The return on an AI investment (ROI) is the comparison between the total cost of owning that investment over 12 months and the measurable gain it produces in the same period; for the calculation to hold up, where the gain will be read from has to be written down before the money is spent.
Most AI proposals state the cost as a single figure and the gain as a single sentence: "this much per month, and in return you save time." You cannot make a decision on those two pieces of information. The cost side is not just the subscription fee, and the gain side is only as real as what you actually measure.
This guide is built so that you can put the proposal in front of you into a table and assess it with your own numbers. The aim is not to convince you; it is to get you to a figure you can stand behind, whether you end up defending the investment or turning it down. There is a fictional worked example at the end, but first let us build the skeleton of the calculation.
How Do You Set Up the ROI Formula in Business Terms?
The classic formula is simple: ROI = (Net gain − Total cost) / Total cost. The problem is not the formula; it is that both variables get filled in with guesswork.
To make the calculation usable, tie the formula to three rules:
- Pick a period and do not change it. For AI investments, 12 months is a reasonable horizon. Setup cost lands in the first month while the gain usually only arrives after the third; look through a three-month window and almost every investment shows a loss.
- Only add up the gain items you can actually measure. The sentence "brand perception improves" has no place in the calculation. If you cannot measure it, write zero; your number gets worse, but it becomes trustworthy.
- Include your own team's time in the cost. Because no money leaves the business, this item is forgotten again and again, and it is the single biggest source of distortion in the calculation.
You also need a baseline: the numbers from the three months before the investment. How many enquiries came in, how many went unanswered, how many turned into sales, how many hours went into which job. Without that line you cannot prove what changed after the investment; you can only feel good about it.
Which Items Sit on the Cost Side?
Total cost of ownership (TCO) is broader than the figure on the invoice. The table below lists the items you run into in a typical AI setup, whether each one is one-off or ongoing, and the detail most often skipped when reading a proposal.
| Cost item | One-off / Ongoing | Most often skipped detail |
|---|---|---|
| Setup and configuration | One-off | The setup fee is usually quoted for a "standard scenario"; exceptions specific to your business are billed as extra line items |
| Software subscription / licence | Ongoing (monthly or annual) | Pricing that jumps a tier as user count, message volume or transaction count grows |
| Usage-based transaction fee | Ongoing, variable | In a busy month the bill can come in higher than the fixed subscription |
| Integration development | One-off (sometimes recurring) | If your existing software has no API (application interface), the cost multiplies several times over |
| Data preparation and cleaning | One-off, heavy | Merging scattered customer records is usually the longest step of the project |
| Team training | One-off + repeat | Recurring mini-sessions for new hires |
| Process change | One-off, hidden | Writing the new flow, dropping the old habit, the slowdown of the first few weeks |
| Maintenance and improvement | Ongoing | Monitoring answer quality, correcting faulty output |
| Internal labour (your own team) | Ongoing | The hours the project owner spends each month — uninvoiced but real |
| Compliance and data security | One-off + ongoing | Where data covered by KVKK (Turkey's personal data protection law) is processed, and updating your privacy notices |
The first seven items in the table get discussed in most proposals. The last three do not, and those are the ones that actually surprise the budget. If you want to see how service pricing is built up item by item, the guide to GEO cost line items does a similar breakdown on the service side.
How Do You Turn Invisible Costs into Numbers?
An invisible cost is not a cost that does not exist; it is simply a cost with no invoice attached. Four of them stand out.
The learning curve. In the first weeks the team uses the new tool slowly and does some jobs twice. A crude but workable assumption: accept that part of your productivity disappears in the first month and largely recovers in the second, and write that loss into the calculation in hours.
Process change. For the tool to be useful, the existing flow has to change. If your appointment form is changing, if customer records are landing somewhere new, someone has to design and write that. That someone is usually the business owner or the most senior person on the team — in other words, the most expensive hour you have.
Data cleaning. AI does not produce tidy results from messy data. If you start with a list where the same customer sits in three separate records and phone numbers are written in different formats, the first months of gain get pushed back. It also helps to decide which metrics you will track before doing this work; the customer data analysis guide for small businesses can help narrow down which fields are genuinely needed.
Integration. If two systems need to talk to each other, the cost depends on how open those systems are. In modern software this is a few days of work; in closed or very old software the project stops here. Before you take a proposal, ask whether your existing software supports data export and has an API.
What Do You Measure on the Gain Side, and from Which Data?
The gain side is harder than the cost side, because the numbers do not sit on a single invoice. For every item, the answer to "where will I read this number from" has to be settled in advance.
| Gain item | How it is measured | Which data it is read from |
|---|---|---|
| Time saved | (Hours spent before − hours spent now) × hourly cost | Task tracking records, or a simple time log kept for one week |
| Fewer missed enquiries | Drop in the number of unanswered enquiries × conversion rate × average customer value | Call logs, inbox, form submissions |
| Shorter response time | Average first response time and the conversion difference tied to it | CRM or messaging platform timestamps |
| Higher conversion | Change in the quote → sale rate × enquiry volume × average customer value | Sales records, quote files |
| Fewer errors and less rework | Drop in corrected records / returns / repeat jobs × unit cost of a correction | Return records, complaint log, accounting corrections |
| Average basket or job size | Comparison of the average invoice amount before and after the investment | Accounting or e-commerce report |
| Customer retention | Change in the repeat customer rate × average customer value | Sales history, repeat appointments |
| Visibility in AI search | Change in how often the brand name is cited in answers | Query tests repeated at regular intervals |
The last row needs particular care: turning a visibility gain directly into money is hard, and it is very easy to inflate with guesswork. Before you put this item into your ROI calculation, at least make it measurable; we set out how to measure AI visibility with five metrics in a separate guide. If you cannot measure it, the most honest approach is to write zero and keep the item aside as a bonus.
