An invoice extraction demo saves five minutes. Does that make it a good investment? Only if the saving survives review, exceptions, integration costs, and the actual volume of work.
Use a unit of completed business work: an invoice accepted into the accounting system, a support issue resolved, or an enquiry correctly routed. Model calls completed is not a useful denominator for a business case.
A worked invoice example
The figures below are illustrative planning assumptions in Singapore dollars, not ModelShifts client results or supplier quotes.
| Input | Assumption |
|---|---|
| Monthly invoices | 2,000 |
| Current handling time | 6 minutes each |
| Loaded staff cost | S$30/hour |
| Invoices needing exception handling | 20% |
| Handling after automation, routine invoice | 1 minute |
| Handling after automation, exception invoice | 5 minutes |
| Monthly software and model cost | S$400 |
| Monthly maintenance and evaluation | S$600 |
| Initial implementation cost | S$18,000 |
The current workload is 2,000 x 6 / 60 = 200 hours, valued at S$6,000 per month.
With automation, routine handling takes 1,600 x 1 / 60 = 26.7 hours. Exceptions take 400 x 5 / 60 = 33.3 hours. Total handling is 60 hours, valued at S$1,800.
The monthly capacity saving is S$4,200. Subtract S$1,000 in recurring costs to get S$3,200 monthly net benefit. Simple payback is 18,000 / 3,200 = 5.6 months. Twelve-month net ROI is (12 x 3,200 - 18,000) / 18,000, or approximately 113%.
This is a planning result, not a promise. It excludes financing, tax, and discounting, and assumes volume and handling times remain stable.
Capacity released is not automatically cash saved
If the same team remains employed, payroll may not fall. The immediate benefit is capacity: staff can clear a backlog, absorb growth, or improve service. Treat it as cash savings only when spending actually falls or an otherwise necessary hire is avoided.
Keep separate columns for cash cost reduction, capacity released, and revenue contribution. Do not count a salary saving and the value of the same released hours twice.
Find the break-even volume
Weighted handling after automation is 0.8 x 1 + 0.2 x 5 = 1.8 minutes. The saving is 4.2 minutes, worth S$2.10 per invoice at the assumed hourly rate.
For twelve-month cost recovery, allocate S$1,500 per month to the build. Add S$1,000 recurring costs. Break-even volume is approximately 1,191 invoices per month: 2,500 / 2.10, rounded up.
This holds recurring costs constant. If OCR or model costs scale with volume, subtract their per-invoice cost from S$2.10 before calculating break-even.
Stress-test the weakest assumption
If exceptions rise to 40%, handling becomes 2.6 minutes per invoice. Monthly capacity saving falls to S$3,400 and net benefit to S$2,400. Simple payback becomes 7.5 months.
If volume drops to 800 invoices at the original exception rate, monthly capacity saving is S$1,680. Net benefit is only S$680, stretching payback to 26.5 months. Volume can matter more than the advertised model price.
What to measure during the pilot
Sample the current process before building. Record document type, page count, handling time, correction time, and final acceptance. Include messy scans, credit notes, unfamiliar suppliers, and duplicates.
For the new process, measure the same end-to-end outcome. Include time spent finding errors, correcting exports, and resolving failed integrations. Agree which errors require review; a wrong bank account is not equivalent to punctuation.
Close the pilot with a measured input table, exception breakdown, and sensitivity calculation. If the business case depends on perfect automation, narrow the scope before increasing the budget.
Discuss a document AI pilot with a measurable acceptance test and a cost model you can inspect.
