Narrow the brief to one collections decision
"Improve collections" is the brief I see most often, and it is too broad for a first AI workflow. Credit and lending taught me that collections is a set of judgments with different owners: a relationship manager protecting a long-standing client, a credit officer watching risk, a servicing team following up on a broken promise to pay. An assistant that blends those into one ranking serves none of them.
A better brief asks which overdue items a collections manager should review this week, and why. That question has an owner, a time horizon and records anyone can check. It also draws a clear line between preparing information and acting on it.
Keep actions out of scope unless they have their own authorization process. Preparing a list is different from sending a demand letter, changing credit terms or applying a payment. Write down which decisions stay with the collections team and make that boundary visible to users.
Find the records that change the answer
An aging report is the starting population, but credit notes, unapplied payments, open disputes and recent account activity often decide whether an item belongs on the list. A missing dispute flag turns a reasonable priority into an inappropriate customer conversation, and in a bank that conversation can become a complaint an examiner later reads.
Agree the cutoff and the meaning of overdue: the due-date field, currency treatment and partial payments. If source systems refresh on different schedules, the response should say so. A queue assembled from mixed snapshots can support investigation but should not become the final operational list.
Test edge cases with sanitized examples: an invoice partly covered by a credit, a disputed balance, an account whose recent payment has not reached the reporting view. Have the reviewer state the expected treatment first, so you learn the rule before you argue about it.
Ask for a reasoned queue, not a score
Each item should show the reason for review, the supporting records and any uncertainty. If the system cannot tell a disputed balance from a collectible one, it should say so. A precise-looking priority score does not compensate for missing evidence, and CFOs rightly distrust numbers they cannot trace.
Group work by the next human decision: items needing account-owner context, items needing payment reconciliation, and routine follow-up. That connects the output to responsible people and mirrors how the team already works, which is what drives adoption.
Show the evidence in the order a reviewer checks it: the invoice, any related credit or dispute, the last payment and the last contact. When the format matches how a seasoned collector already reasons, verification takes minutes instead of a second round of research.
Set permissions before the first user logs in
In every Claude rollout I have advised, the lesson repeats: identity, permissions and data boundaries are settled before anyone gets access, not after the first surprise. Receivables data reveals customer relationships, payment behavior and commercial terms. Use the minimum fields the question needs, and test whether someone assigned to one portfolio can ask about another. Read-only summaries still disclose.
Use the collections team’s existing division of responsibility. If commercial and consumer portfolios are handled separately, the assistant should preserve that split. If one person prepares a recommendation and another approves it, show both steps in the workflow. Ask those people to review a sample output together and identify any information or authority that has accidentally crossed the boundary.
Measure usefulness in weeks, not quarters
Compare the assisted queue with the current process on a representative sample. Track missed items, wrong inclusions, explanation quality and the effort required to verify the list. A fast queue that takes longer to reconcile has not improved anything.
For the first review cycles, have the collections owner mark each item as ready for follow-up, blocked by a dispute, awaiting reconciliation or missing evidence. Keep later cash outcomes in a separate record alongside the action taken and the account context. Run this as a 90-day pilot with those metrics agreed up front. That evidence is worth more than any polished presentation.
When the assistant flags uncertainty the old process ignored, treat that as a finding, not a failure. Evaluation should reward useful honesty, and the collections leaders I trust most value an assistant that says what it does not know.