Calculating AI ROI Without Lying to Yourself

AI ROI decks are almost always optimistic. The pattern is the same: a vendor counts every hour of "time saved" as cash, ignores adoption drop-off, and forgets the cost of running the model.

The 4-line model

Gross value   = hours_saved × loaded_cost_per_hour × adoption_rate
Run cost      = licences + infra + support
Change cost   = one-time rollout + training
Net annual    = Gross value − Run cost − (Change cost / payback_years)

Why each line matters

• adoption_rate: a tool used by 40% of the team is worth 40% of the saving. Assume <70% unless you have proof.
• loaded_cost_per_hour: salary + benefits + opportunity. Not just CTC.
• Run cost: includes the API/inference cost, which is often the surprise.
• Change cost amortised: rollout is not free. Spread it across the expected useful life (we use 2 years).

A worked example

A 50-person support team saves 30 minutes per agent per day with an AI assistant.

• hours_saved/yr = 50 × 0.5 × 240 = 6,000
• loaded cost = ₹900/hr
• adoption = 60%
• Gross = 6,000 × 900 × 0.6 = ₹32.4L
• Run cost = ₹6L licences + ₹2L infra = ₹8L
• Change cost = ₹10L over 2 years = ₹5L/yr
• Net = ₹19.4L/yr

That is a real number you can defend to a CFO. "AI will save us crores" is not.