Annual workload
annual hours = annual units × total team minutes per unit ÷ 60
+ rework hours
AI value measurement
How to establish an AI ROI baseline before implementation
Define one unit of work, observe a representative period, and record the old workflow before the new system changes behavior. Without that contract, almost any later ROI story can sound true.
Short answer: record volume, touch time, wait time, rework, loaded cost, quality, exceptions, and the downstream revenue or capacity constraint for one clearly defined unit of work. Name the measurement window and owner. Use the same definitions before and after implementation.
Baseline contract
Measure the operation from six angles.
| Measure | Definition | Example | Common mistake |
|---|---|---|---|
| Unit and volume | The repeatable case and completed/received count per period | Invoices processed per week | Mixing different case types |
| Touch time | Minutes people actively work on each unit | 18 minutes across two roles | Using elapsed time as labor |
| Wait time | Elapsed time between required steps | 1.6 days waiting for approval | Ignoring customer or revenue delay |
| Rework and exceptions | Share repeated, corrected, escalated, or rejected | 12% need a second pass | Counting only happy-path work |
| Loaded cost and quality | Labor cost plus defects, credits, leakage, or risk | $68/hour and 2.3% error rate | Calling capacity “cash savings” |
| Outcome constraint | The business result the workflow limits | Renewal reviews completed before risk window | Optimizing activity without outcome |
Formulas
Keep the math simple enough to audit.
Loaded labor cost
loaded cost = annual hours × fully loaded hourly cost
Before/after change
improvement % = (baseline − observed result) ÷ baseline × 100
Report capacity released, cash cost avoided, revenue accelerated, and risk reduced separately. They are different economic mechanisms. Capacity becomes cash savings only if spending falls; it becomes revenue only if the released constraint is redeployed and produces revenue.
Worked example
When does human review erase the return?
These invented assumptions show the calculation, not a client result, price quote or performance benchmark. All dollar figures are USD. A workflow completes 1,000 cases a month at 12 human minutes per case, including normal rework. At $60 per loaded hour, that is 200 hours and $12,000 of monthly labor capacity.
Assume AI is actually used for 600 cases (60% adoption); the other 400 stay manual. Each AI-assisted case still needs 3 minutes of human handling, plus review and correction. Both cases below must meet the same accepted quality.
| Monthly measure | 2 minutes of review per AI-assisted case | 8 minutes of review per AI-assisted case |
|---|---|---|
| Human workload after AI, all 1,000 cases | 130 hours | 190 hours |
| Capacity released from the 200-hour baseline | 70 hours | 10 hours |
| Released capacity valued at $60/hour | $4,200 | $600 |
| New recurring cash spend for AI and support | $800 | $800 |
| Capacity value minus new recurring spend | $3,400 | −$200 |
hours released = AI-assisted cases × (old human minutes − remaining handling minutes − review minutes) ÷ 60
At about 7 minutes 40 seconds of review per AI-assisted case, the released capacity value only covers the $800 running cost. A one-off $6,000 setup cost is additional and excluded from these monthly figures. Adoption ramp-up and lost quality would weaken the result further.
Capacity value is not cash saved. If staffing, contractor bills and revenue stay unchanged, the cash impact is still an extra $800 a month plus the $6,000 setup bill. Record which spending can actually stop or which additional work can produce revenue before calculating financial ROI or payback.
Start with the workflow calculator to estimate current effort. Its capacity range does not deduct implementation or running costs. Replace directional assumptions with observed adoption, review time and quality, then use the cost-per-accepted-result worksheet to check the operating cost.
Collection protocol
A baseline leadership can defend
Choose a representative window
Use enough cycles to include normal volume and meaningful exceptions. Record seasonality, backlog, launches, and policy changes.
Triangulate
Combine system timestamps, a sample of direct observation, and operator review. Any one source can misstate the real work.
Freeze definitions
Name the owner, formulas, exclusions, data source, review window, and acceptable quality floor before the pilot starts.
Use one time convention
If two people spend 8 and 10 minutes on a case, enter 18 total team minutes. Multiply that total by case volume once. Add rework only if its time is not already included. Use the workflow calculator.