How Do You Calculate the ROI of an AI Investment?
The short answer
The return on an AI investment is calculated as: ROI = (Value gained − Total cost) ÷ Total cost × 100. Value gained has three components: hours saved × the true cost of those hours, additional revenue from being faster, and the cost of errors avoided. Total cost is subscriptions + internal time spent on setup + training + the ongoing review burden. The critical part is counting the internal setup time as cost; most calculations skip it and therefore produce fantasy numbers. A properly scoped first project typically breaks even in 2–4 months.
Why most AI ROI calculations are wrong
I see two kinds of bad math. The first is too optimistic: "We saved 40 hours a month at $20 an hour, so that's $800 in profit." This ignores the weeks spent on setup and the ongoing burden of reviewing output.
The second is too cautious: looking only at the subscription and saying "we pay $200 a month and it's unclear what we get". That happens when nobody tracks where the saved time went — the gain exists but stays invisible.
I paid the price of not measuring in my first startup. At Edvays — the online consulting platform I launched before the pandemic — we hosted over 20,000 minutes of calls in the first six months, and it made the press. But I'll admit it: back then I couldn't cleanly measure which channel brought how many calls at what cost. The number looked great and I still couldn't make decisions. Today the first thing I do on any project, before building anything, is open the measurement sheet.
The formula and its line items
The formula is simple; the difficulty is filling in the line items honestly:
ROI (%) = (Value gained − Total cost) ÷ Total cost × 100
| Line item | How to calculate | Common mistake |
|---|---|---|
| Hours saved | (Old duration − New duration) × monthly frequency | Not adding review time to the new duration |
| True cost of an hour | Gross salary + benefits + overhead ÷ hours worked | Using take-home pay (true cost is ~1.5–2×) |
| Revenue from speed | Extra deals closed × average deal size × margin | Using revenue instead of margin |
| Errors avoided | Errors in past 12 months × average remediation cost | Guessing instead of checking history |
| Subscription cost | Per-seat price × users + automation tools | Annual/monthly mix-ups, unused seats |
| Setup cost (internal time) | Hours spent × those people's hourly cost | Skipping it entirely — the biggest error |
| Ongoing review burden | Review time per output × monthly volume | Dismissing it as "we'd look anyway" |
Leave one item out entirely: "the strategic advantage we'll gain in future". It may well be real, but it isn't measurable — and an unmeasurable line turns the whole calculation into fiction. Keep that expectation outside the math, as a separate sentence.
A realistic example: a 12-person services company
The numbers below are the magnitudes I typically see in a team this size. When you plug in your own, watch the ratios rather than the absolute figures.
| Line item | Calculation | Monthly ($) |
|---|---|---|
| Quote preparation savings | 30 quotes × 45 min = 22.5 hrs × $25 | +562 |
| Meeting notes savings | 40 meetings × 25 min = 16.6 hrs × $25 | +415 |
| Content production savings | 24 hrs × $20 | +480 |
| Faster quotes → extra deals | 1 extra deal × $2,000 × 35% margin | +700 |
| Subscriptions | 12 seats × $25 + automation $50 | −350 |
| Setup (amortised over 3 months) | 60 hrs × $30 ÷ 3 months | −600 |
| Review burden | 12 hrs/month × $25 | −300 |
| NET (first 3 months) | 2,157 − 1,250 | +907 · 73% ROI |
| NET (from month 4) | Once setup cost is amortised | +1,507 · 232% ROI |
Note the shape: 73% ROI in the first three months, 232% after. Judging the project in month two and concluding "this wasn't what I expected" is the most common way people quit early. Setup is paid once; the gain repeats every month.
At Gurizon we count the total time we've saved clients: over 14,200 hours today. I didn't start tracking it for marketing — I tracked it to decide what we should automate next. Later I realised showing that sheet to a client is more persuasive than any deck. The moment you measure, the argument ends; until you do, everyone's opinion looks equally valid.
Three things you can't measure but shouldn't ignore
- Speed advantage. Getting your quote in a day before a competitor doesn't fit in the sheet, but it moves your close rate.
- Team morale. Freeing people from the work they hate is hard to measure but shows up in retention.
- Institutional learning. What you learn on the first project makes the second one three times faster to build. It doesn't belong in this month's ROI, but it lowers the next project's cost.
Don't blend these into the math; list them as three bullets underneath it. In a board meeting — or a conversation with your co-founder — that separation is what buys you credibility.
When should you say "this investment didn't work"?
- If by the end of month three you still can't measure hours saved — the problem isn't the tool, it's the process you chose.
- If the review burden eats more than half the time saved — output quality isn't good enough for your work; pull that process back.
- If the system stops when the person who built it leaves — it was never a system, just one person's habit.
- If six months in only one person is using it — it isn't producing value for the team.
If any of those four apply, shut the process down rather than swapping the tool. Changing tools doesn't rescue a badly chosen process; it just grows the bill.
Key takeaways
- ROI = (Value gained − Total cost) ÷ Total cost × 100. The formula isn't hard; filling it in honestly is.
- Count internal setup time as cost — the most skipped line and the one that most distorts the result.
- The true cost of an hour isn't take-home pay; with benefits and overhead it's 1.5–2× that.
- The first three months look weak and the rest look strong — don't judge the project in month two.
- If review burden eats half the gain, shut the process down; swapping tools won't fix it.
Frequently asked
How long does a typical AI project take to pay for itself?
For a well-chosen, repetitive, low-risk process: 2–4 months. For a custom setup over your own data (RAG, integration projects) it stretches to 6–12 months. For anything beyond six months, I'd re-ask whether you picked the right process before re-running the math.
What tool do I need to measure the time saved?
None. Three columns in a spreadsheet will do: task, old duration, new duration. Fill it in by hand for two weeks. Buying time-tracking software is one of the ways people postpone starting to measure — I've watched it happen repeatedly.
If AI prices rise, does my calculation break?
Per-unit model costs have trended downward in recent years, but run the scenario where subscription costs double anyway. If your case only stands up at today's prices, the time you're saving isn't big enough. A solid project stays positive even at 2× cost.
How should I present this to a board?
One page: the formula and line items at the top, three months of actuals in the middle, the unmeasurables as three bullets at the bottom. Separating out what you can't measure increases trust in everything else. And bring real data from a two-week pilot rather than a projection — projections get debated, data doesn't.