How to use AI for invoice chasing and credit control

AI can automate most of invoice chasing. It drafts polite, firm reminder emails in your own tone, escalates them on a schedule you set, personalises each message with the customer's payment history, and flags the accounts that need a human conversation. For a UK small business the practical setup is an escalation ladder of five or six touchpoints, from a friendly nudge before the due date through to a formal letter quoting statutory interest, with AI writing and sending the early steps and a person handling disputes, payment plans and anything legal. Done well, it recovers the hours your team spends chasing, gets invoices paid sooner, and does it without souring customer relationships.

"Most late payment is disorganisation, not malice. A reminder ladder that never forgets and never gets embarrassed beats the most determined credit controller, because it actually sends the emails."

Dean Cookson, founder, Operosus

Why is invoice chasing worth automating?

Because the manual version is expensive in ways that never show up on an invoice. Research commissioned by the Department for Business and Trade and the Office of the Small Business Commissioner found that businesses affected by late payment spend an average of 86 hours of staff time per year chasing it. That is more than two working weeks of someone's time spent writing awkward emails and making awkward calls.

The same research puts the scale of the problem in plain numbers:

• UK businesses are owed an estimated £26 billion in late payments at any given time, an average of £17,000 per affected business
• Late payment costs the UK economy around £11 billion a year and closes 38 businesses every day
• Over 1.5 million businesses, 28% of the total, are affected by late payment each year

There is also a quieter cost: the chase that never happens. In most small businesses the person responsible for credit control is also responsible for three other jobs, so reminders go out late or not at all, and the oldest invoices quietly become the hardest to collect. Automation fixes the consistency problem before it fixes anything else.

Announcing the government's late payment crackdown in July 2025, Small Business Minister Gareth Thomas put it bluntly: "Too many small firms go under each year because they aren't paid on time - that is completely unacceptable."

What can AI actually do in credit control?

It helps to separate two layers, because they solve different problems.

Plain automation handles the schedule. Your accounting software (Xero, QuickBooks, Sage, FreeAgent) can already send a templated reminder when an invoice goes overdue. This is worth switching on today and costs nothing. Its weakness is that it sends the same message to everyone: the customer who has paid on time for five years gets the same template as the one who is 40 days late for the third quarter running.

AI adds judgement to that schedule. Specifically, it can:

• Draft each reminder in your voice, referencing the actual invoice, the work it covers and the customer's payment history, so messages read like a person wrote them
• Adjust tone by stage: warm before the due date, neutral at seven days, firm at twenty-one
• Segment customers, so a first-time late payer gets a softer sequence than a repeat offender
• Summarise an account's full history in seconds when a human does need to pick up the phone
• Spot reply intent: distinguish "paying Friday" from "we dispute this" and route each one differently
• Flag the accounts where chasing should stop and a conversation should start

The last two matter most. The biggest failure mode in automated chasing is not tone, it is a sequence that keeps firing at someone who has already replied. AI reading and classifying replies is what makes the rest safe.

What does a good escalation ladder look like?

Write the ladder down before you automate anything. Ours typically runs in six steps:

Three design notes that make the ladder work:

1. Step 1 is the highest-value step. A pre-due reminder with a one-click payment link prevents more late payments than any chase email. It is also the easiest to automate because nobody minds receiving it.
2. The approval gate moves down over time. Start with a human approving everything from step 3 onwards. After a few weeks of editing nothing, move the gate to step 4, then step 5. You earn automation, you do not assume it.
3. Any reply pauses the ladder. A dispute, a promise to pay, a question, an out-of-office from someone who has left: all of them should stop the sequence until a decision is made. Payment received should kill it instantly, which means checking the ledger immediately before every send, not once a day.

How do you get the tone right?

The awkwardness of chasing is mostly imagined, but the emails still have to be written carefully. The rules we give AI systems for credit control copy:

• Assume cock-up, not conspiracy, until step 4. Most late payment is disorganisation, not malice
• One ask per email: pay, or tell us when you will
• Make paying effortless. Payment link, bank details and invoice number in every single message
• Give an easy out: "if this has already been paid, please ignore this" costs nothing and saves face
• Never threaten anything you will not do. If you mention statutory interest at step 5, be prepared to charge it

Feed the model three or four of your real past emails as examples and tell it what to keep. Generic AI chasing copy is instantly recognisable; copy trained on your actual correspondence is not.

Should reminders come from a person or from accounts@?

Both, deliberately. Early steps should come from the name the customer already deals with, because they get opened and they keep the relationship warm. From step 4, switch the sender to accounts@ or credit.control@. It depersonalises the firmness, protects the working relationship, and signals that the matter has moved from a favour to a process. The customer's contact can then play the good cop: "accounts are chasing me about this one, can you help me out?"

