirty rungs approved by the office manager from a short daily list.
- A live status check against QuickBooks immediately before every send, with paid, disputed, plan, and “handle personally” accounts excluded.
- Text reminders for the first two rungs only, only for customers who opted in, with email for everything else.
- A reply-reading step that classifies every response and routes it: claims-paid to the bookkeeper, disputes to the account manager with the ladder paused, anything hostile to the owner.
- A log of every send written to the QuickBooks customer note, and a quarterly summary of days-to-pay and rung counts.
What changes: collections becomes a ten-minute daily approval list. Nobody who has paid gets a reminder again. Days to pay come down, mostly because of the day-before and day-one messages that were never being sent before, and the accounts that need a call get one on day fourteen instead of day sixty.
What it costs to run
The accounting software you already pay for. An automation platform: a modest subscription on Make or Zapier, or n8n on a small server for ten to twenty dollars a month (n8n vs Make vs Zapier for a Business Without a Developer covers which). Email through your existing provider or a transactional service such as Postmark or Resend, in the low tens of dollars a month at small-business volumes. Text messages through Twilio or your CRM at roughly a cent or two each in the US.
AI usage is tiny because the model only reads replies and occasionally drafts; expect a few dollars a month at most. The meaningful cost is the person who approves the later rungs and makes the calls, time now spent on the accounts that matter instead of on the routine.
The mistakes we see most
Reminding from a copy. The list was exported on Monday, the customer paid on Tuesday, the reminder went out on Wednesday.
Letting the model write the numbers. A drafted reminder states the wrong amount.
Escalating too fast. A day-one message that reads like a day-thirty message, and a customer lost.
No reply handling. Customers reply “paid” to an address nobody reads, and the reminders keep coming.
When to bring in help
If you invoice a few dozen customers a month and use a modern accounting tool, its built-in reminder feature plus a few Zapier or Make steps can give you the early rungs of the ladder without a developer. Set it up and watch it for a month.
A developer earns their fee when you need the live status check before every send, exclusions and pause rules that read tags on the customer record, reply classification with routing, multiple channels with consent handling, and a proper log. That is a small, well-defined project, and because it touches money and customer relationships, it is one worth getting right the first time.
Levelbrook builds collections automation like this for businesses, on a fixed price from a written scope, running in your own accounting system and your own accounts. If your receivables are older than they should be, the form below is where to start.