# How to Calculate AI ROI: The Payback Multiple and the 3x Bar

*Divide the yearly value a system creates by its build cost plus its first-year running cost. Count only revenue at the margin you keep, costs you can cancel and hours that go somewhere. If the result is under 3, cut the scope or don't build it.*

By Cameron Sullivan, Founder & CEO, Techify AI · 15 October 2026 · 10 min read

> **Key takeaways**
>
> - The AI ROI figure that survives a CFO is a payback multiple. It is the yearly value created, divided by build cost plus first-year run cost. $120,000 of value on $40,000 all-in scores 3.0.
> - Classic ROI is the multiple minus 1, times 100. So 3.0 is a 200% first-year return. It is earned back in four months of full-rate value.
> - The bar is three to four because every estimate is too rosy. Halve the value of a 4x system and it still returns 2x. Halve a 1.5x system and it returns 0.75x, a loss.
> - In IBM's 2025 CEO Study, only 25% of AI projects had paid back what was hoped.
> - Count hours only if they go somewhere. Count revenue at the margin you keep. Count a cost only when you can cancel the line.

The short version: calculate AI ROI as a payback multiple. Take the dollars a year the system will add or save. Divide that by the build cost plus the cost to run it for twelve months. Count three kinds of value only: revenue added, costs cut and hours saved. Put each one through a trap test. Then hold the result to a bar of three to four. Under 3, cut the scope until it clears, or walk away.

That bar sounds steep until you see the record. IBM asked 2,000 CEOs in 2025. Only 25% of their AI projects had earned the return they expected in the last few years. Gartner expects over 40% of agentic AI projects to be scrapped by the end of 2027. Agentic AI is AI that takes steps on its own. Two of Gartner's three reasons are rising costs and unclear business value. Both are a missing division sign.

## The AI ROI formula

**Payback multiple = annual value created divided by (build cost + first-year run cost)**

Here are the three parts, in plain words.

- **Annual value created.** The dollars a year the system adds or saves. It comes from three kinds of value only: revenue added, hours saved and costs cut. If it does none of the three, it is a hobby.
- **Build cost.** The fixed price on the quote: one number, written down. On an hourly quote, use the estimate. Then add the overrun from your last three hourly projects.
- **First-year run cost.** Twelve months of model and hosting bills. Add the hours your people spend watching over the system.

Finance will ask for two more numbers. Classic ROI is the multiple minus 1, times 100. So a multiple of 3.0 is a 200% return in year one. The payback period in months is 12 divided by the multiple. That is four months at 3, and three months at 4. At 1.0 it takes a full year. That means the system has earned nothing beyond its own cost. Then there is ramp-up: the slow start before the system runs at full speed. It comes on top of both numbers, so ask for it in writing.

Why a multiple and not a percent? First, it keeps the running cost in the math. Vendors like to leave it out. Second, "3.7" is easier to defend in a meeting than "270%". Some systems pay off for years. For those, your finance team may also want a net present value. That figure counts money in later years as worth less than money today. The first-year multiple is the gate before that talk. It doesn't replace it.

## What counts as value, and the trap in each

"What does it return?" A CFO asks it in a meeting of six. The scene is invented, but you have sat in it. The sales director has just shown a tool that writes up each sales call. The room nodded.

"It saves the team about twenty hours a week."

"Whose twenty hours," says the CFO, "and what do they do with them?"

No one has a slide for that. Here is the slide.

| Kind of value | How to count it | The trap |
|---|---|---|
| Revenue added | Extra orders x average order x the margin you keep | Counting revenue you would have won anyway |
| Costs cut | A line on a bill you can really cancel | Counting a cost that stays on the books |
| Hours saved | Hours x loaded hourly cost x 52 | Hours count only if they go somewhere: a canceled contractor, a skipped hire, work that earns |

A CFO will ask if "revenue" means revenue or margin. Margin is the part of each sale you keep after costs. Show both, and let the multiple stand on the margin. It survives the meeting that way.

