Automation takes the chasing sequence you already run and runs it to the day, so whatever your sequence was doing, you now get more of it. Four patterns come up often enough to describe: a sequence nobody had reviewed, frequency treated as pressure, a ledger-wide finding applied to one of your customers, and prediction bought where consistency was the constraint.
In this article
The short version
- Independent research puts the value of a reminder at roughly 25 percent against sending nothing, and shifting its timing inside a three-week window changed how fast people paid without changing whether they paid.
- Section 40(1)(a) of the Administration of Justice Act 1970 keys the offence of harassing a debtor to the frequency of demands, and frequency is the setting automation makes free. Section 40(3) gives a creditor acting reasonably a broad defence.
- A pattern across your whole ledger often doesn't reproduce inside one customer's record, and the two get quoted as if they were one claim.
- Peer-reviewed work shows machine learning can predict payment dates. It tests no intervention, so it says nothing about whether acting on the forecast changes behaviour.
- Nothing published shows AI-written or AI-timed chasing collecting faster than a well-configured schedule, and that includes anything we could publish.
Automating a sequence nobody had reviewed
Your chasing sequence was probably written once, by whoever had the ledger at the time, and then inherited. Switching it on properly is a genuine gain in consistency, because a message that used to go out when somebody remembered now goes out on the day. Your messages, their order and their wording tend to travel across unexamined.
Independent evidence says the consistency is worth having. In a field experiment on Australian business tax debts, a reminder letter raised the probability of payment by roughly 25 percent against sending nothing, and varying its timing inside a three-week window made no difference to whether payment arrived within seven weeks of the due date, though earlier letters did prompt faster payment. Those are tax debts owed to a revenue authority, sent as letters, in 2018, so treat it as the closest independent evidence available and not as a study of your invoices.
Read together, they point at your gaps before your calendar. A customer who never got a first reminder is worth more of your attention than a second reminder moving from Tuesday to Monday.
Your due dates deserve the same look. Of customer-initiated payments in Adfin's own book, 22.0% land on the due date itself, and among those paid on time, 33.5% arrive on that day and 40.5% in the final two days (Adfin platform data, 59,777 paid invoices). The book leans towards small accountancy practices billing small recurring fees, so read that as these businesses and not as UK businesses. Either way the date you print carries a lot before any message goes anywhere, and how to chase an unpaid invoice covers the sequence around it.
Frequency is the setting automation turns
One piece of law has always applied to chasing, and it's about cadence instead of authorship. Section 40(1)(a) of the Administration of Justice Act 1970 makes it an offence, where the object is to coerce somebody into paying a contractual debt, to harass them with demands for payment which, "in respect of their frequency or the manner or occasion of making any such demand, or of any threat or publicity by which any demand is accompanied, are calculated to subject him or members of his family or household to alarm, distress or humiliation".
Frequency comes first in that sentence, and frequency is what automation makes free. Chasing by hand, you're rationed by your own week, so the fourth message tends not to happen. A scheduler has no such limit, and nor does a second channel running alongside the first.
The defence is broad, and section 40(3) states it. Subsection (1)(a) "does not apply to anything done by a person which is reasonable (and otherwise permissible in law) for the purpose" of securing the discharge of an obligation due to them, protecting themselves from future loss, or enforcing a liability by legal process. If you're pursuing a genuine debt proportionately, you have a plain answer.
Two things narrow the section further. It speaks of alarm, distress or humiliation to a person "or members of his family or household", so it reads as aimed at individual debtors and not at companies. And we found no prosecution, case law or regulatory action applying it to automated or AI-generated chasing, so this isn't an enforcement story. Our reading of the wording, labelled as a reading: proportionate business chasing is a long way from the offence, while high-frequency automated chasing of sole traders and individuals is where the wording starts to bear on what you've configured.
What follows from that is practical. A cap on contact per customer per period belongs in your software instead of in your policy document, because a document doesn't send anything. Count email, SMS and WhatsApp together too, since three channels each running their own polite schedule add up to a cadence you didn't choose. How often should you chase an unpaid invoice sets out reasonable intervals.
A ledger-wide pattern and one customer's record
Two questions get collapsed into one whenever chasing and data meet on the same page, and keeping them apart is most of the discipline here.
One is about everybody. Is there a best channel, or a best hour, across all businesses and all customers? No reliable public evidence settles it, and any confident universal answer is a guess, ours as much as anybody's. The other is about one of your accounts: does this client reply to a text and leave your email unopened? Their own record answers that, from what they've replied to before and how quickly they paid.
Our own data shows why one doesn't substitute for the other. Pooled across Adfin's book, on-time payment falls from 69.6% on invoices under £100 to 60.1% on invoices of £2,500 and over (Adfin platform data, 59,736 paid customer-initiated invoices), and you could read that as an instruction to chase your large invoices harder. Inside a single business's own ledger the effect nearly vanishes: 39 businesses ran later on their large invoices and 43 ran later on their small ones (Adfin platform data). The pooled gradient describes which businesses issue which invoices more than it describes any one of your customers.
The same caution applies to a league table of payment methods or send times. Hold one business constant and the differences between methods largely disappear, so a table like that is mostly about who uses what. A rule you take from the aggregate can be wrong for the account in front of you, and how to set different chasing rules for different clients works from the record instead.