A Worked Example: A Scenario Built on Assumptions
The figures below are not a real client case; they are assumptions constructed to show how the formula works. Substitute your own numbers and the result will change completely.
Assumptions: a service business with 6 people, 120 incoming enquiries a month, 25 of which go unanswered because they arrive outside working hours. The enquiry-to-customer conversion rate is 20%, and the average customer value is 8,000 TL (Turkish lira). The office spends 10 hours a week filling in forms and entering records; the person doing that work costs the employer 300 TL per hour.
Annual cost assumption: setup and integration 60,000 TL (one-off), subscription 6,000 TL/month (72,000 TL), training and data cleaning 25,000 TL, 4 hours a month from the project owner (48 hours × 300 TL = 14,400 TL). Total: 171,400 TL.
Annual gain assumption: assume half of the unanswered enquiries are recovered — 12.5 enquiries a month × 20% conversion = 2.5 extra customers × 8,000 TL = 20,000 TL a month, 240,000 TL a year. Time saved: 6 hours a week × 52 weeks × 300 TL = 93,600 TL. But assume that in the first two months the learning curve stops that gain from materialising, and multiply the annual figure by 10/12: 278,000 TL.
On those assumptions, ROI = (278,000 − 171,400) / 171,400 ≈ 62%. Payback lands at the point where the monthly net contribution covers the setup burden, roughly around the fifth or sixth month.
The real message here is not the percentage but how fragile the calculation is: take the conversion rate as 10% instead of 20% and the ROI turns negative. If you want to try your own scenario quickly, the ROI calculator on our solutions page offers a starting point for a rough first estimate.
How Do You Work Out the Payback Period?
The payback period is the month in which the investment amortises itself, and it is usually more convincing than an ROI percentage, because it speaks in cash flow.
The calculation is this: Payback month = One-off costs / (Monthly gain − Monthly ongoing cost). In the example above the one-off items come to 85,000 TL while the monthly net contribution is roughly 20,600 TL, which lands at the end of the fourth month. Add the fact that the learning curve eats the first two months and the realistic expectation is the sixth.
A practical threshold to work with: payback periods longer than 12 months should be treated as risky at SME scale. A tool that does not pay for itself within a year struggles to hold its assumptions, because pricing, technology and the team will all have changed within that time.
What Are the Common Mistakes That Break the Calculation?
- Writing the gain on gross revenue. An extra customer worth 8,000 TL is not 8,000 TL of gain; the gain is your profit margin on it. If you do not know your margin, build the calculation on margin.
- Counting time saved as money automatically. An hour freed up is a real saving only if that hour was shifted onto revenue-generating work or postponed a hire. Otherwise it is a comfort gain, not cash.
- Dumping the one-off cost into the first month and panicking. The setup fee should be assessed spread across 12 months.
- Multiplying a pilot result across the whole business. The productivity seen in a two-person pilot does not repeat linearly in a ten-person team.
- Ignoring the chance that the tool dies. Unused subscriptions are the quietest enemy of an ROI calculation; measure the usage rate in the third month.
Which Threshold Should You Set for Saying "No" to an Investment?
Before you decide, put three thresholds in writing: the longest payback period you find acceptable, the date on which you will verify the gain, and which number has to reach which level by that date. A single clear sentence is enough — for example, "by the end of the third month the number of unanswered enquiries will have at least halved."
If the threshold is missed there are two options: correct the assumption and continue, or stop. The bad one is the third option — forgetting the threshold and carrying on paying. If you are struggling to set these thresholds for your own business, tell us your situation; clarifying together which items are measurable is the cheapest piece of work you can do before an investment decision.
Frequently Asked Questions
Which hourly cost should I use when converting time saved into money?
Use the total cost to the employer, not the gross salary: take the annual total including salary, insurance, tax and side costs, and divide it by the hours worked. That figure usually comes out noticeably higher than the hourly equivalent of the gross salary. Count an hour saved at its full value only if that hour was moved onto revenue-generating work; otherwise apply a more cautious rate.
If my ROI calculation comes out negative, should I drop the investment entirely?
No — first look at which assumption produced the negative result. Most of the time the problem is not the tool but the scope: a smaller setup focused on a single process rather than the whole business can produce a significant share of the same gain at a far lower cost. If it is still negative after you narrow the scope, that investment is not right for your business at this point.
How long does it take for the return on an AI investment to become visible?
That depends entirely on the type of item. With a solution that reduces missed enquiries the effect becomes measurable within the first weeks, while gains on the content and visibility side usually need a few months to mature. The healthy approach is to pick a fast-measuring item as your early verification point and track the slow items on a separate schedule.
Is a calculation this detailed worth doing for a very small business?
It is worth it, but shrink the table. In a one-person or few-person business three lines are enough: total monthly payment, hours saved per month, and enquiries recovered per month. The real difference is between doing the calculation and not doing it; how many lines it runs to is a secondary matter.
Why does the ROI of a pilot come out different from the ROI of a full rollout?
In a pilot the setup cost is spread across a small group of users and the people who take part are usually the most willing; those two factors make the ROI look more optimistic than it is. When you move to a full rollout, re-price the training, process change and resistance items. Rather than multiplying the pilot result directly, it is safer to pull the gain assumption down somewhat and build a new calculation.
An AI investment, calculated properly, is an easy thing to argue about: the numbers either hold up or they do not. At Next GEO Agency we prefer to clarify which items are measurable and which data you already have before proposing a solution — because a gain that cannot be measured can neither be defended nor repeated.