Clients pay late and I hate chasing: can an AI agent do it safely?

By Ranjit Bhinge, founder of Blur Studio · Updated · 7 min read

Short answer: Yes, for the early, polite reminders, and not for everything. An AI agent here is software that reads your list of unpaid invoices, writes a reminder in your voice and sends it on a schedule. It should handle the friendly nudges alone, draft the firmer notices for you to approve, and stop completely when a client disputes an invoice or says they are in trouble. Accounting tools such as QuickBooks, accounting software used by many small firms, already send plain scheduled reminders. The agent adds the writing, the reading of replies and the handoff to you.

Ranjit Bhinge

You are tired of chasing late payers by hand, and you want to know which parts an AI agent can take over without damaging a client relationship.

I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.

Chasing late invoices is a good job to hand to software, for one reason: it is repetitive, rule-based and unpleasant, so it often gets done late or not at all. The risk is not the typing. The risk is a cheerful automated email landing on a client whose invoice you got wrong, or who has just lost a family member.

So the useful question is not "can AI chase invoices?" It is "which steps are safe to automate, and where must a person decide?" If you want the manual version first, read how to stop chasing late invoices by hand. For the bigger picture, see business process automation and AI workflow automation for small business.

What your accounting tool already does

Before you add an AI agent, switch on what you already have. QuickBooks Online lets you turn on up to three automatic reminders, each set up to 90 days before or after the due date, and it only sends them for invoices you have already emailed from QuickBooks. You can edit the subject line and the message. Xero, another accounting tool, has its own invoice reminder feature; check Xero's help centre for the current options on your plan.

These built-in reminders are plain and fixed. They do not read replies, know that a client always pays on the 15th, or notice that three invoices to the same client are now late. That is the gap an agent fills.

What an AI agent adds

An AI agent is software that uses an AI model to carry out tasks, not only answer questions (see what an AI agent is). For invoices, four jobs are realistic today:

  1. Reading the data. It pulls the list of overdue invoices from your accounting tool, with amounts, due dates, client and past reminders.
  2. Writing the message. It drafts a reminder in your tone that mentions the right invoice, amount and payment link, and that changes with how late it is.
  3. Reading replies. It sorts answers into "will pay on a date", "already paid", "dispute", "question" and "angry or distressed". A language model is good at this sorting, though not perfect.
  4. Handing off. It puts the tricky ones in front of you with a short summary, and it logs everything.

Reading and sorting your invoices as documents is also possible with AI, but that is a separate job from chasing. If you want that side, see accounts payable automation and AI bookkeeping.

The chase sequence

This is the sequence I would set up. The days are a starting point. Change them to fit your terms and your clients.

WhenMessageWho sends it
3 days before dueFriendly heads-up with the payment linkAgent alone
1 day after due"This may have crossed with your payment." Invoice attached.Agent alone
7 days afterPolite reminder, asks for a payment dateAgent alone
14 days afterFirmer note that restates the termsAgent drafts, you approve
30 days afterPhone call or personal email from youYou
45 days and overFormal notice, pause on new work, or collections decisionYou, with a lawyer if needed

Two rules make this safe. Every message from day 14 onward needs your approval. And the agent stops the sequence the moment a reply arrives that is not a plain "will pay".

What the agent must never say or do

  • Threats. No mention of legal action, credit reports, late fees or collections unless you wrote that exact wording and your contract allows it.
  • Claims about the amount. It quotes the figure from your accounting tool and nothing else. It never computes interest or "adjusts" a balance.
  • Anything about a dispute. Once a client says "this is wrong", the agent pauses all messages on that client and alerts you.
  • Anything about hardship. If a client mentions illness, a death, a job loss or money trouble, it stops and hands over to you.
  • Promises. It never agrees a payment plan, discount or extension.
  • Pretending to be you. Sign it from the business, and say who to contact for a person.

OWASP, a nonprofit that publishes security standards, says in its guidance for AI apps that a person should approve high-impact actions such as sending, and that an agent should have only the access its job needs. For this agent that means read access to invoices, permission to send from one email address, and no access to edit, delete or write off anything. See AI agent guardrails for small business for the full rule sheet.

Debt collection rules in the United States are mainly aimed at consumer debts. The Federal Trade Commission says the Fair Debt Collection Practices Act covers credit card debt, car loans, medical bills and other household debts, and that business debts are not covered. The Consumer Financial Protection Bureau's Regulation F implements that law for debt collectors. So if you invoice other businesses, those rules mostly do not apply to your own reminders. If you invoice individual consumers, or you hand overdue accounts to a collection company, they may. This is general information, not legal advice. Ask a lawyer before you automate firm or final notices, and check your state's rules and your own contract terms.

Worked example with made-up numbers

A six-person design studio bills about 40 invoices a month, averaging $2,500. Say 12 of them are paid late and the owner spends about 4 hours a month chasing. Those figures are invented for illustration.