## What counts as cost

Most business cases get the build cost right and the rest wrong. Four lines belong on the cost side.

1. **The build.** The fixed price, or the hourly estimate plus your real overrun record.
2. **Model and hosting.** Take the systems in our builds that handle 1,000+ AI interactions a day. Their model costs tend to run $200 to $1,200 a month, all in. That range assumes tiered routing: each job goes to the cheapest model that can do it. Send every query to the priciest model and the same volume can cost many times more. Tiered routing typically cuts token costs by 40 to 70%. Tokens are the small chunks of text a model bills by. The full picture of run costs is in [what it costs to run an AI system](/insights/cost-to-run-an-ai-system).
3. **Supervision.** The hours your people spend checking its work and taking the cases it passes up. Count them at their loaded cost. Say a system drafts all its work for a person to approve. It saves less time than one that sends on its own. And it should, where a mistake is costly.
4. **The way in.** Nearly every system has one. It might be an API, a door one program opens for others. It might be middleware, a file export, or a screen a system can drive. The question is what it costs. If linking the systems eats half the value, the score halves with it.

## Why the bar is three to four, not 1.5

Every estimate in a business case is too rosy. The bar is the margin that pays for being wrong. So halve the value. A slow rollout plus one bad guess will do that. Then watch what happens:

| Scored | Value cut in half | Verdict |
|---|---|---|
| 1.5 | 0.75 | Loses money |
| 3 | 1.5 | Still in profit |
| 4 | 2 | Comfortable |

A team that would take 1.5 and call it a win is signing up for risk. One wrong guess turns its profit into a loss. The bar is the cheapest insurance in the budget, because using it costs nothing. It is also the bar we hold our own scopes to. Under 3, we cut the scope until it clears, or we don't build it.

## A worked example: two proposals, one budget

Same budget, two proposals. Each comes in at $40,000 all-in for the first year, build and running costs combined. Every number is invented. Neither is a quote from anyone.

| | Option A: summarize every sales call | Option B: chase overdue invoices |
|---|---|---|
| What it does | Writes the call notes | Reads every unpaid invoice, drafts the reminder, hands the stubborn ones to a person |
| Kind of value | Hours saved | Cash recovered and a cost cut |
| The math | 12 reps x 5 calls a week x 20 minutes = 20 hours a week. 20 x $50 x 52 | $360,000 a year written off x 35% recovered = $126,000. Plus the contractor who chases today: 10 hours a week x $45 x 52 = $23,400 |
| Annual value | $52,000 | $149,400 |
| All-in cost, year one | $40,000 | $40,000 |
| Payback multiple | 1.3 | 3.7 |
| Value cut in half | 0.65 | 1.9 |

Option A is what the sales team asks for, and the demo looks great. Its math is generous too. It assumes every saved minute turns into money. It still returns only 1.3 on a bar of 3. Halve it and it drops to 0.65, a loss.

Option B doesn't need all $40,000 on day one. It needs a pilot on the 40 oldest unpaid accounts. You judge that pilot against a pass line: the result it must reach to go on. The math gives you the line before the pilot starts.

- To score 3, it must recover 27% of the write-offs. ($120,000 needed, minus $23,400 of contractor saving, leaves $96,600. That is 27% of $360,000.)
- To score 4, it must recover 38%. ($160,000 minus $23,400 is $136,600.)
- At 10%, it scores 1.5. Stop.

Write the pass line down before the pilot starts. Moving it later is how a 1.5 becomes a success story. On paper, B earns back its cost in about three months of full-rate value. (12 divided by 3.7 is 3.2.) Halved, it still returns 1.9. It is an invoice reminder. It will never be on a conference slide. It will be on the bank statement.

## Capacity value: the half the formula misses

The Payback Rule measures what a system replaces. It says nothing about what the system makes possible. Think of the coach who can take on more clients. Or the rep who can take more first calls. Or the founder who gets Fridays back. That value is real. A rule that ignores it will turn down some good builds.

