Late payment is the quiet tax on a small business. You did the work, you sent the invoice, and now you are spending Friday afternoon writing awkward emails to people you actually like. The money exists — it is just sitting in someone else's bank account, earning them interest instead of you.
Most advice on this is useless. "Automate your reminders" is not a strategy, it is a setting. What follows is the system: how to use AI to see which debts matter, write the sequence that recovers them, and keep the relationship intact — plus where letting AI near a debtor is a genuinely bad idea.
Why late payment is a process problem, not a chasing problem
Owners tend to treat collections as a willpower issue: if they were just a bit more disciplined about chasing, the money would come in. It almost never works, because the problem is structural.
Run the arithmetic on your own business. If you invoice €40,000 a month and your average invoice is settled 18 days past terms, roughly €24,000 of your cash is permanently parked elsewhere — a standing overdraft you grant your customers for free, renewed monthly at your own expense.
Three things cause it, and only two of them are fixable with AI:
- No visibility. You know "a few invoices are late." You do not know that six accounts hold 70% of the exposure, or which three customers have been late every single quarter for two years.
- No ladder. There is no agreed sequence of what gets sent on day 3, day 14, day 30 and day 60 — so chasing depends on whoever remembers, in whatever mood they are in.
- Weak terms. Thirty-day terms, no deposit, no late-payment clause, no card option. AI cannot fix this one. Your contract can.
AI is good at the first two: it turns a messy ledger into a ranked list, and an awkward email into a 40-second task. Both are the jobs that get dropped when you are busy.
Step 1: Build your ageing and risk picture
Export your open invoices from Xero, QuickBooks or Sage. You want a CSV with six columns: customer, invoice number, issue date, due date, amount, days overdue — plus date of last contact if your system tracks it.
Then hand it over with a triage prompt rather than a vague request for "analysis":
Here is my open invoice ledger as a CSV. Group the overdue invoices into 1–30, 31–60, 61–90 and 90+ day bands, with the total value of each band. Then produce a chase list of the top 10 invoices ranked by amount multiplied by days overdue. Flag any customer appearing more than twice, and tell me what percentage of my total overdue value sits with my three largest debtors. Output as two tables. No commentary.
The output changes how you spend the next hour. Most owners chase chronologically, oldest first. Ranking by exposure usually reveals that a handful of accounts hold most of the money, while the long tail of €200 invoices consumes most of the effort for almost none of the cash.
Two patterns change your commercial decisions rather than just your admin, so ask about them explicitly: repeat offenders (a customer who is late every time is telling you to change their terms, not to send another reminder) and concentration risk (if one client holds 45% of your receivables, that is a business risk, not a collections task). Feed the same export into your forecast to see the cash impact — our guide to using AI for cash flow forecasting covers the modelling.
Do this monthly, on the same day. The discipline matters more than the sophistication.
Step 2: Build a four-stage escalation ladder
Write the ladder once, in your own voice, with AI's help — then stop improvising. A sequence that works for most B2B small businesses:
Stage 1 — Due date + 3 days: the assumed-error nudge
Warm, short, no accusation. Assume it is sitting in an approvals queue or went to the wrong inbox. Attach the invoice copy and ask one question: "Can you confirm it is scheduled?" This single email recovers a surprising share of overdue invoices, because most late payment is genuinely administrative.
Stage 2 — Day 14: the firm request for a date
Now you need a named person and a commitment. Shift the ask from "please pay" to "which date will it be paid?" A date is something you can hold someone to; "soon" is not. Keep it to five lines.
Stage 3 — Day 30: escalate the recipient, not the volume
Same calm tone, different person — the finance lead, the director, the account owner. State the agreed terms, restate the amount, note that statutory late-payment interest is available to you, and offer a payment plan if you would accept one. Escalating who reads it beats escalating how annoyed you sound.
Stage 4 — Day 60: the formal step
A letter before action, pausing work, or passing it to a collections agency or solicitor. A human writes and sends this one. Have the template reviewed by a lawyer once, then reuse it.
AI's job here is to draft all four messages in your voice as reusable templates with variables ({first name}, {invoice number}, {amount}, {days overdue}). Its job is not to decide when to escalate — that depends on the relationship, the pipeline and the size of the debt, which only you can weigh.
