How to measure AI ROI in engineering

TL;DR

  • AI ROI is (Outcome − Rework) ÷ Investment. It is the working software AI helps you ship, minus the cost of redoing what came back, divided by what the AI tools cost.
  • Hours saved, active seats and acceptance rates measure activity. All three can rise while delivery stays flat.
  • Rework is the term most teams skip. Leave it out and a team that returns 2.0x looks like 3.0x.
  • Compare each team with 12 months of its own history, not with another company's benchmark. Then price each leak in dollars per month and fix the biggest one first.

AI spend is no longer an experiment. Gartner expects worldwide AI spending to reach $2.59 trillion in 2026 (May 2026). Yet 78% of finance executives say they cannot fully tie AI spend to business outcomes, and 66% of boards now make more AI funding depend on proof of return (CloudZero, June 2026). If you run engineering or own the budget, the question has arrived before the evidence. This guide gives you a method to answer it.

What AI ROI actually is

AI ROI in engineering is how much of your AI investment turns into shipped, working software. The formula has three parts:

AI ROI = (Outcome − Rework) ÷ Investment

  • Outcome is the work that reaches production and stays there, above what the team delivered before. Code generated is not an outcome. Code merged, deployed and not reverted is.
  • Rework is work that has to be done again after it was called done: a ticket that comes back, a pull request that bounces in review, a change that breaks and needs a fix. Its cost is the hours people spend on it, at their real hourly cost, plus the AI spend of both attempts.
  • Investment is what you pay for AI tools: seats, tokens and agents.

The goal is efficiency. More working software per dollar, with less of it coming back.

A worked example

Illustrative example. Round numbers, not customer data.

A team of 40 engineers spends $20,000 a month on AI coding tools.

LineHow it is calculatedAmount
Outcome80 more working changes a month than the team's baseline, at the baseline cost of $750 per change$60,000
Rework25 of those changes came back: hours to fix them at real cost, plus the AI spend of both attempts$20,000
InvestmentSeats, tokens and agents$20,000
AI ROI($60,000 − $20,000) ÷ $20,0002.0x

Now drop the rework line. The same team reports ($60,000 ÷ $20,000) = 3.0x. That gap is not a rounding error. It is a third of the reported return, and it decides whether this team gets more budget or a fix first.

Why hours saved is not ROI

Most AI dashboards report one of three things: hours saved, seats in use, or how often developers accept a suggestion. Each is useful. None is a return.

Hours saved is self-reported, and people misjudge it. In a randomized trial, METR found experienced open-source developers took 19% longer on tasks with early-2025 AI tools, while believing they had been 20% faster (METR, July 2025). METR later labeled that result historical and changed its method, so do not read it as "AI slows developers down." Read it as proof that felt speed and measured speed can be far apart.

Usage is not delivery. A seat that is used every day can still produce code that waits in review or comes back. Telemetry from more than 10,000 developers showed teams with high AI adoption completed 21% more tasks and merged 98% more pull requests. The same data showed pull requests 154% larger, reviews 91% longer and 9% more bugs per developer, with no link between AI adoption and faster delivery at the company level (Faros AI, July 2025, vendor data, directional).

Acceptance rate stops at the keyboard. It tells you a suggestion was taken. It does not tell you whether that code shipped and held.

The pattern is the same in all three. They measure the moment code is generated. ROI is measured where the work lands.

Rework: the cost that erases the return

Rework is where AI gains leak back out. A change written in twenty minutes can wait nine days for review, or bounce back twice over an unclear acceptance criterion. The AI budget was spent. The value did not arrive.

It is not only an engineering problem. BetterUp Labs and Stanford's Social Media Lab found 41% of workers had received AI output that looked polished but lacked substance. Each case took 1 hour 56 minutes to deal with, an estimated $186 per employee per month, based on self-reported time (Harvard Business Review, September 2025).

In engineering, rework leaves a trail in tools you already use:

  • Tickets reopened after they were closed
  • Pull requests that return from review more than once
  • Changes reverted or hot-fixed after deploy
  • Work items that move backwards on the board

Two things make rework easy to miss. The person who generated the code feels the speed, while someone else pays the fix. And the cost shows up days later, in another sprint, under another ticket. That is why you have to put a price on rework in dollars per month. A rework rate in a chart gets noted. A dollar figure gets a decision.

Compare each team with its own history

Benchmarks are tempting. "Teams like yours ship X per engineer" sounds objective. For AI ROI it misleads, for three reasons.