So count it in a second column, on three conditions:

1. There is demand waiting. Not "we could sell more." A queue you can see.
2. You can name what the freed time will do, and who owns it.
3. You value it at the margin you keep, not the revenue.

If all three hold, add it and label it "made possible". Then report the multiple twice: with it and without it.

Some builds clear the bar only with the second column. That is a bet, so fund it like one. Start with a first phase sized to test the demand, with its own pass line. Release the rest of the build only if the demand shows up.

## When AI isn't worth building

Each of these fails the same division.

1. **A workflow that runs ten times a year.** Ten runs x 3 hours x $60 = $1,800 a year. At 3x, the all-in ceiling is $600. Use a checklist.
2. **A chatbot with no traffic to answer.** 25 questions a month at six minutes each is two and a half hours. At $40 an hour, that is $1,200 a year. The ceiling is $400 at 3x and $300 at 4x.
3. **AI where a process change fixes it.** If two systems ask for the same form, delete the second form. It costs nothing and beats every build.
4. **A landing page or a basic CRM setup.** No AI creates the return. A good freelancer will do either well.

## When not to use the payback multiple

The rule is for builds. It is the wrong tool in three places.

**You are exploring.** Say the honest answer to "what is it worth?" is "we don't know yet". Then you are buying facts, not a system. Cap the spend and call it research. Write down the question the test will answer.

**The return is a loss you avoid.** Compliance, security and brand risk don't show up as revenue. Price them anyway: the cost of the event x the chance it happens in a year. Then run the same math. Say in the first line of the case which numbers are estimates.

**You want the lowest quote.** The rule doesn't find the cheapest vendor. It tells you if the thing is worth buying at all.

## FAQ

### How do you calculate ROI on an AI project?

Divide the yearly value by the all-in first-year cost. Yearly value is revenue added, at the margin you keep. Add costs cut that you can cancel, and hours saved that go somewhere. All-in cost is the build price plus twelve months of model, hosting and supervision costs. A system worth $120,000 a year that costs $40,000 all-in scores 3.0. That is a 200% first-year ROI.

### What is a good ROI for an AI project?

Three to four times the all-in first-year cost. Below that, one wrong guess can turn the project into a loss. Halve the value of a 1.5x project and it returns 0.75x. A 4x project still returns 2x. If you add capacity value, report the multiple twice, with and without it. Then whoever signs knows which part is a bet.

### How long until an AI system pays for itself?

Twelve divided by the payback multiple gives the months of full-rate value it takes. At 3, that is four months. At 4, it is three. Ramp-up comes on top, so ask for it in writing. A system that takes twelve months to pay back scores 1.0. It has earned back its own cost and nothing more.

### How do you measure AI ROI after launch?

Measure the number the system was built to move. Check it against the baseline, the number you wrote down before the build. That might be write-offs won back, hours a week on the task, or tickets answered without a person. Read the real model and hosting bills each month. Count the hours people spend checking it. Then work out the multiple again on actuals at three and twelve months. Do it the same way you did it on estimates.

## Sources

- IBM Institute for Business Value, 2025 CEO Study, 6 May 2025 (2,000 CEOs; 25% of AI initiatives delivered expected ROI): [newsroom.ibm.com](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles)
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025, citing escalating costs, unclear business value or inadequate risk controls: [gartner.com](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)
- The three-to-four times bar we apply to our own scopes: [what we won't do](/commitments#refusals), Techify commitments

All worked numbers in this post are hypothetical arithmetic, shown on the page. The $200 to $1,200 monthly model-cost range and the 40 to 70% routing saving are our figures for production systems at 1,000+ daily AI interactions.

## Related

- [How to find AI use cases in your business](/insights/how-to-find-ai-use-cases)
- [What it costs to run an AI system](/insights/cost-to-run-an-ai-system)
- [Why AI projects fail](/insights/why-ai-projects-fail)