AI ROI calculator, a person versus a model
One task. Walk through what changes, the scoreboard keeps count.
Your numbers
With taxes and overhead the hour costs €16.25.
Advanced · minutes per task, build, hosting, tokens
Since you opened this page, the task has cost
The same day, two ways
At 4,000 tickets a month, that is 192 a working day. A person spends 8 minutes on each.
Each square is one ticket. The gold sliver is the model writing all 192 answers, about 6 seconds each. The hatched part is the person reading 20 of every 100 before they go out.
Twelve months, the build paid once
The build pays for itself in 17 days. After that every month keeps €7,928. Twelve months, €90,641.
Out of 100 answers, how many does a person read?
Click a square. This is the honest number in the whole calculation, and it costs more than the model does.
Honestly, it barely matters
Seven models at list price, per ticket with the same checking. Cheapest to dearest, the whole price list fits in €0.0138. The SaaS agent billed per resolution is 5.2× away.
With Claude Sonnet 5.5 the tokens are €0.0071 of the €0.17 per ticket. The rest is the person reading. Pick the model for quality, decide the checking for cost.
Your inputs and published prices. Nothing else.
8 min / 60 × €16.25 = €2.17(1,200 × €1.84 + 400 × €9.20) / 1M × 1.2 = €0.007120% × 3 min / 60 × €16.25 = €0.16(4,000 × (€2.17 − €0.17) − €60) × 12 − €4,500 = €90,641All formulas, model prices and task defaults
Every formula with your values
8 min / 60 × €16.25 = €2.17(1,200 × €1.84 + 400 × €9.20) / 1M × 1.2 = €0.007120% × 3 min / 60 × €16.25 = €0.16(4,000 × (€2.17 − €0.17) − €60) × 12 − €4,500 = €90,641€4,500 / €7,928 a month = 17 days(€4,500 / 12 + €60) / (€2.17 − €0.17) = 218 a monthPublished list prices, USD per million tokens
Prices as of 2026-10-21. Copied from the vendors' pricing pages on that date. Check the linked page before you sign anything.
| Model | Input | Output | Source |
|---|---|---|---|
| Claude Haiku 5.5Anthropic | $0.1 | $0.5 | Pricing page ↗ |
| Claude Sonnet 5.5Anthropic | $2 | $10 | Pricing page ↗ |
| Claude Opus 5.5Anthropic | $4 | $20 | Pricing page ↗ |
| Gemini 3.8 FlashGoogle | $0.75 | $3.75 | Pricing page ↗ |
| Gemini 3.1 ProGoogle | $2 | $12 | Pricing page ↗ |
| GPT-6 LunaOpenAI | $0.1 | $0.5 | Pricing page ↗ |
| GPT-6 SolOpenAI | $2 | $10 | Pricing page ↗ |
| Intercom Fin (per resolution)SaaS, per resolution | $0.99 / task | Pricing page ↗ | |
What each task assumes
| Task | Minutes | Tokens in / out | Checked | Build | Saving a year |
|---|---|---|---|---|---|
| Support ticket reply | 8 | 1,200 / 400 | 20% × 3 min | €4,500 | €90,641 |
| IT request (reset, access) | 12 | 900 / 300 | 15% × 2 min | €5,000 | €51,222 |
| Lead research and qualification | 15 | 2,500 / 600 | 30% × 4 min | €3,500 | €31,543 |
| Sales email or proposal draft | 40 | 3,000 / 1,200 | 100% × 8 min | €3,000 | €27,408 |
| Invoice parsing and entry | 6 | 2,000 / 500 | 25% × 2 min | €6,000 | €46,547 |
| Email classification and routing | 2 | 600 / 100 | 5% × 1 min | €2,500 | €47,247 |
| RFP / security questionnaire answer | 25 | 4,000 / 800 | 60% × 6 min | €5,000 | €8,148 |
| Payment reminder and collections touch | 6 | 1,500 / 300 | 30% × 2 min | €4,000 | €21,486 |
| Your own task | 10 | 1,500 / 500 | 30% × 3 min | €4,000 | €24,749 |
Working estimates from projects delivered in 2025 and 2026, not benchmarks. Every number is editable above. The saving uses your salary, currency and model.
Email me this estimate
One email with the figures and the link to this exact calculation. No newsletter.
Questions
How do you calculate AI ROI?
Take what the task costs with a person per month, subtract what it costs with the model, the person checking and hosting. Multiply by 12 and subtract the one-time build. That is the saving in year one. Divide the build by the monthly difference and you get the payback. This page does exactly that with your numbers and shows every formula with the values substituted.
What is a good ROI for an AI project?
Payback inside six months and a first-year saving of at least three times the build are a healthy result for a repetitive task with a clear monthly volume. Support and back office tasks at a few thousand items a month usually pay back in weeks. Below that line, check the share a person reads and the volume before you sign anything.
Why is there a human review line at all?
Because a model without review is not a process anyone signs off. The review share is the honest number in the model: at 20 percent one output in five is read by a person, at 100 percent every one is. Raise it when mistakes are expensive and the AI cost rises accordingly.
Where do the task defaults come from?
They are our working estimates from projects delivered in 2025 and 2026: how long a person takes, how many tokens a task uses and how much of the output still gets reviewed. They are not benchmarks. Every default is editable in the advanced mode.
What does break-even volume mean?
The monthly number of tasks from which AI is cheaper than the person once the one-time build is spread over a year and the monthly infrastructure is paid. Below that volume you are better off leaving the task with a person.
Are the model prices current?
They are copied from the vendors' pricing pages on the date shown, with a link to each page. When the date is more than two months old the page tells you to check. Prices fall more often than they rise, so the estimate tends to be conservative.
Can I get a tailored estimate?
Yes. Send us the link with your inputs or book a short call and we will pressure-test the minutes, tokens and review share against your actual process, then come back with a fixed scope and price.
Ready scenarios
Turn the estimate into a fixed price
We check the minutes, tokens and the share a person reads against your real process, then send a fixed scope and price.