HomeGuidesHow to calculate AI ROI — the honest guide
Advertisement
Business guide · ROI and business case

How to calculate AI ROI — the honest guide

Most AI ROI calculators give you a number. This guide explains what that number actually means, how to make it defensible, and why ranges are more credible than single-point projections.

By Sarah KendrickPublished April 202610 min read

The problem with most AI ROI calculations

Most AI ROI calculations are wrong in the same direction: they're too optimistic, built on time savings that don't materialise because adoption is lower than projected, and presented as precise numbers when they're really educated guesses dressed up in a spreadsheet.

The CFOs and operations directors we've spoken to are increasingly sceptical of AI ROI projections — not because ROI doesn't exist, but because they've seen too many projections that didn't hold up. The way to build credibility is counterintuitive: be more honest about uncertainty, not less.

The most important thing

A credible AI ROI calculation doesn't tell your leadership what AI will save. It tells them what AI needs to be true for the investment to pay off — and then lets them assess whether those conditions are achievable.

Getting your inputs right

Every AI ROI calculation depends on two variables that people routinely get wrong: time savings estimates and adoption rate. Here's how to approach each honestly.

1
Measure actual time, don't estimate it
Most people estimate how much time AI will save based on how much time they think they spend on tasks AI will help with. Both numbers are wrong. Actual time-tracking studies consistently show people underestimate time on routine tasks and overestimate time on strategic work. Before projecting savings, track actual time spent on the specific tasks AI will help with for one week. Use a simple time log — nothing sophisticated. The resulting numbers will be meaningfully different from your estimates.
2
Use a conservative adoption rate
Adoption is the variable that kills most AI ROI projections. Teams don't use AI tools as consistently as projections assume. A reasonable baseline for a new AI tool: 60% of intended users using it actively after 90 days. Not the 100% your calculation might assume. Build your projection on 60% and treat everything above that as upside.
3
Define what 'saving time' means for the business
Time saved on AI-assisted tasks only produces ROI if that time is redirected to higher-value work. If a marketing manager spends 2 fewer hours per week on draft emails because Claude handles first drafts — that 2 hours only creates value if it's redirected to strategy, client time, or work that drives revenue. Be specific about what the recaptured time will actually be used for.

Industry benchmarks to calibrate against

Rather than letting teams guess at time savings, use published benchmarks as starting points. These are from published research and should be treated as central estimates, not guarantees:

  • Marketing content creation: 30-45% time reduction on first-draft writing tasks (McKinsey, 2024)
  • Software development: 55% faster on certain coding tasks with AI assistance (GitHub, 2023); our own testing shows 40-60% on multi-file tasks with Cursor
  • Customer support: 14% productivity increase for agents using AI assistance (NBER, 2023); Fin AI autonomously resolves 54-71% of tickets in our real deployments
  • Data analysis and reporting: 25-40% time reduction on report generation tasks
  • Email and communication: 30-40% reduction in drafting time for complex responses

The important caveat: these are averages across many deployments. Your actual results will depend on implementation quality, adoption rate, and task specificity. They're useful for calibrating estimates, not for projecting outcomes.

Why you should present ranges, not single numbers

Single-point ROI projections are either wrong on the day you present them or wrong within six months. Ranges are more honest and, counterintuitively, more persuasive with financially sophisticated decision-makers.

1
Build three scenarios
Conservative: 50% of central adoption rate, 60% of projected time savings, tool costs at current pricing. Central: your best estimate. Optimistic: 80% adoption, full time savings, no pricing increases.
2
Calculate break-even, not ROI
The most useful number is not 'we'll save $X.' It's 'this pays back in Y weeks if the team uses it Z hours per week.' That framing lets decision-makers assess the investment against their own confidence in the conditions.
3
State your assumptions explicitly
What adoption rate are you assuming? What salary are you using? What time saving per task? Stating assumptions explicitly is not a weakness — it demonstrates rigor and makes your projection trustworthy.
Use our ROI calculator

Our business AI ROI calculator walks through this methodology for your specific departments and tools, generates conservative/central/optimistic ranges, and exports a presentation-ready summary. Try it free →

How to present AI ROI to leadership

Framing matters as much as the numbers. The most successful AI investment cases we've seen share a common structure:

  • Open with the business problem, not the technology
  • Present the break-even condition before the projected upside
  • Show the conservative scenario prominently, not buried
  • Include a pilot plan that generates real data within 30 days
  • Commit to a review date with actual vs projected comparison

Tracking actual vs projected

The most important step most organisations skip: returning after 90 days to compare actual outcomes against projections. This does two things. First, it improves future projections by grounding them in real adoption and time saving data. Second, it demonstrates accountability — which is the foundation of getting future AI investments approved more easily.

A simple tracking template: weekly time log for the same tasks from your original estimate, adoption rate from tool usage data, time redirected to higher-value work (qualitative, from manager check-ins). Compare to your conservative projection monthly for the first quarter.

Frequently asked questions

What's a realistic ROI for AI tools in year one?
Honestly: wide range. Customer support tools like Fin AI with good knowledge bases can achieve payback in weeks. General AI assistants for knowledge workers typically show meaningful productivity gains in 90-180 days with good adoption. The variance comes from adoption rate more than anything else.
Should I include cost of implementation in my ROI calculation?
Yes, always. Implementation time, training time, process redesign time, and ongoing tool administration are real costs. Many ROI projections undercount these significantly, which is why projections disappoint.
How do I handle it when actual results differ from projections?
Investigate adoption first — lower adoption than projected explains most shortfalls. Then look at task specificity — AI works best on high-repetition, moderately-complex tasks. It delivers less value on highly novel or relationship-dependent work. Use the gap to recalibrate your next projection.