Estimating AI ROI and cost before you commit
Measure the current process first: volume, time per item, error rate and cost. Estimate value as the realistic change in those numbers, not the best case, and count all costs: model usage, infrastructure, build effort, ongoing engineering, human review and change management. Keep estimated value and measured results in separate columns, and remember that time freed isn't money saved unless it's used for something else.
Why business cases mislead
Business cases often assume full automation, no review and instant adoption, so every project looks profitable on paper. Cheaper AI also tends to lead to more usage, not smaller bills, so estimate usage growth as well as unit cost.
Value
Real business impact falls into four groups: revenue, cost reduction, strategic position and risk reduction. For each use case, ask three questions: what is it meant to do, how do you know it does it, and which business numbers (cycle time, cost per case, churn, error rate) move as a result.
Formulas
monthly_value = volume * time_saved_per_item * loaded_hourly_cost * adoption_rate
+ error_reduction_value
monthly_cost = model_and_infra + engineering_upkeep + human_review + licences
payback_months = build_cost / (monthly_value - monthly_cost)
Use the realistic adoption rate, not 100%, and include review time for outputs a person must check.
Count the full cost
The visible cost is the model and cloud bill. The real cost adds engineering hours at their loaded rate, vendor and contractor fees, and the time of the domain experts who label and review.
true_annual_ai_spend = compute + engineering_hours * loaded_rate + vendor_fees + contractor_fees
A worked example
A team handles 4,000 documents a month at 20 minutes each, at a loaded cost of $40 an hour, so the process costs about $53,000 a month. If AI cuts handling to 8 minutes with review, at 80% adoption, the value is about $25,600 a month. If the system costs $6,000 a month to run and maintain and $90,000 to build, payback is under 5 months. These are illustrative numbers, so replace each one with your own measured baseline.
Estimated vs measured
Keep two columns in every business case: what you estimated before launch, and what you measured at 3 and 6 months. The gap between them is how you get better at estimating, and it's what a board will trust.
From the floor
In the AI foundations and business alignment session for leaders, I said even a 1 or 2% improvement can matter: in a supply chain problem we worked on, the client had 200 vehicles, and optimising routes so that two fewer vehicles were needed reduced their cost for a long period, which is the kind of return you plan over a year, not a quarter.
In the AI leadership cohort, I put value ahead of cost: a client may happily create budget for an AI product, but if only two customers end up paying for it, the value isn't there however good the product is.