AI in credit control does a small number of things well today. It ranks your ledger by who is likely to pay late, drafts your messages, works out when each one goes out, triages the replies, and hands you anything needing a decision. The judgement stays with you, and the distance between what has been demonstrated and what gets marketed is wide.
In this article
The short version
- Prediction and segmentation are genuine machine learning. The email going out on day seven is usually a rules engine carrying the same badge.
- Independent research supports two things: a reminder changes payment behaviour a lot, and a model can predict when a customer will pay. Nothing published shows AI-written or AI-timed chasing collecting faster than a competent schedule, so every confident percentage here is vendor marketing.
- Article 22 UK GDPR no longer exists in the form most articles quote. Articles 22A to 22D replaced it, fully in force 5 February 2026, and the model moved from prohibition to permission with safeguards.
- Among repeat payers in Adfin's own book, the 24.3% who are late at least 80% of the time produce 51.9% of all late payments (Adfin platform data, 1,359 business-customer pairs).
What still needs a person
Start with the parts that don't belong to software, because everything after this is easier to weigh once they're on the table. The call where your client explains they're in difficulty needs you, and so does the decision that follows it. Whether to extend terms, whether to pause work, whether a disputed invoice was right in the first place: those are commercial judgements about your business, and a model has no view worth having on them.
The professional standard points the same way. The PCRT bodies' January 2026 guidance on the ethical use of AI tools says AI output "should also be regarded as if it were prepared by a less experienced junior colleague and reviewed with appropriate scepticism". Difficult payment conversations covers the calls that are yours, and what's left once you've set them aside is a lot of repetition.
Three different things get sold as AI
Three different mechanisms share one word here, and telling them apart makes every claim you read easier to judge.
| Mechanism | What it actually does | Typical wording |
|---|---|---|
| Statistical prediction | Learns from historical payment behaviour to rank or score your customers and invoices | "payment prediction", "risk brackets", "customer segmentation" |
| Generative drafting | A language model writes or rewrites a message, or reads an incoming one | "AI email generator", "drafts replies", "adapts your tone" |
| Rules and triggers | Sends message two on day seven unless a payment lands first | "automated reminders", "collection paths", "smart templates" |
The enterprise vendors name their algorithms: HighRadius publishes "Predict Payment Date Random Forest & Gradient Tree" and "Customer Segmentation K-Means ML", and Kolleno names Microsoft Azure OpenAI and OpenAI's GPT models. Chaser's late payment predictor sorts your due invoices into risk brackets from low to high, using the due date, the value and past behaviour.
At the other end, Trove's claim that its AI will "vary sentences using AI to avoid identical weekly emails" describes a system choosing between strings somebody wrote earlier. Paidnice publishes no AI claim at all on its reminder pages, and neither does Satago. On the two accounting platforms most UK small businesses use, the reminders are rules: Xero's run a fixed schedule, and its AI product makes no claim to chase anybody. Best AI credit control software in 2026 goes vendor by vendor on which is which.
Prediction and ranking, before anything goes out
The best-evidenced use of machine learning in receivables happens before a single message is written. Kureljusic and Metz tested a range of algorithms on real data from one German DAX-40 company, more than a million invoices covering 2017 to 2019, and found neural networks in particular suitable for predicting customers' payment dates, beating linear and multivariate regression. The paper forecasts and doesn't intervene: nothing in it says acting on the forecast changes anybody's behaviour, so anyone citing it as proof that AI collects faster has read it too quickly.
Ranking earns its place for a simpler reason: your lateness almost certainly isn't spread evenly. Take every business-customer pair in Adfin's book with at least five payments the customer made themselves, 1,359 pairs across 12,863 payments, and 27.4% were never late once while the worst 24.3% produced 51.9% of every late payment in the group (Adfin platform data). That book is weighted towards small accountancy practices billing small recurring fees, so test the shape against your own ledger. On a two hundred client list, roughly fifty names generate half your chasing work. Building that ranking by hand is a one-off; keeping it current is where a model earns its keep, and what credit control is actually deciding works through the concentration.
Drafting the message
A language model writes a competent, courteous reminder, and it writes your fortieth as carefully as your first. Given the invoice, the history and your tone, it produces something that reads as though somebody at your firm wrote it, sent from your own domain.
Two things deserve your attention before you switch drafting on. A model asked to be warmer will happily add a sentence promoting your services, and on the ICO's own wording that changes the legal character of the message: "if your service message has elements that are direct marketing, even if that is not the main purpose of your message, then it will count as direct marketing". A neutral reminder is a service message and PECR's consent rules don't touch it. Put a cross-sell inside one and it has moved category, and PECR's email rule does reach your sole trader and partnership customers.
The second is arithmetic. Any message asserting an interest figure, a fixed compensation sum or a contractual fee is making a legal statement, and a wrong one costs you more than a vague one would, so those are your clearest candidates for review. Should invoice reminders come from a person or an AI? takes the authorship question on its own, and can AI chase invoices for you? draws the line between the parts it drafts and the parts you decide.
Timing and channel
Keep two questions apart here, because most of the marketing in this category collapses them into one.