  • The agent sends the day 3, day 1 and day 7 messages on its own. Say it resolves 8 of the 12 with no human time.
  • Four reach day 14. The agent drafts four firmer notes. The owner reads and approves them in about 10 minutes.
  • One client replies that a line item was billed twice. The agent pauses, tags it "dispute" and sends the owner a two-line summary. The owner fixes the invoice and the client pays.

The owner's time falls from about 4 hours to under an hour. The larger gain is that chasing happens on time, every time. Your real numbers will differ, so measure your own late-payment count for a month before you decide.

How to set it up, step by step

  1. Clean the data. Check that client emails, due dates and payment links are right in your accounting tool. An agent will repeat any mistake faithfully.
  2. Switch on the built-in reminders first. Use them for the day-before and day-after steps.
  3. Write your tone sheet. Three sample emails that sound like you, plus a list of phrases you never use.
  4. Define the stop list. Disputes, hardship, key clients, anyone with an open question. Put these accounts on a hold list.
  5. Run it in draft mode. Anthropic, the company that makes Claude, warns in its guide to building agents of "the potential for compounding errors" and recommends extensive testing in sandboxes (safe practice copies). For a month, the agent writes and you send. Compare its drafts with what you would have sent.
  6. Let it send the first three steps alone. Keep approval for everything firmer.
  7. Review weekly. Read the log, check the sorting of replies, and adjust the wording.

You can build a version of this with automation tools plus an AI model, or ask a developer. Our guide to the AI email assistant covers drafting replies, and AI employee explains how to think about handing work to software.

Common failure points

  • Reminding on a paid invoice. If a client paid by bank transfer and you have not matched it yet, they get a nag. Match payments first, or add a three-day delay before the day-1 message.
  • Wrong contact. The invoice goes to a person who left. Keep an accounts-payable contact on file.
  • Tone drift. The AI gets stiffer or friendlier than you. Re-read samples monthly.
  • Chasing a good client like a bad one. Put your best clients on a hold list so they get a personal note from you instead.
  • No one owns it. Name one person who reads the log.

What AI cannot do here

It cannot judge a relationship. It cannot tell that a long-term client is having a bad quarter and deserves a call. It cannot decide to write off a balance. And it cannot know that an invoice is wrong unless the client tells it. Those are your decisions, so the sequence brings you in from day 30.

How I'd approach this

I would start smaller than you think. Turn on your accounting tool's built-in reminders this week. After a month, look at which invoices still go late and why. If most are late because of reminders that sound cold or arrive too rarely, a drafting agent will help. If most are late because of disputes or slow approvals on the client's side, no reminder fixes that, and the real problem is your invoice process.

When you build the agent, I would keep it read-only on your books and keep a human yes on every message from day 14 on. Chasing money is personal. The software should make you faster at it, not speak for you in the moments that matter.

Questions people ask

Can an AI agent send overdue invoice reminders automatically?

Yes, for early and polite reminders. Your accounting tool may already do plain scheduled reminders, for example QuickBooks Online offers up to three automatic reminders. An agent adds tailored wording and reading of replies, and should hand firm notices and disputes to a person.

Is it safe to let AI write payment demands?

Let it draft them, and have a person approve every firm notice. The agent should never threaten legal action, calculate interest or mention collections unless you supplied the exact wording. See the guardrails rule sheet.

Does debt collection law apply to my invoice reminders?

The Federal Trade Commission says the Fair Debt Collection Practices Act covers household debts such as credit cards and medical bills, and that business debts are not covered. Rules can differ if you bill consumers or use a collection agency, and by state. Ask a lawyer before you automate final notices.

What should the agent do if a client disputes an invoice?

Stop all messages to that client, tag the account as a dispute and alert you with a short summary. A person fixes the invoice or replies. Restart the sequence only after you approve it.

How do I automate invoicing, not just chasing?

Creating the invoice is a separate step. Start with business process automation to see which steps from quote to payment can be linked, and use your accounting tool's recurring invoices where they fit.

Do I need an AI agent, or are built-in reminders enough?

If most late payers are simply forgetful, built-in reminders are enough. An agent pays off when you want messages that sound like you, replies sorted for you, or reminders that adapt to each client. Try the built-in option for a month first.

More in this topic: Problem solving: AI strategy and adoption (24 guides) →

Sources

Ready for the next step?

A 30 minute call. If you do not need me, I will tell you.

Ranjit Bhinge You talk to me, Ranjit. I reply within a few hours.

Not sure what you need? Tell me what is going on. I will point you the right way, even if it is not me.

Blur Studio · 30 minute call

Book a 30 minute call

You talk to me, Ranjit. If you do not need me, I will tell you.

  • We talk about your business and what is going on.
  • You leave with a clear next step, even if it is not me.
  • I reply to every message within a few hours.
Loading the calendar…