AI Cost Visibility & Optimization Understand, allocate & reduce your AI costs - Learn More

Introducing AI ROI in nOps

See whether your AI coding spend is turning into shipped work

Today we’re announcing AI ROI in nOps.

AI coding tools are one of the fastest-growing lines on the engineering bill, and every leader who approves that spend gets the same question from finance: what are we getting for it? AI ROI answers it with numbers. It joins two things nOps already sees, your AI coding spend (Cursor, Claude Code and GitHub Copilot) and the pull requests your engineers merge (GitHub, GitLab or Bitbucket), and shows what the spend produced for the whole organization, for each team and for each developer.

We ran it on our own engineering team first. In the first 47 days of Q4, 18 nOps engineers spent $27.3K on AI coding tools and merged 682 pull requests, 60% more than during the same period in Q3. 84% of that spend went into work that merged, at $37 per merged PR. AI ROI also showed us where it fell short: AI-assisted PRs needed rework 30% of the time, compared with 24% for the rest, and that rework cost $6.6K of avoidable spend.

Why AI coding spend is so hard to explain

  • The bill counts tokens, not work. Billing tells you who used which model and what it cost. It doesn’t tell you whether the money built a feature, fixed a bug or went into work that was never merged.
  • Every engineer uses a different mix of tools. One team works in Cursor, another in Claude Code, a third in Copilot, each with its own console and its own idea of usage.
  • Nobody ties the spend to outcomes. Finance asks whether AI is paying off, and engineering answers with PR counts and surveys that aren’t tied to cost. The answer is an opinion rather than a number.

How AI ROI works

AI ROI uses data nOps already collects and adds the piece most teams are missing: the link between a dollar of AI spend and the pull request it helped produce.

  • Spend per person per day comes from your Cursor and Claude Enterprise integrations, and GitHub Copilot through the GitHub integration.
  • Merged pull requests come from GitHub, GitLab or Bitbucket Cloud. nOps reads PR metadata only and never stores code.
  • Each merged PR gets an AI cost and a type of work: Build, Fix, Ship, Docs or Understand. The type comes from the PR title or branch, the linked Linear or Jira ticket, or the nOps agent.
  • The nOps agent on each developer’s machine makes cost per PR exact instead of estimated and assigns a type of work to AI spend that never reached a PR. A small model on the machine labels each prompt, so prompts and code never leave the laptop.

Everything rolls up into three scores out of 100. The scale is the same for every team, person and quarter, so a 70 always means the same thing:

  • Velocity: how fast work gets from its first commit to production.
  • Productivity: how much AI spend turns into shipped work, and at what cost per PR.
  • Efficiency: how much AI spend was avoidable. Today that is rework: AI-assisted PRs that a later fix or revert had to correct.

AI ROI has a tab for each audience: the Executive Summary for leaders, the Engineering Overview for engineering managers, and Developers for everyone who writes code.

Results: what AI ROI found on our own team

Executive Summary: is the spend paying off?

The Executive Summary answers that on one screen. Our team scored 82 on Velocity, 86 on Productivity and 71 on Efficiency. Work reached production a median of 7 hours after its first commit, and 79% of AI spend went into work that shipped, at $37 per PR. Compared with the same period in Q3, Productivity rose 11 points and Efficiency fell 29, as more AI-assisted work needed rework.

Below the scores, every team sits side by side with its spend, merged PRs, cost per PR and scores. AI ROI flags a team when its cost per PR rises by a quarter or more, or when its spend grows well ahead of its output. Three of our six teams were flagged. Platform Engineering’s cost per PR rose 166% while its spend rose 556%, which is a conversation to have now rather than at the end of the quarter.

Engineering Overview: where a team’s money goes

Engineering managers can select a team or all teams to see how they perform. Across our teams: 682 PRs merged, 84% of spend tied to merged PRs, $967 per engineer a month, and an estimated 96% of merged lines written by AI.

Where the spend goes follows every dollar from team to type of work to outcome. For us, Build took $13.9K, and $4.3K never reached a merged PR. Where the PR cost goes groups merged PRs by Linear project or Jira epic: the Recommendations project took $3.5K of the $7.1K that is linked to tickets.

Engineering & PRs: what AI-assisted work returns

Every merged PR is listed with its AI cost, output and rework, and AI-assisted PRs are compared with the rest. Nearly two-thirds of our PRs (438 of 682) were AI-assisted. They had a median AI cost of $11 and merged as fast as the rest, but needed rework 30% of the time against 24%. That gap is the next thing for us to fix, and it is now a number we can track.

By type of work, Build was 59% of the PRs’ AI cost at $37 per PR, while Fix PRs grew 148% and cost $5.97 each. Use cases shows the product areas, technical areas and repositories the work went to.

Developers: your own scores, and why

Each developer sees their own numbers, and managers can pick anyone on their teams. Pick a score and Why this score shows which figures moved it and by how many points, with tips for the weakest one. In the example below, a Velocity of 91 sits 10 points above all teams thanks to a 5-hour lead time and 21.4 PRs a week, while an Efficiency of 61 points to rework.