Is there a best channel, or a best hour, across all businesses and all customers? No reliable public evidence settles that, Adfin isn't in a position to publish a finding either, and any confident universal answer is a guess.
Does this particular client of yours reply to a text and ignore email? Their own record answers that: the channels they've responded on, how quickly payment followed, which invoices they queried. A claim about one customer needs no general rule behind it.
The independent evidence is about the reminder and not its optimisation. Gillitzer and Sinning ran a field experiment on Australian business tax debts and found that a reminder letter raised the probability of payment by roughly 25 percent against no letter at all, while moving the letter within a three-week window changed how quickly people paid without changing whether they paid, in work published by the Tax and Transfer Policy Institute at ANU. Tax debts and letters aren't invoices and email, so read that as a direction and not a transferable number.
Related, from Adfin's own data: across customer-initiated payments, 22.0% land on the due date itself and the median on-time payment arrives three days early (Adfin platform data). Your customers work to your due date, so the date you choose does more for you than the hour you send, and how AI decides when to chase covers the mechanics.
Query triage and the inbound side
Chasing generates replies, and your replies are where the time goes. Somebody promises to pay on Friday, somebody disputes a line, and each wants a different next step from you. Language models are genuinely good at sorting them, and vendors describe this one accurately for once: Chaser's inbound tool "reads each debtor's message, detects intent (promise-to-pay, dispute, document request, etc.)" and drafts a response. Sorting an inbox by intent is safer than composing an outbound demand, because a misread reply gives you a bad draft you can see instead of a message already sent.
Adfin's Customer Agents handle query triage in the same spirit, and can send a statement covering everything a customer owes when an account goes quiet. What matters in the design is that a promise to pay pauses your sequence and a dispute stops it, because a reminder arriving after a query does more damage than the one you never sent.
Escalation, and where a person signs
Escalation is where automation should get slower instead of faster, for legal reasons as much as commercial. A reminder doesn't change your customer's position: the debt, the due date and the statutory entitlements exist because of your contract and the Late Payment of Commercial Debts (Interest) Act 1998, whether anybody emails anybody or not. An account stop, a fee or an agency referral does change it, and where your customer is a sole trader, that starts to look like the automated outcomes data protection law is interested in.
Frequency deserves separate attention, because automation makes it free. Section 40(1)(a) of the Administration of Justice Act 1970 makes it an offence to harass a debtor with demands for payment which, "in respect of their frequency or the manner or occasion of making any such demand", are calculated to subject them to alarm, distress or humiliation. Section 40(3) gives a creditor acting reasonably a broad defence, and our reading is that proportionate business chasing is a long way from the offence. High-frequency automated chasing of sole traders is closer to it than most people running a sequence have considered, so cap the contact rate in your software.
So draw the boundary by consequence. A first polite reminder on a small, undisputed, in-terms invoice is the safest thing an automation can do unattended, and anything changing what your customer owes or what they can buy is where a person signs. What is agentic credit control? covers how those boundaries get set, and is it safe to let AI email your clients? the approval gate.
Reconciliation and the work around the edges
Chasing somebody who has already paid you is the most avoidable failure here, and it's a matching problem: payments arrive with the wrong reference or amount, and until they're matched your ledger thinks the money is missing.
Kolleno publishes "AI Powered Matching", and Adfin auto-matches payments including underpayments, overpayments, and payments arriving with no reference, syncing both ways with Xero. Failed collections are the same kind of work: a retry and a fallback to another method recover money your sequence would otherwise start asking for, and failed payments covers that side.
What the independent evidence supports
The position as at August 2026, after a deliberate search. Independent evidence says a reminder works, from Gillitzer and Sinning, and peer-reviewed evidence says machine learning can predict payment dates, from Kureljusic and Metz. There's no independent published study of AI-written against human-written reminders, none on AI-chosen send timing for commercial invoices, and none on channel choice in business-to-business chasing.
The chartered body for the profession is instructive by its silence: the CICM's own article on AI and automation in credit management carries no quantified claim about effectiveness.
Every confident number here comes from a vendor. HighRadius publishes "10% Reduction In DSO", Sidetrade publishes "30–40% average DSO reduction" and "68% workflow efficiency", and none of those pages carries a methodology, a sample or a period. Intuit's payments page tells you that you'll "get paid 4 days faster on average when you send invoice reminders with Payments AI" with a superscript footnote marker, and no footnote appears on the page. What finance teams get wrong about automating collections picks up what follows from an empty evidence base.
Where the law stands in August 2026
Most of what you'll read about GDPR and automated chasing quotes a provision that has gone. Article 22 UK GDPR, cited everywhere for a right not to be subject to a solely automated decision, was replaced by new Articles 22A to 22D by section 80 of the Data (Use and Access) Act 2025, fully in force on 5 February 2026.
The change was structural. The old model prohibited solely automated significant decisions except in three cases; the new one permits them on any lawful basis except the recognised legitimate interests basis in Article 6(1)(ea), provided the four safeguards in Article 22C(2) are in place: information about the decision, the ability to make representations about it, the ability to obtain human intervention, and the ability to contest it. The surviving prohibitions cover special category data and Article 6(1)(ea) processing, and debt recovery isn't one of the recognised legitimate interests, so your chasing runs on ordinary legitimate interests or on contract.
