Turning a Daily File of Returned IRA Contributions Into a Controlled, Auditable Reversal Run
Thread AI Editorial · June 11, 2026
At a large financial services firm, a treasury operations team starts each day with a bank’s report of rejected IRA contributions. Each one has to be traced back to its original deposit and reversed by hand, in two separate systems.
With Lemma, a Worker carries each returned payment from the bank’s report to a finished reversal. It enters each reversal once, stops on anything ambiguous, and records every decision along the way.
What Lemma delivers
One record writes to multiple systems
With no manual retyping.
100%
Of deposit matches, amounts, and reversal reasons decided by fixed rules, never by AI.
The Challenge
When a participant funds an IRA by electronic transfer, the firm’s bank credits the money right away. If the payment later bounces, for reasons like insufficient funds or a closed account, the bank pulls the money back. The participant’s account stays out of balance until someone reverses the original deposit.
That work starts each morning with a report from a legacy archive, typically 50 to 70 rows and sometimes as many as 180. For each row, an analyst looks up the participant, finds the original deposit, confirms the reason for the reversal, and then types the same details into a second system. At about four minutes a row, a typical day is several hours of careful copy-and-paste.
The routine rows aren’t the hard part. The exceptions are:
Look-alike deposits.
Two deposits can match on contract, amount, and date, and differ only by bank routing number.
Notices of change.
Some rows are bank notices with no money attached that should be handled elsewhere and must never trigger a reversal.
Restricted accounts.
A row belonging to someone on the treasury team has to go to an authorized reviewer, not the analyst on duty.
Weekend files.
Monday’s report can cover several days at once, which changes where the analyst has to look.
The Solution
The Worker follows the team’s own sequence: find the returned payments, research them, reverse them, and report back. Each morning it pulls the bank’s report, checks its totals, archives a copy, and forwards it unchanged to the pay-in team. For each row, it looks up the participant, searches both systems for candidate deposits, and translates the bank’s return code into the right reversal reason.
Every row is then sorted for an analyst to review:
- Clear matches are pre-filled and ready to approve.
- Ambiguous matches come with copy-ready values for the analyst to decide.
- Restricted accounts go to an authorized reviewer.
- Bank notices are set aside and never reversed.
Nothing is entered until it’s approved. After approval, the Worker enters each reversal in both systems and confirms it went through. At the end of the run, it reconciles the report against both systems and emails a summary.
Lemma Capabilities
Deterministic where money moves.
Matching, amounts, dates, and reason codes all follow explicit rules. The Worker never asks an AI model to choose a deposit or invent a reason, and when the evidence isn’t clear, it stops.
People hold the authority.
Handoff steps in the Worker put approval, exceptions, and restricted accounts in front of the right person, and reviewer corrections to the reason mapping are logged so the rules improve.
Audit by default.
Every search, candidate count, decision, write, and confirmation lands in an append-only event log with recordings of each system session, ready for control testing.
What’s Next
Analysts can spend their mornings on the exceptions that need judgment instead of retyping deposits, and reversals are handled the same way whether a veteran or a new hire is reviewing them. Because other returned payments follow the same steps, the firm can bring that consistency to annuity and institutional transfers, returned checks, and wires as well.
Does this sound familiar?
Your team types the same details into two systems every day.
Most items are routine, but the exceptions are where mistakes happen.
Key judgment lives with a few long-tenured people.
You need the process to work the same no matter who’s handling it.