ROI, NPV and Payback

Count the whole cost

Benefits (time saved, errors avoided, revenue gained) must be set against all costs: building and integration, model usage, infrastructure, monitoring, evaluation, training users, legal review and ongoing maintenance. Payback period is how long until cumulative net benefit covers the upfront cost. NPV (net present value) discounts future money because a rupee next year is worth less than a rupee today. Be conservative with benefits, show a range (pessimistic, expected, optimistic) and state your assumptions so others can challenge them.

Costs, benefits, ownership

A funded AI project needs a believable business case and a clear owner, team and way of working.

NPV and payback, run

I ran this with illustrative figures in rupees: build 1,000,000, run cost 300,000 a year, benefit 900,000 a year, 10% discount rate, 3 years. NPV is about 492,111 and payback takes 20 months.

def roi(build, run_per_year, benefit_per_year, years=3, rate=0.10):
    flows = [-build] + [benefit_per_year - run_per_year] * years
    npv = sum(f / (1 + rate) ** i for i, f in enumerate(flows))
    payback_months = build / ((benefit_per_year - run_per_year) / 12)
    return round(npv), round(payback_months, 1)

print(roi(build=1_000_000, run_per_year=300_000, benefit_per_year=900_000))

Output:

(492111, 20.0)

Stress-test the benefit

Re-run the model with the benefit cut by 30% and the cost raised by 30%. If the project only works under the optimistic numbers, it is fragile.

Quick check: What does payback period measure?

• How fast the model replies
• How long until net benefits cover the upfront cost
• The number of users
• The size of the dataset