The prompts that actually work
The difference between a usable draft and something you rewrite entirely is the constraints you put in the prompt. Two worth saving.
The reminder, in your voice. Paste three of your own best past emails first, then:
Above are three emails I have written, so you can match my tone. Now draft a stage-2 payment reminder for invoice {number}, €{amount}, {days} days overdue, to {first name} at {company}. Constraints: British English, under 120 words, warm but direct, no guilt-tripping, no exclamation marks, no "I hope this finds you well." End with a single question asking which date it will be paid.
The difficult reply. When a client says they cannot pay yet, this is where relationships are won or lost:
A client has replied saying they are having cash flow problems and cannot pay a €6,400 invoice that is 38 days overdue. Draft a reply offering three options: payment in full within 7 days, two instalments over 30 days, or four over 60 days with work paused until the first clears. Tone: collaborative, not threatening. Make clear which option I prefer. Under 150 words.
Keep these as saved prompts or a project in whichever assistant you use, rather than retyping them. If your reminders already run out of your accounting software, our guide to using AI for invoicing and billing covers the automated side; this post is about the judgement calls that sit on top of it.
The guardrails that keep you out of trouble
Collections is one of the few AI use cases where a bad output has legal consequences. Four rules.
Never send stage 3 or 4 unread. A sentence implying legal action you do not intend to take, or wording that reads as a threat, can create real exposure — and consumer debt collection in particular is tightly regulated in most countries. Human eyes on anything beyond a polite reminder.
Never let AI invent legal figures. Models will cheerfully produce a plausible-looking interest rate or statutory reference. Look up the entitlement that applies to you, then give the model the numbers. In the UK, the Late Payment of Commercial Debts (Interest) Act 1998 provides statutory interest at 8% above the Bank of England base rate plus fixed compensation; across the EU, Directive 2011/7/EU sets at least 8 percentage points above the relevant reference rate plus a minimum of €40. Verify the current figures for your jurisdiction before they go in a letter — and treat this as general information, not legal advice.
Mind the personal data. A debtor ledger containing named contacts is personal data under the GDPR. Use a business or team tier whose terms confirm your inputs are not used for training, strip details you do not need, and never paste a full customer ledger into a free consumer chatbot.
Do not let AI decide who is creditworthy. If your customers are individuals or sole traders, this is regulated territory: the EU AI Act treats AI used to evaluate the creditworthiness of natural persons as high-risk, and the GDPR gives people the right not to be subject to purely automated decisions with significant effects. Use AI to summarise payment history. Let a person make the call.
Stop late payment before it starts
The highest-return hour here is not spent chasing at all. Collections is downstream of terms.
- Shorten terms. Fourteen days instead of thirty, on new contracts. Nobody pays faster than the number on the invoice.
- Take deposits. Forty to fifty per cent up front on project work removes most of your downside.
- Make paying frictionless. A payment link on the invoice, card accepted, direct debit for retainers. Every extra step is a day of delay.
- Add a late-payment clause that states your interest rate and recovery costs, so stage 3 is enforcing an agreement rather than making a new demand.
- Credit-check new accounts above a threshold you set, say €5,000.
AI is useful here as a reviewer. Paste your standard terms and ask it to list every payment-related clause you are missing compared with normal practice for your sector, then have your solicitor approve the additions. Our guide to AI clauses for client contracts covers the wider contract review workflow.
What good looks like after 90 days
Three months in, a working collections system looks boring. That is the point:
- Your average days-to-payment has dropped by 10–15 days, mostly from stage 1 going out reliably rather than from any clever wording.
- No reminder is written from scratch. Everything comes from a template with variables filled in.
- Chasing takes under 30 minutes a week, on a fixed slot in the calendar, not scattered through your Fridays.
- Two or three chronically late customers have been moved onto deposits, shorter terms, or off your books entirely.
Note what is not on that list: fully automated collections with no human involved. That is not the win available to a small business, and chasing it is how people end up sending a legally unwise email to their largest client at 2am.
The real return is quieter. As one owner put it, AI frees up time for relationships, not just reporting — and collections is exactly that trade. You are not replacing the conversation with a machine. You are trying to stop dreading it, so the awkward email actually goes out on day three, when it still works.
Where does your business stand on AI?
Take the free 3-minute AI Readiness Quiz and get a personalised score with your next steps.
Take the Free Quiz →