A professional services firm's CFO should have the billing manager record, on every client invoice, the date it was paid and whether the client disputed it, before buying an AI collections tool that predicts which invoices will be paid late.
Researchers at IBM built models of that kind in a paper presented at the ACM's KDD conference in August 2008. IBM's annual report for 2025 says it is offering new products and services associated with AI development, deployment, governance and management. The team used invoices created from March 2004 to February 2005 at four firms, two of them Fortune 500 companies: three supplying high-tech equipment for telecommunication, networking and IT services, and one specializing in online advertising placement and scheduling services. Each invoice was labelled on time or late from its due date and the date it closed. Of 54 fields recorded per invoice, three were left once identifiers, facts not known at issue and fields with too many missing or unique values came out: the amount, the payment terms and whether the invoice was under dispute. Those three improved only marginally on always predicting the most common outcome at two of the firms, while adding each customer's payment history raised accuracy at all four.
A second IBM Research team worked with a multinational bank in a paper posted in August 2020 for a KDD workshop on machine learning in finance. It covered 91,562 invoices to 2,229 of the bank's customers in six Latin American countries and the United States, from November 2018 to November 2019. Its model reached up to 81% accuracy in predicting whether an invoice would be paid late, where always predicting late scored 54.51% on the test invoices. The team found a large number of missing values and incorrect data in several fields, and says the field showing whether an invoice was under dispute had consistency issues, due to being manually entered information.
The engagement partner should pass each client dispute, including after payment, to the billing manager within five working days, and the billing manager records it on the invoice from one fixed list (hours, rates, a description of work, scope, other), naming the time entries concerned where hours or descriptions are questioned. A corrected invoice carries the original invoice number and stays marked disputed, so a client who disputes the corrected bill too stays one record, with the partner and off any reminder schedule.
In month one the billing manager exports every invoice from the last twelve months with its due date, any credit note and a paid date only once the balance is paid or credited, and the controller counts closed invoices with no paid date and invoices with no dispute entry. During month two the billing manager fills those gaps from credit notes and partner emails, and the list goes live. By month three the CFO gives each shortlisted vendor the first nine months to train on and the last three to test, split by date as the bank study was. An operating partner can hold every professional services company in a portfolio to that same count of missing entries.
If no vendor's model beats, on those three months, the accuracy of always predicting the more common outcome, the CFO buys none.
In the bank study, three of the six model types tested, the two most accurate among them, scored best when each customer's payment history covered the previous four months, of windows from three to twelve; two others did best at ten, and the authors attribute the short window to customers paying less late as the months went on.