How you use AI splits the developer’s spend across every tool by type of work, and lists what drove the cost: each tool’s spend and cache use, reasoning effort, Cursor fast mode, MCP tool results and background requests. nOps recommendations for the developer’s AI use, such as a cheaper model for the same work, appear beside it with a link to open each one.

The bottom line: AI coding spend reported the way engineering already works, by team, by type of work and by pull request, with three scores that say whether it is getting better.

Get started

  • Connect an AI coding tool. In Settings → Integrations, connect Cursor or Claude Enterprise. GitHub Copilot spend comes in through the GitHub integration.
  • Connect your code host: GitHub, GitLab or Bitbucket Cloud, with a read-only token.
  • Group people into teams (optional). Open AI ROI → Setup and click Set up teams. Teams unlock the team view and the Engineering Overview.
  • Connect Linear or Jira (optional), so PRs whose title or branch doesn’t name a type of work get one from their ticket.
  • Install the nOps agent (optional). Each developer runs npx @nops-io/clara and approves their machine once in the browser. It makes cost per PR exact.

The Setup tab shows each step and what is still missing. AI ROI opens on it until spend and pull requests are arriving.

To try it out in your own environment, you can get started with a personalized demo.

nOps manages $5 billion in cloud and AI spending and was recently ranked #1 in G2’s Cloud Cost Management Category.

Frequently Asked Questions

What does AI ROI measure?

AI coding spend per person, the pull requests they merged, and how the two connect: AI cost per PR, the share of spend that reached merged work, how fast work reaches production, and how often AI-assisted work needed rework. These roll up into Velocity, Productivity and Efficiency scores for the organization, each team and each developer.

Which tools are supported?

AI coding tools: Cursor, Claude Code (through Claude Enterprise) and GitHub Copilot. Code hosts: github.com, gitlab.com and Bitbucket Cloud. Linear and Jira are optional. GitHub Enterprise Server, GitLab self-managed and Azure DevOps are not supported yet. API spend, such as Amazon Bedrock or OpenAI, stays in AI Tracking and isn’t counted as coding spend.

Do we need the nOps agent?

No. Without it, AI ROI estimates each PR’s cost from its author’s spend while the PR was in the making, and shows the cost with a “~”. With it, the cost comes from the sessions that produced the PR, and AI work that never reached a PR gets a type of work too.

Can nOps read our prompts or code?

No. AI ROI never reads prompt text, AI responses, file contents or commit messages. The agent labels each prompt with a small model on the developer’s machine and sends only the label, counts and IDs. It records only in repositories your admin allows, and each developer approves their own machine.

How is cost per PR calculated?

With the agent, it is the recorded cost of the AI sessions linked to the PR. Without it, an author’s spend since their previous merge, up to 14 days back, is credited to the PRs they merge next, split by size. Claude Code spend for the last three days is still settling, so recent PRs show “Arriving” until it is final.

How is the type of work decided?

Using the strongest available evidence first: the PR title or branch (such as fix: or feat/), the linked Linear or Jira ticket, the commit types, the agent’s prompt labels, and finally which files changed. Anything that can’t be typed shows as Not categorized, so you can always see how much of the split rests on each source.

Does AI ROI measure how much time developers save?

No. It measures cost and engineering output: merged PRs, time to production, output and rework. It doesn’t track how long anyone works.

Who can see what?

Executives (organization admins, or a role with the ai-roi:executive permission) see everything. Engineering managers (ai-roi:manager) see every team and its people. Developers see their own numbers, organization totals, and their own team’s totals without names. Totals for a group of fewer than five people who used AI are hidden from anyone who can’t already see its members.

How do I get started?

AI ROI is in preview. Ask your nOps account team to turn it on, then connect an AI coding tool and a code host in Settings. New to nOps? Book a demo and we’ll walk you through it on your own AI spend.

Jordan Stein

Jordan Stein

Published Date: November 19, 2026, Announcement

Related Posts

Introducing Cursor Integration in nOps

Announcement

Introducing Cursor Integration in nOps

byRick HaggartRick Haggart•Published Date: Jul 13, 2026
Introducing Claude.ai (Enterprise) Integration in nOps

Announcement

Introducing Claude.ai (Enterprise) Integration in nOps

byRick HaggartRick Haggart•Published Date: Jul 17, 2026
Introducing Oracle Cloud Infrastructure (OCI) in nOps

Announcement

Introducing Oracle Cloud Infrastructure (OCI) in nOps

byJordan SteinJordan Stein•Published Date: Nov 18, 2026
Introducing the nOps MCP & Skills

Announcement

Introducing the nOps MCP & Skills

byBenjamin White•Published Date: Nov 5, 2026
nOps AI Optimization Recommendations: Closing the Gap From Insight to Verified Savings

AI & Tokenomics

nOps AI Optimization Recommendations: Closing the Gap From Insight to Verified Savings

byRick HaggartRick Haggart•Published Date: Nov 3, 2026
Introducing Databricks Optimization in nOps: Find and fix idle Databricks compute in a few clicks

Databricks

Introducing Databricks Optimization in nOps: Find and fix idle Databricks compute in a few clicks

byJordan SteinJordan Stein•Published Date: Nov 4, 2026