  1. Teams are not comparable. A platform team and a mobile team have different review norms, change sizes and rework patterns. A cross-company average blends them into a number that fits nobody.
  2. What you need to know is what changed. The useful question is not "is this team's rework high?" It is "did this team's rework move since AI arrived, compared with its normal range?"
  3. Noise is real. Releases, holidays and on-call weeks move metrics every month. You need enough history to tell a real change from a normal swing.

That is why the baseline should be 12 months of the team's own history: the transitions where its work usually bounces back, the states that behave like queues on its board, the cycle times that are normal for it. Measure each change against that, team by team. A company-wide average hides the one team that generates most of the rework.

A six-step method you can run

This method works whether you are a VP of Engineering preparing a board slide or a CFO deciding next year's AI line.

Step 1. Add up the investment per team. Include every AI cost: coding assistant seats, API tokens, agents. Assign it to the team that uses it. One company-wide AI line cannot be defended team by team.

Step 2. Set each team's baseline. Take 12 months of history from the board, the repository and the CI/CD pipeline. Record the normal range for throughput, cycle time, review time and rework.

Step 3. Define the outcome. Pick a unit that means delivered work: merged changes that reach production. Add a quality filter, for example "not reverted or reopened within 14 days." Count only what clears it.

Step 4. Price the rework. For each reopened ticket, bounced pull request or revert, count the hours spent at real hourly cost, plus the AI spend of both attempts. Sum it per team per month.

Step 5. Calculate and place each team. Compute (Outcome − Rework) ÷ Investment and the cost per outcome. Then place each team on a cost × rework matrix: scale or fund, optimize cost, fix the rework, or trim.

Step 6. Act on the largest leak, then measure again. Pick the problem with the highest monthly cost. Assign an owner and an action. Check next month whether the metric returned to its normal range. Repeat.

The first two steps are where most attempts stall, because the data lives in several tools. Do them once and the rest becomes a monthly read.

From measuring to recovering

Measurement alone does not return money. A dashboard that says rework rose 30% still leaves someone to find the cause, decide the fix and follow up. The value is in the last mile: a price, an owner and an action for each leak.

This is the part Leanmote was built for. It connects to 50+ tools read-only and compares each team's metrics with 12 months of its own history. It prices every problem it finds (rework, queues, stalled work, AI spend without return) in dollars per month. Then AI agents work on recovering it, at the level of autonomy you choose. Connecting takes 2 minutes, and the first problems show up on day 1. Across customers, Leanmote detects $6M+ in losses per year and governs 6T+ AI tokens per year.

You can run the method in this guide with spreadsheets. Leanmote makes it continuous, and puts it to work on the fix.

Frequently asked questions

How do you measure the ROI of AI in engineering?

Use (Outcome − Rework) ÷ Investment, per team. Outcome is working software shipped above the team's own baseline. Rework is the cost of redoing work that came back, in hours at real cost plus AI spend. Investment is what the AI tools cost. Compare each team with 12 months of its own history.

Why is "hours saved" not a measure of AI ROI?

Hours saved is self-reported and ignores the hours added back by rework. In METR's early-2025 trial, experienced developers believed they were 20% faster while they took 19% longer. Measure delivered work net of rework instead.

What is AI rework?

AI rework is work that has to be done again after AI-assisted output was called done: a reopened ticket, a pull request that bounces in review, a reverted change. Its cost is the hours spent fixing it plus the AI spend of both attempts. It is the term most AI ROI calculations leave out.

Should I compare my teams with an industry benchmark?

No. Teams differ in review norms, change size and rework patterns, so a cross-company average fits none of them. Compare each team with its own 12-month history to see what changed since AI arrived.

What is a good AI ROI?

There is no universal target. A return above 1.0x means the outcome, net of rework, is worth more than the AI tools cost. More useful than one number is the trend per team and which teams are trending down.

How long does it take to measure AI ROI?

With 12 months of history already in your board, repository and pipeline, you do not need to wait for a new baseline. Leanmote connects in 2 minutes and shows the first priced problems on day 1.

Sources

  1. Gartner, Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026, May 2026.
  2. CloudZero, Finding the ROI of AI: The Finance Perspective, June 2026.
  3. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025.
  4. Faros AI, Lab vs. Reality: What METR's Study Can't Tell You About AI Productivity in the Wild, July 2025.
  5. BetterUp Labs and Stanford Social Media Lab, AI-Generated 'Workslop' Is Destroying Productivity, Harvard Business Review, September 2025.

Keep reading

Put a price on what leaks. Recover it.

Book a demo