Cost Benefit Analysis Formula for AI Projects

What is the cost benefit analysis formula?

Present value of benefits (PVB) =
  Σ [Benefit at time t ÷ (1 + discount rate)^t]

Present value of costs (PVC) =
  Σ [Cost at time t ÷ (1 + discount rate)^t]

Net present value (NPV) = PVB − PVC

Benefit-cost ratio (BCR) = PVB ÷ PVC

Return on investment (ROI) =
  (PVB − PVC) ÷ PVC × 100%

The formulas are simple; the difficult work is defining a credible “with-project” change against business as usual, assigning costs and benefits to the correct period, avoiding double counting, selecting consistent units and discount rates, and preserving non-monetized impacts beside the summary metrics.

Choose the metric that answers the decision

Use NPV and BCR together. With identical discounted inputs, ROI equals BCR minus one, expressed as a percentage: a BCR of 1.70 corresponds to approximately 70% ROI. Payback adds a liquidity and timing view but should not replace lifetime value. No single metric determines strategic fit, operational feasibility, legal permission, or acceptable risk.

Calculate incremental value against business as usual

Cost-benefit analysis compares options, not a project against zero. Define business as usual as the credible future without the proposed investment, including existing contracts, expected demand growth, required maintenance, known hiring, and planned process changes. Then calculate only the difference caused by each option.

Incremental benefit at time t =
  outcome with project at t − outcome under business as usual at t

Incremental cost at time t =
  cost with project at t − cost under business as usual at t

For example, if analysis demand would rise 15% without AI and the team would already add two contractors, avoided contractor cost may be incremental. If a contract was ending regardless of the project, its removal is not an AI benefit. Record the owner and evidence for every business-as-usual assumption.

Include the complete cost and benefit boundary

Use opportunity cost for internal labor: time spent implementing the project cannot simultaneously deliver another priority. Keep sunk costs—past spending that cannot change—outside the forward decision, while still recording them for governance or lessons learned.

Discount future costs and benefits consistently

Discounting converts values occurring in different years to a common present-value basis. A future value is divided by one plus the discount rate raised to the period number. Year 0 is not discounted. Use the organization’s approved rate and valuation policy; do not copy a public-sector social discount rate into a private investment model without confirming that it answers the same question.

Present value at time t =
  future amount at t ÷ (1 + discount rate)^t

Discount factor at time t =
  1 ÷ (1 + discount rate)^t

Do not add inflation to a real discount rate and then apply it to nominal cash flows without a consistent conversion. State the currency, price year, nominal or real basis, tax treatment, discount rate, timing convention, and horizon at the top of the model.

Apply six rules before calculating BCR or NPV

1. Use incremental flows.Subtract the business-as-usual outcome from every option.
2. Use one price basis.Keep all values real or all values nominal with a matching rate.
3. Assign the correct period.Record when resources are consumed and benefits are realized, not when a slide predicts them.
4. Use mutually exclusive value routes.Do not count the same recovered hour as avoided cost, added output, and strategic capacity.
5. Include negative effects.Model correction, downtime, incidents, transition loss, and residual risk.
6. Keep an evidence register.Attach source, owner, method, date, confidence, and refresh rule to each input.

Calculate a three-year AI project example

This hypothetical example evaluates an AI-assisted data-analysis workflow over three years using an 8% discount rate. Year 0 includes implementation, integration, and training. Annual benefits combine mutually exclusive, verified routes: avoided external analysis, realized value from reused capacity, and lower rework. All figures are illustrative and rounded.

The base case has positive NPV, BCR above 1, and discounted payback during year 2. That is not an automatic approval. Review whether the benefit evidence is attributable, whether the option meets nonfinancial constraints, whether a lower-cost option creates better value, and whether downside scenarios remain acceptable.

Calculate discounted payback with uneven cash flows

Do not divide initial cost by an average annual benefit when cash flows vary. Accumulate discounted net benefits period by period. In the example, the unrecovered balance after year 1 is $54,815. Year 2 contributes $114,883 of discounted net benefit, so interpolation gives approximately 1 + 54,815 ÷ 114,883 = 1.48 years.

Discounted net flow at time t =
  (benefit at t − cost at t) ÷ (1 + rate)^t

Payback =
  full periods before crossing
  + unrecovered balance before crossing
    ÷ discounted net flow in crossing period

Interpolation assumes the crossing-period flow arrives evenly. If benefits arrive at a known date—such as annual contract avoidance—use monthly or quarterly periods instead. State whether the reported payback is simple or discounted because they answer different questions.

