AI automation ROI is the net monthly benefit compared with the one-time build cost. Net monthly benefit = (hours saved × loaded hourly cost × share that turns into real value) + (errors avoided × cost per error) − monthly running costs, including the time people spend reviewing. Payback in months = build cost ÷ net monthly benefit. Every input should be measured in a pilot, not assumed.
A business case for automation usually fails in one of two ways. Either it is a single optimistic number that nobody can trace, or it is an elaborate spreadsheet whose inputs are all guesses. This guide gives a formula that fits on one line, explains how to measure each input, runs three scenarios with made-up numbers so you can see how sensitive the answer is, and calculates a break-even volume. It extends the shorter method on our AI automation for business page. All figures below are illustrative arithmetic. They are not Timeline Digital results, client results, benchmarks or prices.
The formula, input by input
Net monthly benefit = (Hours saved × Loaded hourly cost × Value share) + (Errors avoided × Cost per error) − Running costs
Payback (months) = Build cost ÷ Net monthly benefit
| Input | How to measure it | Common trap |
|---|---|---|
| Monthly volume | Count items from the system of record for the last three months | Using a busy month as the average |
| Baseline minutes per item | Time a sample of real items, exceptions included | Estimating from memory; using the easy items only |
| New minutes per item | Measure in the pilot: items handled automatically take a little time, reviewed items take review time | Ignoring review time |
| Loaded hourly cost | Salary plus employer costs and overhead for the people doing the work | Using the salary alone, or the highest-paid person |
| Value share | The part of the saved hours that becomes something real: reduced overtime, a hire avoided, or capacity used on measurable work | Counting every saved minute as cash |
| Errors avoided | Baseline error rate minus new error rate, times volume | Comparing the AI’s error rate on a clean sample with people’s error rate on everything |
| Cost per error | Time to find and fix it, plus any credit, penalty or lost sale | Leaving it at zero or at a guess |
| Running costs | Model usage, hosting, monitoring, evaluation upkeep, reviewer admin | Counting only model usage |
| Build cost | The quoted or estimated cost to build and integrate | Leaving out integration, data preparation and training |
The two inputs that most often make or break a case are new minutes per item and value share. Both can be measured, and neither should be assumed.
How to compute new minutes per item
Not every item needs a person. Let r be the share of items that go to review, m_r the minutes a reviewer takes, and m_a the minutes a person still spends on an item handled automatically (for example, a quick glance at a log or a single click). Then:
New minutes per item = r × m_r + (1 − r) × m_a
If 30% of items are reviewed at 3 minutes each and the rest take half a minute of touch time, new minutes per item = 0.3 × 3 + 0.7 × 0.5 = 0.9 + 0.35 = 1.25 minutes. The review share r comes from the confidence-threshold design described in human-in-the-loop design for AI agents; the more you route to review, the lower the saving but the lower the risk.
A worked example with three scenarios (illustrative)
Imagine an invoice-handling process with these invented baseline figures:
- 3,000 items a month
- 8 minutes per item today
- 4% of items need rework today (120 items), at a cost of 20 each
- A loaded cost of 28 per hour
- An assumed build cost of 60,000, a placeholder chosen to show the arithmetic and not a quote or a market price
| Conservative | Expected | Optimistic | |
|---|---|---|---|
| Share of items sent to review | 60% | 30% | 20% |
| Review minutes per item | 4 | 3 | 3 |
| Touch minutes on automatic items | 0.5 | 0.5 | 0.5 |
| New minutes per item | 2.60 | 1.25 | 1.00 |
| Hours saved per month | 270 | 337.5 | 350 |
| Value of hours saved at 28 an hour | 7,560 | 9,450 | 9,800 |
| Value share | 40% | 70% | 100% |
| Real value from time saved | 3,024 | 6,615 | 9,800 |
| Error rate after | 2.5% | 1.5% | 1.0% |
| Errors avoided per month | 45 | 75 | 90 |
| Value of errors avoided at 20 each | 900 | 1,500 | 1,800 |
| Running costs per month | 1,500 | 1,280 | 1,280 |
| Net monthly benefit | 2,424 | 6,835 | 10,320 |
| Payback on 60,000 | 24.8 months | 8.8 months | 5.8 months |
Work one column by hand to check it. Expected: hours saved = 3,000 × (8 − 1.25) ÷ 60 = 337.5. Value = 337.5 × 28 = 9,450. Real value at a 70% share = 6,615. Errors avoided = 120 − 45 = 75, worth 1,500. Net = 6,615 + 1,500 − 1,280 = 6,835. Payback = 60,000 ÷ 6,835 = 8.8 months.
What does the spread teach? The same build is a two-year payback in one reading and under six months in another. The difference comes from three inputs: how much goes to review, how many saved hours become real value, and running costs. That is the point of measuring in a pilot. A business case that shows only the optimistic column is not one a careful finance team will accept, and one that shows all three is easier to defend.