When should a human take over?

Automate the routine, never the judgement. Hand the account to a person the moment any of these happen:

• The customer disputes the invoice or the quality of the work. Chasing a disputed invoice by robot is how you lose a client and the money
• They ask for a payment plan. Agreeing terms is a negotiation, not a template
• A promise to pay is broken twice. The pattern matters more than the words
• The account is strategically important. Your largest customer being 20 days late is a relationship conversation, not step 4 of a ladder
• You are approaching a letter before action. From here every word has legal weight
• You see signs of insolvency: sudden silence, staff departures, other suppliers complaining. Speed and a phone call beat any email sequence

The AI's job at handover is to make the human effective in thirty seconds: full payment history, every message sent and received, promises made and broken, all in one summary.

What is your legal backstop?

You have more leverage than most small businesses use. Under UK late payment legislation you can charge statutory interest at 8% plus the Bank of England base rate on overdue business-to-business invoices, unless your contract sets a different rate. You can also add fixed debt recovery costs: £40 on debts under £1,000, £70 from £1,000 to £9,999.99, and £100 on £10,000 or more.

Most businesses never charge it, and that is fine. Its value is in the mentioning. A step 5 email that calculates the interest accruing on a specific invoice, to the penny, changes the conversation without a single threat. This is something AI does well: the sum is mechanical, the framing is everything.

The direction of travel helps you too. The government's Time to Pay Up package, announced in July 2025, proposes a 60-day cap on payment terms and mandatory statutory interest on late payments. Expecting to be paid on time is becoming the legally reinforced default, not an awkward request.

How do you wire it together?

The architecture is simpler than the vendors make it sound. Across the systems we build at Operosus, credit control automation follows the same pattern as every other reliable business automation: a trigger from the source system, context gathered, an AI draft, a human gate where the stakes justify one, then send and log. It is the same shape we use for client onboarding automation and lead follow-up; only the trigger and the tone change.

In practice:

1. Source of truth: your accounting package. Xero, QuickBooks and Sage all expose invoice status via API or webhooks. Never maintain a separate list of who owes what
2. Orchestration: an automation layer (n8n, Make or similar) checks for overdue invoices daily and works out where each one sits on the ladder
3. Drafting: the model receives the invoice details, the customer's history and your tone rules, and produces the email for that step
4. Approval: drafts for the firmer steps land in a queue, in your inbox or a dashboard, for a one-click approve or edit
5. Sending: from your normal mailbox, not a bulk platform, so replies thread into the conversation and deliverability stays clean
6. Logging: every send, reply, pause and handover recorded against the customer

Build it fail-soft. If the ledger check fails, the API times out or the model errors, the correct behaviour is silence, never a wrong or duplicate chase. A reminder that arrives a day late is a non-event; a chase sent for an invoice paid yesterday morning costs you goodwill you cannot easily buy back.

What are the common mistakes?

• Auto-sending the angry steps. Steps 1 to 3 can run unattended. Steps 4 onwards need eyes on, at least until trust is earned
• No pause on replies. The single most damaging failure. Classify every inbound message before the next send
• Stale payment data. Check paid status at send time, not on yesterday's sync
• One sequence for everyone. Segment by history and value. Your best customers should experience your gentlest ladder
• Chasing disputed invoices. Disputes leave the ladder entirely and go to a person
• Burying the payment route. Every message carries the link, the bank details and the reference

Where to start

You do not need to build all of this in week one. The order that works:

1. Tighten the basics first. Clear payment terms on every invoice, correct contact, purchase order number where required. Automation amplifies whatever hygiene you already have
2. Switch on built-in reminders today. Xero and QuickBooks reminders are crude, but a crude reminder beats a forgotten one
3. Write your escalation ladder as a one-page policy. Steps, timings, tone, handover triggers. This document is the specification for everything that follows
4. Add AI drafting with human approval. Let the system write every chase for two or three weeks while a person presses send. Edit the drafts; the edits become your tone rules
5. Automate the early steps, keep the gate on the late ones. Most businesses end up auto-sending steps 1 to 3 and approving 4 onwards. That balance recovers nearly all the time without surrendering judgement

If you would rather have it built than build it, this is exactly the kind of system we put together at Operosus: connected to your accounting package, writing in your voice, with a human gate where it matters. If you are an accountancy practice chasing on behalf of clients, our guide to AI for accountants covers the practice-side version. Either way, start this month. The research above says the chase is costing you two working weeks a year, and that is only the time you can see.