Prevent double counting in AI benefit formulas

Create a value ledger with one row per outcome, a unique identifier, causal mechanism, baseline, formula, owner, and overlap group. If two rows share the same overlap group, explain why both remain or remove one.

Test sensitivity, scenarios, and switching values

A point estimate hides decision risk. Use one-way sensitivity to identify influential inputs, coherent scenarios to show combinations of conditions, and switching values to show how far an assumption can move before the decision rule changes.

With costs unchanged, the example reaches BCR 1 when realized benefits fall to about 58.8% of the base forecast. In other words, benefits can fall by roughly 41.2% before NPV reaches zero. This is a switching value, not a probability. Pair it with evidence about how plausible such a shortfall is.

Do not create a “downside” scenario by changing only one convenient variable. Link adoption, productivity, quality, demand, capacity reuse, and support cost in ways that can occur together. Preserve the formula and source behind every scenario value.

Value AI benefits from observed operational evidence

Eligible annual tasks =
  total annual tasks × eligible share

AI-exposed tasks =
  eligible tasks × observed adoption

Net hours recovered =
  exposed tasks ×
  (baseline labor − assisted labor − review − rework) ÷ 60

Realized capacity value =
  net hours recovered × approved reuse rate × value per reused hour

Use accepted output, not generated output. Include all roles that prepare, review, correct, document, or operate the workflow. Apply adoption only to eligible tasks. Value recovered capacity only to the extent that it is reused for approved work; keep unused capacity visible instead of assigning it a fictional cash value.

For risk reduction, model expected loss carefully: change in event probability multiplied by the attributable impact, less control and residual-risk costs. Use ranges when events are rare, and never present the full worst-case loss as an annual benefit.

Keep material non-monetized impacts beside the formula

Not every consequence should be forced into dollars. Data privacy, explainability, employee impact, accessibility, supplier lock-in, resilience, strategic learning, and decision quality may be material even when credible valuation is unavailable. Describe direction, affected stakeholders, scale, timing, evidence strength, and whether the impact is a constraint.

Structure the calculation so another reviewer can reproduce it

Separate inputs, calculations, outputs, and evidence registers into visible areas. Lock formula cells, version assumption changes, and add automated checks: totals reconcile, discount factors decline correctly, all flows have units, no blank owners remain, and low/base/high ordering is logical.

Interpret the formulas as evidence, not a verdict

Compare all feasible options on the same scope and horizon, including a minimum viable change and business as usual. A higher BCR does not always dominate a larger positive NPV when budgets, capacity, strategic objectives, and indivisible investments differ. Document the trade-off and the decision owner’s rationale.

Use a ten-step cost benefit analysis workflow

1. Define the decision.Name the owner, options, deadline, objectives, constraints, and success criteria.
2. Describe business as usual.Forecast the credible future without the investment.
3. Map the causal mechanism.Connect the intervention to operational changes and final outcomes.
4. Set scope and horizon.Include material setup, operation, transition, and end-of-life effects.
5. Quantify incremental flows.Use attributable quantities and mutually exclusive value routes.
6. Choose the valuation basis.State currency, price year, real or nominal treatment, and discount rate.
7. Calculate present values.Apply timing consistently and reconcile undiscounted totals.
8. Calculate NPV, BCR, ROI, and payback.Use the same input boundary and label each definition.
9. Test risk and uncertainty.Run sensitivity, coherent scenarios, and switching values.
10. Plan evaluation.Assign owners, leading indicators, review dates, and stop or scale rules.

Avoid ten cost benefit formula mistakes

Move from formula to evidence and implementation

Use the cost benefit analysis example for a narrative worked case, the cost benefit analysis template for a reusable decision structure, and the business case template to connect economic evidence with strategy, delivery, risk, and governance.

Check the model before presenting the result

• The decision, options, business-as-usual case, owner, deadline, and constraints are explicit.
• All costs and benefits are incremental, attributable, in scope, and placed in the correct period.
• The currency, price year, real or nominal basis, discount rate, horizon, and timing convention are stated.
• AI eligibility, adoption, acceptance, review, rework, and realized capacity reuse are evidenced.
• Every recovered workload unit has one mutually exclusive value route.
• NPV, BCR, ROI, and payback use clearly labeled, consistent input boundaries.
• Low, base, and high scenarios are coherent rather than arbitrary percentage changes.
• Sensitivity identifies influential assumptions and switching values.
• Material non-monetized impacts and mandatory constraints remain visible.
• Every material input has a source, method, date, owner, confidence, and refresh rule.

Cost benefit analysis formula frequently asked questions