Yes, for four parts of the job: drafting the message, working out when it goes out, choosing which channel to use, and handing an account to a person when it stops being routine. What it can't do is the judgement, the difficult call, or the decision about what a relationship will carry.
The short version
- Drafting, timing, channel selection and escalating to you are the parts software genuinely does now.
- Judgement about your own business, the conversation with somebody in difficulty, and what the relationship can take are yours.
- Independent research says a reminder raised the probability of payment by roughly 25 percent against no reminder at all, and that moving it earlier changed how fast people paid without changing whether they paid.
- No independent study shows AI chasing collecting more than a well-run schedule, so treat any percentage you're quoted as marketing.
- Where you put the approval gate depends on the consequence of the action and not on how good the model is.
The four parts it does
Each of these is a capability you can check for yourself in a trial, and not one of them claims an outcome.
| Job | What software does with it |
|---|---|
| Drafting | Writes a specific, courteous message from the invoice, the history and your tone, sent from your own domain |
| Timing | Decides when this customer's next contact goes out, from what they've done on previous invoices |
| Channel | Picks email, SMS or WhatsApp for this customer, and changes medium instead of repeating an email |
| Escalating to you | Recognises a query, a promise, a dispute or a silence and puts the account in front of a person |
The fourth gets underrated, and it's what makes the other three usable. A chasing sequence that runs into a dispute and keeps going does you real damage, so the value of software noticing and stopping is at least as high as the value of it sending.
Two things sit just outside this list and are worth knowing about. Query triage, where inbound replies get sorted by what they actually say, is standard now. And a statement covering everything a customer owes, sent when an account has gone quiet, does a job no individual reminder does.
The three parts that stay with you
Judgement about your own business comes first. Whether to extend terms on a big job, whether to keep working while an invoice ages, whether this client is worth carrying through a bad quarter: a model has no stake in your business and no view worth having on any of that.
Then there's the difficult call, meaning the conversation where somebody tells you they're struggling. That one wants a person on both ends, and it usually produces an instalment arrangement or an honest date rather than a payment. Difficult payment conversations covers what to say.
And the relationship decision. Chasing a client you also advise costs something that chasing a stranger doesn't, and only you can price it. So the useful way to divide the work is by whether the next step needs an opinion about a person.
What does the evidence show?
The strongest independent finding here is about reminders and not about AI. 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 compared with sending nothing, while varying its timing inside a three-week window changed the speed of payment without changing the rate, in work published by the Tax and Transfer Policy Institute at ANU. Tax letters aren't invoice emails, so read the direction and not the number.
Both halves are useful to you. Sending something matters a great deal, and that's the case for automating the sending. Earlier reminders got money in sooner without getting more of it in at all, a corrective to send-time marketing.
Beyond that, the ground is thin. There's no independent published study of AI-written against human-written reminders, and none on AI-chosen timing for commercial invoices. So software chasing your invoices is a capability claim you can verify, and any figure attached to it is somebody's marketing.
Where to put the approval gate
Set the boundary by what an action does to your customer rather than by how confident the software is.
A first polite reminder on a small, undisputed invoice inside terms is the safest thing to let run unattended. A late fee, an account stop, a threat of proceedings or a referral to an agency changes what your customer owes or what they can buy, so a person signs those. Anything to a customer who has raised a dispute or a complaint waits for you as well, along with any message quoting a specific interest or fee figure, since a wrong number in writing is worse than a vague sentence.
One more setting matters more than it looks. Cap how often any one customer can be contacted, in the software itself. UK law has always policed the frequency of demands for payment, and automation makes frequency free. Is it safe to let AI email your clients? works through the data protection and professional side, and what is agentic credit control? covers how boundaries get set.
Whether it's worth handing over
The case for it is about capacity. Every account you have is slightly different, and you can hold that in your head for ten of them: this one answers a text, that one needs the office manager, that one always pays after a statement. Across four hundred accounts the pattern goes, and the sequence quietly stops running in the weeks you're busiest. Businesses affected by late payment already spend an average of 86 hours a year chasing, on the Small Business Commissioner's research.
Keep two questions apart when you assess a tool. Is there a best channel or a best send hour across all businesses and all customers? No reliable public evidence settles that, ours included, so anybody quoting you one is guessing. Does this client of yours answer a text and ignore email? Their own record answers that, and a claim about one customer doesn't need a general rule behind it.
Adfin's Customer Agents run on that basis. You set the rules of engagement, so the channels, the persistence, what may go out unreviewed and what waits for you, and the agent picks the next action per customer from your own email domain, applying late fees or offering an instalment plan where you've allowed it. Whether that collects more than your current sequence is not something anyone can show you yet, and the honest pitch is that the routine part happens without you. AI for credit control covers the whole process stage by stage.
Common questions
Can AI chase invoices for you? Yes, for the routine parts: drafting the message, deciding when it goes out, choosing the channel and escalating to a person when something changes. The judgement about your business and your relationships stays with you, and so does any conversation with a customer in difficulty.
Will AI get your invoices paid faster? No independent study shows that, and every specific figure in this market is a vendor's own. What independent research does show is that sending a reminder matters a lot compared with sending nothing, and that reminding earlier tends to bring payment forward without changing how many people pay.
Does an AI reminder have to be approved before it goes out? No rule requires it, and how you set that up is your decision. A sensible split is by consequence: routine first reminders on undisputed invoices can run unattended, while fees, account stops, agency referrals and anything to a customer who has raised a dispute wait for a person.
Can AI decide which channel to chase on? It can pick per customer from that customer's own history of replies and payments. What no software can tell you is the best channel in general, because there's no reliable evidence for one, so treat a universal recommendation as a guess.
What happens when a customer replies? Query triage reads the reply, works out whether it's a promise to pay, a dispute or a request for a document, and routes it. The setting that matters is what a reply does to the sequence: a promise should pause it and a dispute should stop it.
Is it worth it for a small ledger? If you have ten accounts and you know them all, a simple schedule and your own attention will do. The argument gets stronger with volume, because a pattern per customer is straightforward to hold for ten names and impractical for four hundred.
Sources
- 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)
Reviewed by the Adfin team. This article describes what software can do and summarises the published research and UK legal position as at August 2026 for information, and it isn't legal advice. UK data protection law on automated decisions changed in 2026 and the regulator's updated guidance was still in draft when this was written, so if you're planning to let software apply fees, stop accounts or refer debts without a person approving each one, take advice on your own facts. If you're a member of a professional body, its guidance on AI tools covers your duties of competence and client confidentiality, and no software setting transfers those.
