AI ROI Calculation: How to Compute the Real Return on Investment

Most AI projects don't fail because of the technology. They fail because someone walked into the kickoff with a spreadsheet full of wishful numbers. Here is how to calculate AI ROI honestly – with the formula, three worked examples and the hidden costs vendor decks leave out.

You're staring at a quote for €28,000 for an AI project. The vendor promises "300% ROI in the first year." You sign, the project goes live – and nine months later the monthly cloud bill exceeds the savings. What went wrong?

Answer: the calculation was never honest. Return on investment for AI projects gets systematically inflated, because vendors low-ball setup costs, ignore ongoing costs, and multiply savings by a CEO's hourly rate. Once you understand how that works, you can sanity-check any AI proposal in 15 minutes.

Cost Components: What's actually coming your way

Before you calculate savings, you need a complete list of costs. Five blocks – the second and fifth are the most often forgotten.

1. Setup costs (one-time)

Everything you pay before the system goes live: design, development, integrations, data prep, testing. For a mid-market AI project plan on €5,000 to €30,000. Simple voicebots at the low end, workflow automation with three integrations at the high end.

2. Licence and cloud costs (ongoing)

AI models charge per request. Voice minutes per second. Hosting per month. For 1,000 calls/month on a GPT-based voicebot you quickly pay €150 to €400 just for AI calls, plus €80 to €200 for telephony and hosting. Grow, and you pay more.

3. Integration costs

AI has to write to your CRM, read your calendar, query your ERP. Budget €1,500 to €4,000 per integration for the initial setup, plus annual maintenance whenever the vendor changes their API (more often than you'd think).

4. Maintenance & ongoing development

An AI system is not a toaster. Budget 15 to 25% of setup cost per year for maintenance. Skip it, and after 18 months you have a system nobody understands that breaks every time reality changes (new products, opening hours, staff).

5. Team training & change management

The most expensive blind spot. Plan on 4 to 12 hours per employee for initial onboarding, plus 1 to 2 hours per quarter for updates. For a 10-person team easily 80 to 150 hours in year one.

Savings Components: What you realistically gain

The other side of the equation. Equally important: don't multiply savings by your most expensive hourly rate, and don't count effects that were already in place before AI.

1. Time savings

Hours your team no longer spends on routine work. Realistic: AI handles 50 to 80% of a task – the rest stays human. Multiply hours saved by the real loaded hourly cost, not the price you charge clients.

2. Conversion uplift

When your chatbot or voicebot turns more inquiries into orders. Rule of thumb: 5 to 15% uplift when the system removes a clear bottleneck (e.g. missed calls outside business hours). More only with data to back it up.

3. Error reduction

Fewer typos on data transfers, no forgotten callbacks. For processes where mistakes cost real money (wrong orders, missed deadlines), this is a hard euro figure. Calculate conservatively: how much did one missed deadline cost last year, how often did it happen?

4. Scalability without hiring

The most underestimated lever. With AI you can handle 30 to 50% more volume with the same headcount. Only visible after month 6, but long-term often the biggest. Include conservatively, or the calculation will look unserious.

The ROI Formula: Simple enough to do in your head

With your cost and savings lists, the formula is straightforward:

That covers year one and ignores ongoing costs. For an honest calculation, set monthly savings net – after deducting ongoing costs:

For multi-year projects, the 36-month view is more honest – setup costs are one-off while savings flow year after year:

What counts as a "good" ROI? For SMB AI projects we typically see 150 to 400% ROI over 12 months, with break-even between month 6 and 18. Higher than that is rare; lower usually means the use case is too narrow or the setup costs are too high.

3 Worked Examples: What honest numbers look like

Hidden Costs you mustn't forget

Five line items that show up in almost no proposal – and can wreck your ROI:

• Data cleanup before the project: Messy data (different Excel versions, old CRM leftovers, PDFs without OCR) often costs 20 to 40% of the setup. Plan for it before the project starts.
• Onboarding opportunity cost: 2 to 4 weeks of reduced output on day-to-day work – a real loss.
• Vendor lock-in: Migration after 2 years easily costs 30 to 60% of the original setup.
• Compliance & data protection: PIAs, DPAs, EU-based servers. For complex projects €1,500 to €5,000 one-off, plus annual audits.
• Hidden price increases: OpenAI, Anthropic, cloud providers – all have adjusted multiple times. Plan on 10 to 20% annual cost growth on ongoing costs.

When ROI is not the right metric

Not every AI project can be justified by ROI alone. Four situations where you shouldn't hide behind numbers:

• Learning projects: The first project is partly training. Lower ROI, but the experience makes every subsequent project much cheaper.
• Strategic competitiveness: If competitors use AI and you don't, long-term you lose customers. Risk argument, not a savings argument.
• Compliance & mandates: Accessibility, EU AI Act, industry standards. The metric is compliance, not ROI.
• Employee retention: If your accountant quits because of manual receipt capture, replacement costs €15,000 to €30,000. These "soft returns" are real but rarely in ROI spreadsheets.

At Cogswell IT we discuss before every proposal which case applies. Some projects we sell as "strategic," not "profitable" – and the customer knows exactly what they're buying.

What's your takeaway?

One rule: For every AI proposal, ask whether ongoing costs are in the ROI calculation. Which hourly rate was used for savings? Is cloud cost scaling factored in? If the vendor turns pale – you know enough. More on our approach is on our services page.