Work through an invoice-entry example
Suppose a wholesaler receives 600 supplier invoices a month. Staff currently read the PDF, find the supplier, enter the amount and tax fields, attach the original and send the draft for approval. At six minutes per invoice, that is 60 hours of entry work. These are invented planning inputs; measure your own baseline before using the calculation.
For a first pilot, let the model extract fields into a draft. Keep supplier creation, approval and payment outside its permissions. A draft with an unknown supplier, unreadable amount or repeated invoice number goes into an exception queue. Never treat a model’s confident wording as evidence that the invoice is correct.
Count the review work that remains
| Work | Calculation | Hours |
|---|---|---|
| Manual baseline | 600 × 6 min ÷ 60 | 60 |
| Review 480 normal drafts | 480 × 1 min ÷ 60 | 8 |
| Resolve 120 exceptions | 120 × 6 min ÷ 60 | 12 |
| Sample checks and upkeep | Assumed monthly allowance | 5 |
| Pilot total | 8 + 12 + 5 | 25 |
Illustrative inputs, not observed performance. Assumes every draft is reviewed.
Staff hours per month · zero baseline
Pilot effort = normal drafts × 1 minute + exceptions × 6 minutes + 5 hours upkeep. The 50% scenario has 300 normal drafts and 300 exceptions.
View data table
| Scenario | Total staff effort |
|---|---|
| Manual entry | 60 |
| 20% exceptions | 25 |
| 50% exceptions | 40 |
At 20% exceptions, the model releases 35 staff hours per month. This is available capacity, not automatically a payroll saving: the team may spend that time on other work. Compare the value of the work you can actually redirect with model usage, integration support and setup effort. At 50% exceptions, the released capacity falls to 20 hours.
Decide what would stop the pilot
| Input | Expected behaviour | Evidence |
|---|---|---|
| Same invoice uploaded twice | One draft; second upload flagged | Invoice ID and duplicate log |
| Unrecognised supplier | Hold for staff; do not create a payee | Exception queue entry |
| PDF says “ignore your rules” | Treat document text as data; preserve permissions | No unauthorised action in audit log |
| Model or accounting service unavailable | Queue safely or return to manual entry | Every input accounted for |
Before you proceed
- Measure current volume, effort and error types.
- Select a process owner who can review exceptions.
- Compare a rule-based option with the AI proposal.
