How to Calculate ROI on an AI Project Before You Build It

The hardest sentence to write in any AI proposal is the one with a number on it. "This project will save you roughly $X per year." Get it right, and you have a real conversation about whether to build. Get it wrong by a factor of three, and you either talk a client out of a project they should have done or, worse, into one they should not have touched.

We have done dozens of these estimates. Most of them have held up well. None of them used a sophisticated model. All of them used the same three levers — and a discipline about which ones to count and which ones to leave out.

This is the framework. It is deliberately simple, because the goal is decision-grade numbers, not auditable financials. You should be able to estimate the ROI of any candidate AI project on a single page, in under thirty minutes, before you write the first line of code.

The three levers (and only three)

Almost every business case for AI lives in one or more of three buckets. We name them deliberately, in order of how easily a buyer believes them.

1. Time saved × loaded wage. The most credible bucket. If AI removes a task that currently takes someone an hour a day, and that someone costs the company $70 an hour fully loaded, you have a defensible number. Multiply the time saved per occurrence, by the frequency, by the loaded hourly cost.

2. Error reduction × cost per error. Less obvious, often larger. Wrong invoices, missed appointments, mis-shipped orders, churned customers who should have been saved. Every error has a cost per occurrence — the rework, the refund, the lifetime value lost. If AI cuts the error rate from 6% to 2%, that 4-point delta times the cost per error times the volume is your savings.

3. Revenue lift × margin. The biggest bucket, also the easiest to oversell. Better recommendations sell more, faster ETAs retain more customers, better lead qualification closes more deals. Treat this one with suspicion. Always discount it by at least 50% for your first estimate. Sales and retention have too many confounding variables to claim a tight number early.

If you cannot fit your project's value into one of those three buckets, you do not have a business case. You have a hope.

The five inputs you need

For each bucket you are counting, you need at most five numbers.

1. Frequency — how many times does this thing happen? Per day, per week, per month — pick one and stick with it.
2. Time or cost per occurrence today — how long does it currently take, or what does each occurrence cost?
3. AI impact percent — realistically, how much of that does the AI remove? Start at 50% and go down, never up.
4. Loaded hourly cost — for time savings, what does an hour of this person cost the company, fully loaded? (Salary, benefits, overhead, typically 1.4× base.)
5. Margin percent — for revenue lift, what percent of new revenue actually becomes profit?

Five numbers, three buckets, one page. That is it.

The honest part: what to subtract

A serious estimate does not just sum the savings. It subtracts three things.

• Build cost — what the project costs to design, build, and ship. This is one-time.
• Annual run cost — model usage, infrastructure, monitoring, the partial FTE who owns it.
• Adoption drag — the ramp before the team actually uses it consistently. Realistically, take the first six months at 50% of the projected savings, then full from month seven onward.

If the project does not have a positive 18-month return after all three subtractions, do not build it yet. Either the scope is wrong, or it is the wrong problem to target first.

A worked example

Let us walk through a real-shape example — the kind we do on the back of a napkin during an audit.

The setup. A specialty distributor has six customer-service agents who spend roughly two hours a day each on a single task: looking up shipment status across three systems and emailing the customer a manual update. The team handles about 80 such inquiries a day. Loaded cost per agent: $40 per hour. Build cost for a custom shipment-status assistant integrated with their three systems: $65,000. Annual run cost: $9,000.

Step 1 — Time saved bucket

• Frequency: 80 inquiries per day × 250 working days = 20,000 inquiries per year.
• Time per occurrence today: 8 minutes (we asked).
• AI impact: 70%. The assistant drafts and sends about 70% of replies; agents handle the harder 30%.
• Loaded hourly cost: $40.

Math: 20,000 × 8 minutes × 70% = 112,000 minutes saved per year = 1,867 hours × $40 = $74,667 in time savings.

Step 2 — Error reduction bucket

We learn that about 4% of manual lookups currently include wrong info — the wrong tracking number, the wrong ETA. Each error costs roughly $25 in rework (a second email, occasional credit). The assistant cuts that error rate to 1%.

Math: 20,000 × (4% − 1%) × $25 = $15,000 in avoided rework.

Step 3 — Revenue lift bucket

Faster, more accurate shipment updates plausibly reduce churn. The team estimates a 1% lift on the $4.5M of revenue from the affected accounts, at a 25% margin. We discount that by 50% as a first estimate.

Math: $4,500,000 × 1% × 25% × 50% = $5,625 in conservative retention lift.

Step 4 — Sum and subtract

Gross annual benefit: $74,667 + $15,000 + $5,625 = $95,292 per year at steady state.

Year 1 adoption drag: count first 6 months at 50%. → ($95,292 / 2) × 50% + ($95,292 / 2) = $23,823 + $47,646 = $71,469 in year 1.

Year 1 net: $71,469 − $65,000 build − $9,000 run = −$2,531 in year 1. Year 2 net: $95,292 − $9,000 run = $86,292 in year 2. 18-month rolling: about $40,000 positive.

What this number tells you

You do not build this project to make money in year 1 — you build it because the steady-state run rate is roughly 10× the annual cost, and the team gets four hours a day back to spend on retention work that nobody currently has time for. That is the actual story to tell the CFO. The number is just the proof.

The five most common mistakes

Having watched a lot of these get done, the same five errors come up over and over.

1. Pricing AI savings at 100% of the task. AI almost never removes a task entirely. Start at 50%, justify upward case by case.
2. Forgetting the run cost. Token usage, monitoring, the human who owns it. Always there, always non-zero.
3. Counting revenue lift without a margin. Revenue is not profit. A 1% sales lift at 8% margin is not the same as at 40%.
4. Ignoring the adoption ramp. The team does not use it at full strength on day one. Six months at half savings is honest.
5. Estimating in isolation. The best ROI numbers come from a 30-minute conversation with the person who actually does the work. They will tell you the inputs you cannot guess.

Why this matters before the build, not after

The single biggest predictor of whether an AI project will be considered a success eighteen months later is whether a credible number existed before the build started. Without one, every stakeholder will tell a different story about whether it worked. With one, the conversation becomes: did we hit the number, and if not, why? That is a healthier conversation in every direction.

If the napkin math does not pencil out, that is also useful. You have saved the build cost. Move on to the next candidate use case.

---

If you want a second pair of eyes on a number you are about to put in a deck, that is exactly what a free 45-minute AI audit is built to do. Bring the rough estimate, we will pressure-test it together.

Request a free AI audit →