Prediction, where consistency was the constraint
Prediction is the part of this market with real research behind it. In a peer-reviewed study of one German DAX-40 company's ledger, covering more than a million invoices from 2017 to 2019, neural networks were found suitable for predicting customers' payment dates and beat conventional baselines such as linear and multivariate regression. One firm, one ledger, and a forecasting study: it tests no intervention, so it can't tell you that chasing differently on the strength of a forecast collects anything sooner.
Notice what carries the AI badge and what doesn't. HighRadius names its methods on its collections page, listing "Predict Payment Date Random Forest & Gradient Tree" and "Customer Segmentation K-Means ML" alongside language models for its inbox and co-pilot. Prediction and segmentation there are machine learning. The email leaving on day seven is a rules engine, in that product and in most others, and both get sold under one word.
Ranking earns its place when your attention is what's scarce. Among repeat payers in Adfin's book, 27.4% were never late once, while the 24.3% who are late at least 80% of the time produced 51.9% of all late payments (Adfin platform data, 1,359 business-customer pairs with five or more payments). Those pairs qualified by paying five times, so the group selects for reliable payers and you're reading the shape and not the exact proportions. A short list is easier to work through than a full ledger when you've got an hour.
A score can't repair a sequence that stops running in a busy week. If your chasing slips when the month closes, a better ordering of the same neglected list doesn't recover it, and consistency is the cheaper thing to buy first.
The records question that tends to get skipped
Automated chasing generates a record of demands, and that record does two jobs for you. Evidentially, it proves what you demanded and when, if an invoice ever reaches a statutory demand or a money claim. In data protection terms, it's what lets you answer a question about a decision months later.
One exemption gets read generously. Article 30(5) of the UK GDPR lifts the record-of-processing duty from organisations employing fewer than 250 people, but not where "the processing is not occasional". Credit control runs continuously against most of your customer list, so our reading is that a small firm running a chasing platform is unlikely to be inside that exemption.
The ICO's own screening list for a data protection impact assessment also names innovative technology, including AI, and decisions about access to products or services taken wholly or partly automatically. That list points at your automated escalation before it points at your reminders. Is it safe to let AI email your clients? works through the whole position, including where it's genuinely unsettled.
Keep the trail as something you can read: what went out, to whom, on which trigger, under which template, and who approved anything out of the ordinary.
Where the routine part can go
The conversations that decide anything need you. A client explaining they're in difficulty, a negotiation over instalments, an account you'd rather keep than collect from this month: your judgement is what's valuable there, and no schedule substitutes for it.
Most of your chasing isn't that. Most of it is the same confirmation, the same reminder and the same statement against a different name, in a channel that customer reads, at a moment they can act on. Handing that over is a scheduling job, and it competes with everything else in your week, so it's usually the first thing to slip.
Delegating it well means setting the rules of engagement once: which channels are allowed, how persistent the sequence may be, what can go out without you seeing it, and where you want to review. Adfin's Customer Agents work a ledger inside boundaries like those, from your own email domain, alongside late fees, instalment plans, payment retries and a statement when an account goes quiet.
The claim that holds up is a capability claim. An agent can hold one customer's pattern in a way you can for ten accounts and not for four hundred. Whether acting on that pattern collects money sooner hasn't been measured by anybody, us included, so a vendor quoting you a figure is ahead of the evidence. What is agentic credit control? draws the line between a schedule and an agent properly.
Common questions
What changes when you automate invoice chasing? Consistency, mostly. Messages go out on the day they're due instead of when somebody remembers, and your frequency stops being limited by your own week. The sequence, the wording and the due dates travel across unchanged, so they're all worth a look before you switch anything on.
Can automated payment reminders count as harassment? Section 40(1)(a) of the Administration of Justice Act 1970 keys the offence to the frequency or manner of demands, and section 40(3) gives a broad defence to a creditor acting reasonably to recover a genuine debt. Our reading is that proportionate business chasing is a long way from it, and that very frequent automated chasing of a sole trader is where the wording starts to matter. We found no case applying it to automated chasing.
Does AI know the best time to send a payment reminder? Keep two questions apart. Across all businesses, no reliable evidence settles a best hour or channel. For one of your clients, their own history of replies and payments is a real answer, and that narrower claim is the one worth making.
Is a payment prediction model worth having? It helps when your attention is what's short, because a ranked list is easier to work through than a full ledger. Peer-reviewed work shows machine learning can predict payment dates on one large firm's ledger, but it tested no intervention, so a forecast isn't evidence that chasing differently collects sooner.
How many reminders is too many? There's no statutory number. What the 1970 Act polices is frequency and manner, so in practice you want a cap per customer per period, set in your software and counted across every channel together instead of per channel.
What still needs a person once chasing is automated? The difficult conversation, the decision to escalate or wait, anything asserting a legal consequence or a specific interest figure, anything going to a customer who has raised a dispute, and the first send after you change a template.
Sources
- legislation.gov.uk — Administration of Justice Act 1970, section 40 (accurate as of August 2026)
- legislation.gov.uk — UK GDPR, Article 30 (accurate as of August 2026)
- ICO — when do we need to do a DPIA? (accurate as of August 2026)
- Gillitzer and Sinning, ANU — nudging businesses to pay their taxes, does timing matter? (accurate as of August 2026)
- Kureljusic and Metz, Journal of Applied Accounting Research — predicting customer payment dates (accurate as of August 2026)
This article covers data protection and professional conduct information and is not legal advice. Our reading of the Administration of Justice Act 1970 and of the UK GDPR record-keeping provisions is labelled as ours in the text. If you're deciding how far to automate escalation against an individual or a sole trader, that's a decision worth taking advice on. Last updated August 2026.
