How to Calculate the ROI of an AI Project (With a Worked Example)

The ROI of an AI project is annual benefit minus annual running cost, divided by the build cost — and a project worth doing usually pays back its build cost within 6 to 18 months. The hard part is being honest about all three numbers. Here is the framework, the costs people forget, and a worked example you can adapt.

Step 1 — Quantify the benefit (four types)

• Time saved — hours per week returned to the team, valued at the loaded hourly cost. The most common and easiest to measure.
• Cost avoided — headcount you don't need to add, or external spend (agency, overtime) removed.
• Revenue gained — faster response times that win deals, higher conversion, capacity to serve more customers.
• Risk reduced — fewer compliance incidents, fewer errors, fewer escalations. Real but harder to price; estimate conservatively.

Count only the benefits you can defend with a number. A vague 'better customer experience' doesn't belong in the model — if it matters, find the metric that proves it (CSAT, retention, repeat rate).

Step 2 — Count the full cost

• Build cost — the one-time fixed-price project (typically €15,000–€80,000 for a production system).
• Running cost — LLM API usage, infrastructure, vector DB, observability (€500–€2,500/month for most systems).
• Internal cost — the time your team spends on discovery, data preparation, review, and adoption. Real, and routinely forgotten.
• Maintenance — a retainer or internal owner's time to keep it working as models change.

Step 3 — The payback formula

Payback period (months) = build cost ÷ (monthly benefit − monthly running cost). First-year ROI = (annual benefit − annual running cost − build cost) ÷ build cost. If payback is under 12 months, the project is usually a clear yes; 12–18 months is reasonable for a strategic capability; beyond 24 months, re-scope or drop it.

A worked example

A 40-person company automates first-line customer support with an AI system. Benefit: it handles 70% of tickets, freeing two support agents' worth of time on repetitive work — roughly €5,500/month in loaded cost returned to higher-value work. Build cost: €28,000. Running cost: €900/month. Net monthly benefit: €5,500 − €900 = €4,600. Payback: €28,000 ÷ €4,600 ≈ 6.1 months. First-year ROI: (€55,200 annual benefit − €10,800 annual run − €28,000 build) ÷ €28,000 ≈ 59%.

“If you can't write the ROI on one line, you're not ready to build — and that's useful information, not a failure. The projects worth doing make the number obvious.”

Two honest caveats

First, don't model 100% of the benefit on day one — adoption ramps, so assume a 2–3 month curve to full value. Second, the biggest risk to the ROI is not the model; it's a project that ships late or doesn't ship at all. A fixed-price scope with a defined go-live protects the denominator. Build the ROI case before the project, measure against it after launch with a proper eval suite, and you turn 'we think AI helped' into a number you can defend to the board.