30 supplier invoices a day, each manually verified against GST numbers, PAN records, and work order references. A pipeline extracts, validates, and flags — so the accounts team reviews only what needs them.
Accounts team spending 1.5–2 hours daily on invoice verification — for 30 invoices.
Each invoice required manual cross-checks: GST number, PAN number, work order reference, line item totals.
Work order numbers rarely matched exactly — small formatting differences between vendor and internal systems.
Attention degraded across the batch — errors concentrated at the end when fatigue had set in.
Roughly 1–2 incorrect invoices processed every day, requiring correction cycles: calls, credit notes, reconciliation.
Skilled accounts staff spending most of their verification time on invoices that had no issues.
Three stages. Clear invoices pass straight to the payment queue; only failures reach a human.
Invoices arriving by email are processed through an OCR pipeline. Vendor name, GST number, PAN number, invoice date, line items, totals, and work order reference are extracted and normalised into structured fields — ready for validation. No manual data entry.
Next: structured fields pass to the rules engine.
GST and PAN numbers are validated against format rules and cross-referenced with the approved vendor register. Work order references are fuzzy-matched against open purchase orders — handling differences like "WO-2024/1234" vs "WO2024-1234". Line items are checked against approved rate cards and totals reconciled.
Passes: clear invoices go directly to the payment queue. Triggers Stage 3: any failed check.
Invoices that fail any check are flagged with a specific reason — wrong GST number, unmatched work order, rate card deviation — and queued for review. The accounts team opens the day with a short list of items that actually need attention, not a full batch of 30.
Volume: ~4% of invoices flagged — the rest clear without human touch.
| Metric | Before | After |
|---|---|---|
| Daily verification time | 1.5–2 hours — full batch, every invoice | ~45 minutes — flagged items only |
| Effort reduction | — | 50% of daily verification time recovered |
| Error catch rate | Variable — fatigue-dependent, ~1–2 slip-throughs per day | 4% flagged — consistent, end-to-end |
| Human review scope | All 30 invoices | Flagged invoices only — with reason attached |
| Correction cycles | ~1–2 per day (vendor calls, credit notes, reconciliation) | Near zero — caught before payment |
| End-of-batch accuracy | Lower — attention degraded across the batch | Consistent — validation rules don't fatigue |
"Software handles the volume. Humans handle the ambiguity."
The system doesn't replace human judgement — it focuses it. Accounts staff still review every flagged invoice. Their skill is applied where it matters: on the 4% that have genuine problems.
Humans are poor at sustained, repetitive verification. Attention drifts, patterns stop registering, small discrepancies get missed — especially at the end of a long batch. Software doesn't fatigue. The 50% effort saving came not from working faster but from eliminating the 96% of the task that should never have been human work in the first place.
Invoice session counts are matched against the booking system automatically. A therapist who invoices for 18 sessions when 16 are logged is flagged before payment — not discovered at month-end during reconciliation.
Unit prices on supplier invoices are checked against last agreed rates. A supplier who quietly increases a line item by ₹50/unit across a monthly order is caught at the first invoice, not when someone spots the trend in quarterly accounts.
GST numbers on every incoming invoice are validated against the vendor register. Receiving invoices from an unregistered vendor or a mismatched GST entity is a compliance and reconciliation risk — caught automatically, every time.
Service invoices are matched against work orders raised. A maintenance firm billing for a job with no corresponding work order in the system is flagged immediately — before any payment approval, not after.
90-minute AI Opportunity Audit. No pitch — just findings you can act on.