Whether a payment reminder engages any of this is our reading and not a citation, because no regulator has addressed invoice chasing. Our view: a reminder is a communication and not a significant decision, since the legal effects around a late invoice come from your contract and from the 1998 Act, and those four safeguards make little sense applied to which day an email went out. In business-to-business chasing your debtor is often a company, and a company isn't a data subject, although the contact's details are personal data. Escalation stays unsettled, and two of the terms deciding it can be defined later by regulations under Article 22D.
Treat the regulator's own material carefully, because it currently disagrees with itself. The ICO's page on Article 22 and fairness still states the three repealed exceptions, while its Data (Use and Access) Act pages state the new position, and its updated guidance on automated decision-making was still in draft when this was written, after a consultation that closed on 29 May 2026. No UK rule requires you to tell a customer an email was AI-drafted, and for a practice the PCRT guidance is blunt that putting client data into publicly available AI tools "is likely to constitute a breach of client confidentiality, unless the client has consented". Is it safe to let AI email your clients? works through all of it properly.
Where the routine work can go
The stages above divide into decisions and repetition. The repetition is the same confirmation, the same reminder, the same statement, against a different name every week, and it competes with everything else you're doing. Businesses affected by late payment already spend an average of 86 hours a year on this, on the Small Business Commissioner's research.
Delegating repetition well means more than switching on a schedule. You set the rules of engagement once: which channels are allowed, how persistent the sequence may be, what can go out unreviewed, and where you want to look first. An agent then works your ledger name by name inside those boundaries, adapting to what each account has responded to before, from your own email domain. Adfin's Customer Agents do that, alongside late fees, instalment plans, payment retries, and the statement when an account goes quiet.
What this doesn't do is make anybody pay faster on the strength of the technology. The capability is real and the measured performance isn't, so a vendor telling you otherwise is ahead of the evidence, ourselves included too. You want chasing to feel personal but have no time takes the same problem from the other end.
Common questions
What does AI actually do in credit control? Four things, mostly: it predicts which of your invoices are likely to be paid late, drafts the messages, works out when each one goes out, and sorts the replies by what they mean. Prediction and drafting carry real machine learning; the sequence itself is usually rules.
Is AI chasing legal under UK data protection law? Chasing an unpaid invoice is ordinary processing on legitimate interests or on contract, and drafting the message with software doesn't change that. Escalation is the area to look at: an automated fee, account stop or agency referral against a sole trader is closer to the decisions Article 22C was written for. No regulator has ruled on invoice chasing, so take advice on your own design.
Does AI chasing get invoices paid faster? No independent published study shows AI-written or AI-timed chasing collecting faster than a well-configured reminder schedule. Independent research does show that a reminder itself changes payment behaviour substantially, and that a model can predict when a customer will pay. Every specific speed claim here is vendor marketing.
Do you have to tell customers an email was written by AI? No UK legal duty requires it, and the EU AI Act's disclosure obligation applies to providers of AI systems in the EU, so a UK business chasing UK customers falls outside it. Practices have a professional overlay: the PCRT bodies recommend a statement in the engagement letter about potential AI use.
What's the difference between automated reminders and AI chasing? Automated reminders run a schedule you wrote: message two on day seven, the same for everybody. AI chasing decides within boundaries you set, varying the timing, channel and wording by what each customer has done before. The first is easy to audit; the second is newer and less proven.
Which parts of credit control should stay with a person? The call where your client explains they're in trouble, the decision to escalate or wait, any message asserting a legal consequence or an interest figure, and anything to a customer who has raised a dispute or been flagged as vulnerable.
Sources
- legislation.gov.uk — Administration of Justice Act 1970, section 40 (accurate as of August 2026)
- legislation.gov.uk — Data (Use and Access) Act 2025, section 80 (accurate as of August 2026)
- legislation.gov.uk — Late Payment of Commercial Debts (Interest) Act 1998 (accurate as of August 2026)
- legislation.gov.uk — UK GDPR, Article 22A (accurate as of August 2026)
- legislation.gov.uk — UK GDPR, Article 22C (accurate as of August 2026)
- ICO — direct marketing guidance, identifying direct marketing (accurate as of August 2026)
- Small Business Commissioner — late payments research (accurate as of August 2026)
- Tax and Transfer Policy Institute, ANU — nudging businesses to pay their taxes, does timing matter (accurate as of August 2026)
- Journal of Applied Accounting Research — machine learning for payment date prediction (accurate as of August 2026)
Reviewed by the Adfin team. This summarises UK data protection law and professional guidance on AI as at August 2026 for information, and it isn't legal advice. The Information Commissioner's Office had not published final guidance on automated decision-making when this was written, and no regulator has addressed invoice chasing, so the readings marked as ours are reasoning from the legislation and not settled positions. If you're deciding whether to let software apply fees, stop accounts or refer debts without a person approving each one, take advice on your own facts, and if you belong to a professional body, read its guidance on AI tools too.
