This is one of the jobs AI is genuinely good at. You hand it the invoice facts copied from your books, a line or two about where you stand with the client, and two or three reminders you have sent before, then ask for the next message in a sequence you decided in advance. What comes back sounds like you, and you send it. The email is rarely better than one you would write on a good day. The difference is that it exists on Tuesday, which is when the reminder was supposed to go out.

Very little money goes uncollected because a reminder was badly worded. It goes uncollected because the second reminder never got written. Asking someone you like, or someone you want more work from, to pay you is uncomfortable enough that the task keeps sliding down the list. A draft that is already sitting there on the right day removes most of that friction.

Write the sequence while you are calm

Decide the ladder before you need it. The version you invent at the moment you are annoyed is always worse than the one you wrote in a quiet hour. Ours has three messages and then a decision.

On the due date, a short friendly nudge that assumes the invoice was simply missed. Two weeks later, a firmer message that names the amount, says how many days late it is, and asks for a payment date. At around four weeks, a final notice: plain, calm, and clear about what happens next. After that it stops being an email and becomes a phone call, a pause on the work, or a decision to hand it to someone else.

This is practitioner common sense rather than a research finding, and its value is that it takes the judgment out of the moment. With the ladder written down you are not deciding whether today is the day to sound firm. You are sending the second message because it is day fourteen.

Exhibit 1

Each rung has one job, and a way of failing.

The reminder ladder, and the tone each rung needs On the due date: the friendly nudge Assume it was missed. Short, and no apology for asking. Two weeks late: the firmer ask Name the amount and the days overdue. Ask for a date. Four weeks late: the final notice Calm and plain. Only consequences you will carry out. After that: it stops being an email A phone call, a pause on the work, or handing it on.
Note: this is the sequence we use, not a rule from anywhere official. Move the timings to fit your terms and your trade, then keep them fixed so the dates decide for you.

Tone is the whole job

Each rung fails in its own way. The nudge fails by apologizing: "so sorry to bother you about this" tells the reader your deadline is soft and buys another fortnight of silence. The firmer message fails by turning passive, and "I assume this must have slipped your mind again" lands far worse than saying the invoice is fourteen days late. The final notice fails when it threatens something you have no intention of doing. If you will not put the debt with a collections agency, do not mention one.

A general assistant is useful here precisely because it will produce the same message at three levels of firmness in a few seconds. Read them next to each other and the choice is obvious. Picking a tone is much easier than producing one while irritated, and that is most of what politeness means in this context.

Which tool, and where each one stops

Two classes are in play. Your accounting software almost certainly sends automatic reminders, and for the first rung it is the right tool. It fires on the due date whether or not you feel like it, and the message needs no judgment, because at that point there is nothing to judge. If that setting is off, turn it on today.

From the second rung on, an automatic reminder is the wrong instrument. It cannot know that this client has paid on time for three years, or that there is an unresolved query sitting on the second line of the invoice. That is where a general assistant you paste into does the work: you supply the context, it supplies the wording, you press send. The shape is the same as the rest of the everyday tasks AI can take over.

What to feed it

Five things, and gathering them takes a couple of minutes. The invoice facts: number, amount, what the work was, the date you issued it, the date it fell due, and how many days late it now is. The relationship: how this client has paid before, what work is still in flight with them, whether you want more of it. Anything already said about this particular bill, including a payment date they promised and missed. Two or three of your own past reminders, which teach your tone better than any adjective you could type. And which rung you are on.

Copy the numbers in. Do not ask for them. The amount, the invoice number and the days overdue come out of your own bookkeeping, and a model asked to work them out will produce something plausible instead. A wrong figure in a chasing email hands the client a reason to reply about your mistake rather than their payment, and that is another fortnight gone.

Exhibit 2

The context comes from you, and so do the numbers.

Hand it these The invoice facts What the work was How they have paid before What has been said already Your own past reminders Which rung you are on Never from the tool The amount and number Dates and days overdue Late fees or interest The decision to send
Note: both lists describe the setup in this article. The right-hand list is short on purpose, because every item on it is a way of losing a client you were only trying to invoice.

A prompt you can adapt

Swap your own facts in and it reads roughly like this. Draft the second reminder in my sequence, in my voice, matching the three past emails below. Invoice 1043, $2,400, issued June 3, due July 3, now fourteen days late, for the work described in it as a full kitchen rewire. This client has paid inside a week for two years, so treat this as unusual rather than a habit. They did not reply to the first reminder. Keep it under 120 words, ask for a specific payment date, and do not apologize for asking. Do not mention a late fee. Do not state anything that is not in what I have given you. Give me a warm version and a firmer version so I can pick.

Two of those lines earn their place immediately. The word limit stops the drift into a formal letter that reads as though a lawyer drafted it, which is the register these tools reach for when the subject is money. Asking for two versions turns a writing job into a choice, and a choice takes ten seconds.

Check before it goes out

Start with the one that costs most to get wrong: has it already been paid? Look at the bank, not at your memory. Chasing money that landed on Friday does more damage than the delay you were worried about. Then check the amount, the invoice number and the dates against the record you copied from, and the client's name and its spelling, and whether the person you are writing to is the one who actually pays.

Then look for anything the draft added on its own. A late fee, an interest rate, a reference to your terms, a hint about what you might do next. If it is not in the contract you signed, it does not go in the email. Read the result aloud last. If you would not be comfortable saying it with the client in the room, shorten it rather than softening it, because short and factual reads as polite while long and hedged reads as nervous.

You press send, every time. Do not connect this to anything that mails on its own schedule. An escalation that fires while you are unaware of a phone call from the client is a fast way to lose one, and it is the same approval step you would keep if AI were drafting your routine replies.

What changes after a month or two

Not much in the accounts on week one, and anyone promising otherwise is selling something. What you notice first is that the awkward email stops taking up room in your head. The nudge goes out on the day. The firmer one follows two weeks later without a private argument about whether it is too soon. Fewer invoices drift quietly past ninety days, because most of them were unpaid for the least interesting reason there is: nobody asked twice.

Fitting the sequence to the way your own business bills, and deciding which rungs the software should handle by itself, is an afternoon of work rather than a project. It is the sort of small, unglamorous setup that AI enablement is mostly made of, and the test of it is whether the day-fourteen email goes out on day fourteen.

Start a conversation

All notes Start a conversation