How to Calculate ROI on an AI Project Before You Spend Anything

Most ROI calculations for AI projects are produced by the vendor selling the project. Here is how to do your own, including the parts that make the number smaller.

Step 1: Measure the current cost properly

Not estimate — measure. Ask whoever does the work to time it for two weeks.

Direct cost is instances per month, multiplied by minutes per instance, multiplied by loaded hourly rate divided by sixty. Loaded means salary plus superannuation plus overhead: roughly 1.3 to 1.5 times base.

For a process running 200 times a month at 20 minutes, done by someone on a loaded rate of $55:

200 × 20 ÷ 60 × $55 = $3,667 per month, or about $44,000 a year.

Then add what the direct cost misses:

• Error cost. How often does it go wrong, and what does fixing it cost?
• Delay cost. If work waits, what does the wait cost? For quoting, slower response means lower conversion, and that is usually the largest number on the page.
• Opportunity cost. What would that person do instead? If the honest answer is "leave on time", that is real but it is not revenue.

Be disciplined here. Inflating opportunity cost is the most common way these calculations become fiction.

Step 2: Estimate the realistic saving

Automation rarely removes a process entirely. It removes most of it.

A reasonable planning assumption is 60% to 80% of instances fully handled, with the remainder still needing a person, plus review time on the automated ones.

Same example at 70% automation:

• 140 instances handled: saves $2,567/month
• 60 still manual: costs $1,100/month
• Review time on the 140, say 2 minutes each: costs $257/month

Net saving: about $2,310 per month, or $27,700 a year. Not the $44,000 the naive calculation suggests.

Step 3: Count the full cost

The bottom three are where vendor calculations go quiet. Ongoing cost typically runs 10–20% of build cost annually. A system with no maintenance budget degrades within about three years.

Step 4: Work out payback

Build $25,000, ongoing $800/month, saving $2,310/month.

Net monthly benefit: $1,510. Payback: $25,000 ÷ $1,510 = about 17 months.

How to read that:

• Under 12 months — strong, proceed
• 12 to 24 months — reasonable if the process is stable
• 24 to 36 months — marginal; requirements may change first
• Over 36 months — wrong process, look elsewhere

The three ways these get inflated

Assuming 100% automation. Nothing automates completely. If a proposal assumes it does, the author has not built one.

Counting headcount reduction that will not happen. The most common inflation. AI removes tasks, not people. If the business case needs a redundancy, be sceptical — that saving usually fails to appear.

Ignoring maintenance. A build quoted with no ongoing cost is not cheaper. It is incomplete.

The number that changes everything

For customer-facing processes, response time often dwarfs labour saving.

If quoting takes three days and automation takes it to same-day, and same-day quotes convert at 40% against 25% for three-day quotes, then on 60 quotes a month at $2,000 average:

• Before: 15 jobs, $30,000
• After: 24 jobs, $48,000

$18,000 a month in additional revenue — six times the labour saving.

This is real but requires honesty. Do you know your conversion rate by response time? If not, you are guessing, and you should say so in the calculation rather than presenting a guess as a projection.

A simple test

If the payback depends on the most optimistic assumption in your model, it is not a good enough case.

Rerun with the pessimistic version: 50% automation instead of 70%, ongoing cost 20% higher, no conversion improvement. If it still pays back inside three years, the case is genuine. If it only works optimistically, wait — either for the process to stabilise or for the cost to come down.

Both happen more often than people expect.