When a process speaks directly to clients or payers, the safest design is one in which every outgoing message is produced by rules that people have reviewed and approved, while AI works in the background to confirm that those rules behave as intended. This piece follows up on the invoice-reminder example in Improve First, Then Automate. The assistant, in operation since May 2026, now chases each type of receivable on its own cadence with approved wording, and none of its messages is written by AI. The approach applies well beyond receivables, to any process in which precision and the relationship both matter, such as statements, renewal notices, client correspondence or compliance letters.
Follow-up depended on a few people and on records that had drifted
The firm carries several hundred open invoices at any time, and they sit in quite different collection situations. Some are awaiting payment from a primary payer. Some have been declined by that payer and now fall to a second payer to settle. Some are for a portion of the fee the primary payer does not reimburse, some belong to files that take a long time to resolve, and some are older balances the firm had come close to writing off. Each situation calls for a different message, a different recipient and a different rhythm of follow-up.
Before this work, follow-up depended on the firm's project board and on the small number of people who knew how to run it. Reminders went out from a no-reply address, so a payer with a question about a file had no way to ask it. The automation covered only part of the work: primary payers received reminders on a fixed loop, while second payers were chased by hand and occasionally by fax. Over time the records had also drifted, so that many service labels were wrong and files sat in the wrong groups, and nobody could rely on the board to show which payers had been contacted and which had not.
The consequence was that invoices aged past 120 days and, in a number of cases, well beyond a year.
The firm used fixed rules for everything a payer would see
The firm built an accounts-receivable assistant on the tools it already used, and it has been in operation since May 2026, refined continually as the firm learned what worked. The first decision, and the one that shaped everything after it, was that every message reaching a payer would come from fixed rules and approved wording, so that the firm could always explain exactly what had been sent and why.
The assistant was given its own identity as a named payables administrator, sending from a monitored mailbox in the firm's Microsoft 365 tenant. Replies now reach a person, and every message sent is kept on record. Each weekday morning an automation engine reads the project board, determines which files are due, composes each email from the approved template for that situation, sends it, and records the reminder against the file. Each email carries its invoice as an attachment. An early version sent a link instead, but the links expired within an hour, which was before most recipients opened their email, and the change was made once that became clear.
The collections logic was then matched to how the business actually works. The receivables are now organised into eleven collection groups served by six message types, and each group has its own timing for the first reminder, its own interval between reminders, its own recipient and its own wording, as Exhibit 1 shows. Standard balances are followed up monthly, slow-resolving files quarterly, and an escalation track copies the firm's principals when a balance warrants it. Every message is personalised with the details the payer needs to act on it, including the file reference, the service, the invoice details and payment instructions.
A small set of rules applies across every group. A payer with several files due receives them together in one combined statement, no payer is contacted more than once in ten days, payers who have asked for fewer emails are contacted at half the frequency, and nothing is sent on statutory holidays.
Several hundred files then had to be moved into the correct collection groups, and this was done in waves, each preceded by a snapshot that would allow the change to be reversed. Where a moved file had recently had personal contact from a member of staff, its first automated reminder was deliberately delayed and staggered, so that a payer who had just spoken to someone at the firm would not receive an automated email the following morning.
AI was used to check the system
The emails themselves come from a rules-based engine by design, because in collections a payer needs to receive the same message, with the same terms, every time the same situation arises. AI, in this case Anthropic's Claude, was used for four tasks around that engine, each of which would have been slow or impractical to do by hand.
The first was an audit of the board data, looking for records that were missing, mislabelled or behaving unexpectedly. The second was verification: before more than 300 service records were corrected in bulk, a sample was checked against the underlying invoices to confirm the correction was right. The third was simulation, in which three months of sends were modelled on live data before any change went live, so that the firm could see exactly who would receive what, and when, before any email was sent. The fourth is a daily check, in which an AI agent compares each morning's sends with the forecast and reports any difference.
The audit found two coverage gaps, neither of which had produced an error message. As the board grew, a read limit had left part of the portfolio out of view, and a date check intended to begin monthly reminders on one category of balance had never triggered. Between them, the gaps left about 250 files without the follow-up they were owed, and both were closed. Both were the kind of fault that an automated process can carry for months, because it continues to run without complaint while doing less than its owners believe.
The daily check has already proved its value in operation. In the first week it confirmed that the sends matched the forecast exactly, and on one day it identified that the platform had timed out and nothing had been sent, which was corrected the same morning. Throughout, people approve every change that reaches a payer, and every bulk edit is backed by a snapshot from which it can be restored.
The design was built around the person receiving the email
The audit also made clear that the earlier process had been designed around files, whereas the people receiving the messages experience them as a relationship. The system was therefore refined around the recipient.
Each payer now receives one combined statement listing all of its open files, with the total outstanding and the largest invoices attached, where previously each file generated its own email. No payer hears from the assistant more than once in ten days, payers who have asked for fewer emails are automatically placed on half that frequency, and nothing is sent on statutory holidays. Two safeguards protect privacy and accuracy: a combined statement includes only the files associated with that exact email address, so that no payer ever sees another payer's files, and a file is marked as reminded only once the mail server confirms that the message was sent. The roughly 250 files that had been missed are being released over five weeks, grouped by payer, so that nobody receives a sudden burst of correspondence.
The results so far are in control and consistency, and cash is the next measure
The assistant has been live since May 2026, and the most recent refinements, which brought every account onto its agreed cadence and released the files that had been missed, went live in October 2026. The quarter to January will be the first to reflect the system as it now stands, so the results to date describe how the firm follows up rather than what it collects.
| Measure | Before | Now |
|---|---|---|
| Collection groups with their own cadence | none | eleven |
| Standard message types, each with approved wording | one generic reminder; the rest written by hand | six |
| Escalation to the firm's principals | none | automatic, on a defined track |
| The same payer emailed more than once in a day | regularly | never; at least ten days between emails |
| Service records corrected | n/a | more than 300, each checked against its invoice |
Over the quarter to January, the firm is tracking collections against billing, days sales outstanding, and how often an older account pays within 30 days of a reminder, and we will report the results when that quarter closes.
What this suggests for other client-facing processes
The pattern in this case is not specific to receivables. In any process that sends messages to clients, patients, members or payers, the organisation is accountable for every word that leaves it, and the people receiving those messages judge the organisation by their accuracy and their tone. That is a strong argument for keeping the outward-facing logic in rules that people can read, approve and explain.
It is equally a strong argument for using AI in the work that surrounds those rules. Rule-based systems tend to fail silently, and the checking required to catch those failures, across hundreds of records and months of scheduled activity, is precisely the kind of work that people rarely have time to do and that AI does well. Used in that way, AI makes a client-facing process more dependable without being placed between the firm and the people it serves, which is consistent with the principle we described in the earlier piece: rules where you can, and AI where you must.
Case details are drawn from work led by LEA and are presented with identifying details generalized.