From three things: the rules you set, a prediction of when this customer is likely to pay, and what that customer has done on your previous invoices. The first is a schedule and most products stop there. The second and third are where machine learning does real work, and they produce a claim about one customer rather than a rule for everybody.
The short version
- Timing rests on triggers you configure, a payment-date prediction, and one customer's own record of replies and payments.
- Peer-reviewed work shows machine learning can predict when a customer will pay, but it tests no intervention, so it doesn't show that acting on a forecast changes behaviour.
- Keep two questions apart. Is there a best send hour across all businesses? No reliable evidence settles it. Does your client answer a text? Their own record does.
- Independent research values the reminder itself at roughly 25 percent against none, and moving it earlier changed the speed of payment and not the rate.
- No history means no pattern, so a new client gets your default schedule.
The three inputs
| Input | What it holds | Where it comes from |
|---|---|---|
| Triggers | Due date, day count, invoice size, whether a mandate is in place | Rules you configure |
| A payment-date prediction | When this customer is likely to pay this invoice | A model reading their payment history |
| A response record | Which channel got you a reply, and how quickly payment followed | Your own previous contact with them |
Most products hold the first one only. Xero's reminders and most chasers run entirely on triggers, so if a product tells you it chases intelligently, ask which of the other two it holds. A prediction lets your ledger be ranked by expected lateness instead of by age, and a response record is specific to one of your accounts instead of being general knowledge about businesses.
Two questions, not one
Almost every confident claim about chasing timing collapses these two, so keep them apart while you read anybody's marketing, ours included.
The population question asks whether some day, hour or channel works best across all businesses and all customers. No reliable public evidence settles it, and Adfin can't publish a finding either, so any universal answer is a guess dressed as a discovery.
The per-client question asks whether one client of yours replies to a text and ignores email. Their own record answers that, from previous communications and payment behaviour, and the answer needs no general rule behind it because it isn't claiming one. It also expires, since a change of finance manager can reset the pattern, so treat it as a rolling read of the record and not a permanent profile.
Your software can hold a pattern for four hundred accounts where you can hold one for a dozen, and that's a capacity argument. Whether acting on those patterns collects more money hasn't been measured in public.
What a model learns from one ledger
The best independent evidence for prediction is a peer-reviewed study by Kureljusic and Metz. They tested several machine learning algorithms on real transaction 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.
Read what that does and doesn't establish. Payment behaviour has learnable structure on one large firm's ledger, and a model can use it to forecast a date. The study tests no intervention at all: nobody chased differently, and nothing in it shows that acting on a forecast changes when your customer pays. So it supports ranking your ledger, and not a claim that AI collects faster.
What the independent evidence says about timing
Gillitzer and Sinning ran a field experiment on Australian business tax debts. A reminder letter raised the probability of payment by roughly 25 percent compared with sending nothing, and varying the timing inside a three-week window changed how quickly people paid without changing whether they paid. Those were letters about tax debts owed to a revenue authority, so treat the direction as useful and the number as not transferable.
Both halves are worth having. Sending something matters a great deal against sending nothing, so consistency is the thing to buy. And earlier reminders brought money in sooner without bringing more of it in, which is a useful corrective to any product sold on finding your optimal send time.
What our own data shows about when payment lands
Payments cluster on the date you set. Of 59,777 paid customer-initiated payment requests over 26 months, 22.0% landed exactly on the due date (Adfin platform data). Your due date is doing more work than the reminder schedule behind it.
Lateness also concentrates. Across 1,359 business-customer pairs with five or more customer-initiated payments, 27.4% were never late once, while the 24.3% late at least 80% of the time produced 51.9% of all late payments in that group (Adfin platform data). A pair only enters that cut after five payments, so it leans towards customers nobody moved onto a mandate and the proportions won't transfer to your book.
That second figure argues for ranking your ledger instead of getting clever about hours, since half the lateness in that group came from a quarter of the payers. And we can't generalise our own record into a rule about channels or send times, so we won't tell you which performs best.
What it can't work out
A new customer has no record, so they get your default schedule. Circumstances outside your ledger stay invisible, and a client whose own customer has just failed looks like a normal payer until the day they don't pay. A dispute is a stop condition, so make sure a reply pauses your sequence. And the decision to escalate, add a fee or refer a debt is judgement about a relationship. The law is less settled at that end, so is it safe to let AI email your clients? is the more useful page.
Setting the boundaries yourself
Your timing decision is only as good as the limits around it. How many contacts per customer per fortnight. Which channels are allowed for which client. What pauses on a reply or a promise. What may go out without you.
Adfin's Customer Agents work inside those limits: you set the rules of engagement, the agent picks the next action per customer from that customer's own history, and messages go from your own email domain. Whether that collects more than your current schedule is not something anybody can show you yet, and our claim is narrower than the market's. There's no evidenced best time to send a reminder in general, and there can still be a better time for this client of yours. How often should you chase an unpaid invoice? has a default cadence.
Common questions
How does AI decide when to chase an invoice? It combines the triggers you configure, a prediction of when this customer is likely to pay, and that customer's own record of which channels got a reply. The first is an ordinary schedule and the other two make the decision per customer.
Can AI predict when a customer will pay? Peer-reviewed research on one German firm's ledger of more than a million invoices found neural networks suitable for predicting payment dates. It forecasts and tests no intervention, so it doesn't show that chasing differently changes when anybody pays.
Is there a best time of day to send a payment reminder? There's no reliable public evidence for one, and we can't publish a finding either, so treat a universal recommendation as a guess. What you can get at is narrower: what your client has replied to before, and how quickly they paid.
Does chasing earlier get you paid faster? Independent research on tax reminders found that earlier letters brought payment forward without changing how many people paid. Those were tax debts and letters, so read the direction and not the number, and expect earlier contact to change speed more than rate.
What happens with a brand new customer? No history means no pattern, so they get your default schedule until a record builds. Agreeing terms and the due date up front does more for a new account than any timing decision, since payments cluster on the date you set.
Does the size of the invoice change the timing? It changes what a slip costs you more than your cadence. A large invoice is worth confirming as approved before the due date, and our own data doesn't support chasing large invoices harder.
Sources
- Journal of Applied Accounting Research — Kureljusic and Metz, machine learning for predicting payment dates (accurate as of August 2026)
- Tax and Transfer Policy Institute, ANU — Gillitzer and Sinning, nudging businesses to pay their taxes, does timing matter (accurate as of August 2026)
Reviewed by the Adfin team. This describes how automated and machine learning chasing systems make timing decisions, and summarises the published research as at August 2026. Platform figures are Adfin's own data with the population stated, and they describe payment behaviour on Adfin's book and not the effect of any AI feature. No independent study shows that AI-timed chasing collects faster than a well-configured schedule, so treat any percentage you're quoted as a marketing claim until its method is published.
