AI project ROI equals total benefits minus total costs, divided by total costs, over one period, usually three years. Costs include implementation, integration, maintenance and staff time. Benefits count only savings you will actually capture.
Most AI business cases fail at the finance review for one of two reasons. Some count savings nobody will capture. Others leave out half the cost side. This framework gives the formulas, a full list of cost and benefit lines, and one worked example with every number shown so you can replace ours with your own. If you need the case built from your own workflows and presented to a board, see our AI business case service.
What is the formula for AI project ROI and payback?
Three formulas cover what a finance committee asks for.
- ROI = (total benefits − total costs) ÷ total costs × 100%, over the same period.
- Payback period = months until cumulative net benefit equals the one-time investment. Net benefit per month = monthly benefit − monthly running costs.
- Net present value (NPV) = the sum of each year's net benefit divided by (1 + discount rate) raised to the year number, minus the initial investment. Use it when the finance team discounts other projects.
Use the same horizon for costs and benefits. Three years is common because it captures maintenance and at least one change of model version. A one-year view makes most projects look worse; a five-year view assumes the process and tools stay stable for longer than they usually do.
For a fast first estimate, our AI ROI calculator applies these formulas to your volumes and hourly costs.
Which costs belong in an AI business case?
Every line below is either one-time or recurring, and both kinds belong in the three-year total.
| Cost line | Type | What it includes | Often missed? |
|---|---|---|---|
| Discovery and design | One-time | Process mapping, baseline measurement, use-case selection | Sometimes |
| Build or configuration | One-time | Prompts, workflows, retrieval setup, testing | No |
| Integration | One-time | Connections to CRM, document store, ERP or line-of-business systems | Yes |
| Internal staff time | One-time | Requirements, test cases, user acceptance, training sessions | Yes |
| Privacy and governance | One-time, then yearly | Privacy impact assessment, AI use policy, vendor review | Yes |
| Model usage | Recurring | Tokens per request times volume, billed per million tokens | Rarely large |
| Hosting and logging | Recurring | Cloud resources, log storage, monitoring | Sometimes |
| Maintenance | Recurring | Prompt and rule updates, evaluation runs, model version changes | Yes |
| Change and transition | First months | Lower productivity while staff learn the new process | Yes |
Model usage is easy to estimate once you know volumes. As of September 2026, Anthropic lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens (USD). Google lists Gemini 3.5 Flash-Lite at $0.30 and $2.50. OpenAI lists gpt-6-luna at $0.10 and $0.50. A process with 2,000 documents a month, 5,000 input tokens and 500 output tokens each, uses 10 million input and 1 million output tokens. At Claude Haiku 4.5 prices, that is about $15 USD a month.
Staff time and integration usually cost more than the model.
Which benefits can you count, and how do you measure them?
Count a benefit only if it has a baseline, a unit of measure and an owner who will report it. The table orders benefits from most to least defensible.
| Benefit | How to measure | Formula | Confidence |
|---|---|---|---|
| Staff hours saved and redeployed | Time per unit before and after, volume | (old minutes − new minutes) × volume × loaded hourly cost × capture rate | High if the capture rate is real |
| Rework and error cost avoided | Errors per month, time to fix each | errors avoided × hours per fix × loaded hourly cost | High with a measured error log |
| Backlog cleared or capacity added | Queue size, cost of overtime or temporary staff | cost of the extra capacity you no longer buy | Medium |
| Faster cycle time | Days from request to decision or invoice | value of earlier cash or of work won on speed | Medium |
| Risk reduction | Probability and cost of a specific failure | reduction in probability × cost of the failure | Low: show it separately |
| Staff satisfaction and retention | Turnover in the affected roles | track it; keep it out of the payback figure | Qualitative |
The capture rate is the share of saved hours that turns into value: redeployed to other work, overtime cut, or a hire avoided. Hours that disappear from timesheets with no change in staffing or output are not a saving. Ask the process owner what they will do with the time before you count it.
How do you calculate payback from a 15% improvement in an operational metric?
Multiply the annual cost tied to the metric by the improvement and by the capture rate, then compare the result with costs. The example below uses round, illustrative inputs. It is not a client result; replace every input with your own.
Inputs (assumptions). An accounts payable team handles 2,000 invoices a month at 15 minutes each. Loaded staff cost is $60 an hour. AI extraction and matching cut handling time by 15%.
- Baseline cost: 2,000 × 0.25 h × $60 = $30,000 a month, or $360,000 a year.
- Gross benefit at 15%: $4,500 a month, or $54,000 a year.
- Year one assumes half the benefit while the team adjusts.
| Item | Base case | Conservative case |
|---|---|---|
| Capture rate | 100% | 60% |
| One-time costs: build $10,000, integration $15,000, staff time $4,800, training $1,200 | $31,000 | $31,000 |
| Running costs: model usage about $15, maintenance $240, hosting $100 a month | $4,260 a year | $4,260 a year |
| Benefits, year 1 / 2 / 3 | $27,000 / $54,000 / $54,000 | $16,200 / $32,400 / $32,400 |
| Three-year benefits | $135,000 | $81,000 |
| Three-year costs | $43,780 | $43,780 |
| Net three-year return | $91,220 | $37,220 |
| Three-year ROI | 208% | 85% |
| Payback | about 14 months | about 20 months |
Present the conservative case as the main number. If the project beats it, the decision looks better over time. If it does not, the project still pays back.
What ROI do you need to approve an AI project?
There is no standard threshold for AI projects. Apply the hurdle rate or payback limit your organization already uses for other investments of similar size and risk. Three checks make an AI case easier to approve:
- It passes in the conservative scenario, with a realistic capture rate and a slow first year.
- The baseline is measured, with current volumes, handling times and error counts taken before the pilot.
- A small first step tests the biggest assumption. A six-week pilot on real work shows whether the 15% holds before the full budget is committed.
Which mistakes inflate or deflate an AI business case?
Five mistakes account for most weak cases.
- Counting uncaptured hours. Savings without a plan for the freed time.
- Using vendor ROI figures as the analysis. They tend to assume best-case automation rates and leave out integration and staff time. Use them as a starting point and rebuild with your own data.
- Straight-line benefits from day one. Benefits ramp up; costs arrive first. Model the first year month by month.
- Leaving out maintenance. Prompts, rules and evaluation sets need updates, and vendors retire model versions.
- Undercounting benefits. Teams sometimes ignore rework, backlog and cycle-time gains because they are harder to price. Include them with their own confidence level.
How should you present the case to finance?
Answer four questions on one page: what we spend, when we get it back, what could go wrong, and how we will measure it. Show the base and conservative scenarios side by side, list the assumptions behind each, and name who reports the metrics and how often.
If you need ranked use cases with ROI ranges before you build, the two-week AI Opportunity Audit interviews your team and produces them, at $2,900 CAD + HST. The six-week AI Pilot Sprint ($9,800 CAD + HST) then tests the main assumption on real work. Both are listed with our other fixed-price packages. For document-heavy processes like the example above, AI document processing is the usual first build.