A national medical-staffing agency put thousands of nurses and doctors on shifts across hundreds of hospitals, on contracts full of night, holiday and on-call rates. Every month their timesheets came in however people sent them, an app, a photo in a WhatsApp message, an email, and the payroll team had to verify all of it by hand, so pay ran late and people started leaving.
The contracts were the hard part. A clinician might be paid one rate for a night shift, another for a holiday, another for being on call, and the agency ran thousands of them across hundreds of hospitals. Every month, verifying who worked what came down to a person reading timesheets one by one.
And the timesheets arrived in every form imaginable. Some came through an app, some as a photo in a WhatsApp message, some in an email, and none of them agreed on a format. The payroll team spent the run-up to payday checking thousands of lines by hand, trying to spot a duplicate or a missing shift against the hospitals' own attendance records.
When a number was wrong, someone was paid late, or paid short. In a market already short of clinical staff, a late or incorrect payslip was a reason to leave, and the agency was carrying that risk every single month.
Let AI collect and read every timesheet, confirm it against the place it was worked, and surface only the ones that do not reconcile.
A document-understanding model collects the timesheets from the inbox and the staff portal and reads them whatever the format, an app export, a photographed sheet, an emailed screenshot, and pulls the shifts into one consistent structure instead of a person retyping them.
It applies each clinician's specific contract to the shifts they worked, the night rate, the holiday rate, the on-call rate, so the pay is calculated against the real agreement rather than a single default that never quite fits.
The reported hours are matched against the hospitals' own attendance records, so a shift is confirmed by the place it was worked, not just by the person claiming it, which is what makes the number trustworthy enough to pay.
An anomaly model flags the sheets that do not reconcile, a duplicated shift, a missing entry, an unusual leave or overtime claim, so the team's attention goes to the handful that are wrong instead of the thousands that are fine.
Only the flagged, contested sheets are sent to a payroll accountant to decide. Nothing questionable is paid on the model's word, and the person reviews each exception with both the timesheet and the hospital record in front of them.
The roughly 85% of sheets that reconcile cleanly flow straight into the payroll system for immediate payment, so the month-end scramble shrinks to the exceptions and the rest is simply done.
The clean majority pays itself; a person handles only what does not reconcile.
The monthly payroll run went from five hard days to a single day. Clinical staff are paid accurately and on time, the manual chase through thousands of timesheets is gone, and the team spends payday on the exceptions instead of the whole pile.
Payroll errors fell to almost none, which ended the compensation advances and disputes that used to follow a bad month, and in a market short of clinical staff, being paid right and on time became a reason to stay rather than to leave.
Payroll goes wrong when a person has to read thousands of timesheets in every format and catch the few that do not add up. The value here is an AI that reads any format, prices each shift by its real contract, checks it against the hospital's own record, and surfaces only the anomalies for a person to sign, so the clean majority pays on time and the mistakes are caught before payday. It is the same intake-and-anomaly discipline behind our payroll automation work, aimed at shift timesheets cross-checked against hospital logs rather than a bookkeeping bureau.
A firm's payroll bureau rekeyed inputs from hundreds of clients by hand every month-end. AI intake maps any format, an anomaly model flags outliers for a manager, and one rule change updates every client, turning the month-end scramble into minutes with far fewer errors and no new headcount.
Read the case → Bookkeeping automation · AccountingAn accounting firm's team burned days keying in invoices and reconciling ledgers by hand. AI document capture plus a learning categorisation engine cut month-end close 40% and manual entry 80%, and freed 25% more client capacity with no extra admin headcount.
Read the case →
Reef TRH