The ROI of an AI automation is not a slogan, it is a subtraction. Annual gains minus total costs, divided by the capital committed. The problem is that half of all projects overstate gains and forget two thirds of the costs. The result: a headline ROI of +300% that turns negative after eighteen months.

The formula, no cheating

Start simple: ROI = (net annual gains - total cost of ownership) / total cost of ownership. But the word that matters is total. The cost of an AI agent is never just a model subscription. You must add integration, maintenance, human oversight and the cost of errors.

  • Build cost: design, integration with existing tools, testing (often 40 to 60% of the first-year budget)
  • Usage cost: API/model calls, infrastructure, licenses (recurring, often underestimated 3x)
  • Oversight cost: human time spent checking output, especially the first six months
  • Error cost: manual rework, customer incidents, eroded trust
  • Change cost: team training, change management, internal resistance

Measure gains like an accountant, not a salesperson

A credible gain rests on a baseline measured before the project, not estimated after. If you automate quote processing, time the current process: how many minutes per quote, how many quotes per month, what loaded hourly cost. That is your zero point. Without a baseline, every gain figure is just an opinion.

Time saved that turns into neither revenue nor avoided cost is not a gain: it is free time.

Beware the phantom time-saved trap. Saving 5 minutes for ten people only creates value if that time is reallocated to productive work or removes a hiring need. Otherwise it is comfort, not ROI. Always ask: does the freed-up time fund something concrete?

Payback: your best safeguard

Rather than chasing a hard-to-defend three-year ROI, think in payback: in how many months does the project pay for itself? For an SMB our trigger threshold is clear: a payback under 12 months is healthy, 12 to 18 months is debatable, beyond 24 months we challenge or drop it. A project with a 6-month payback survives even if your estimates were 30% too optimistic.

  • Payback < 12 months: green light, the project absorbs surprises
  • Payback 12-18 months: conditional, demand a solid baseline
  • Payback 18-24 months: rescope and shrink the perimeter
  • Payback > 24 months: do not build, find another lever

Three honest, numbered examples

Case 1, inbound email triage: EUR 6,000 build, EUR 200/month usage, saving 25 h/month at EUR 35 loaded. Annual gain ~EUR 10,500, year-1 cost ~EUR 8,400. Payback ~9 months. Good project. Case 2, custom support chatbot: EUR 40,000 build, heavy oversight, real resolution rate 45% rather than 80%. Real payback: 31 months. We rescope. The difference between the two is not the technology, it is the honesty of the math.

Challenge first, then build: if the ROI does not hold on a sober spreadsheet, it will not hold in production.

Our rule at NexusOS: no project starts without an ROI calculation that survives a pessimistic scenario. If you want us to challenge yours before you commit a single euro, write to contact@nexus-os.fr. Better to kill a bad project on a spreadsheet than on an invoice.