A developer using an AI assistant can open a pull request in minutes, yet the change still sits in a queue until a human reviewer picks it up. That is why agent-written pull requests wait 17.6 hours for a first review against 3.4 hours for human ones, even as enterprises spent $4.0 billion on AI coding tools in 2025, seven times the year before.

Today we sit down with Ori Keren, co-founder and CEO of LinearB, to discuss how the company grew from engineering dashboards into a context layer for agents, how it sells to boards that want proof of AI's return and where developer productivity goes once agents write much of the code.
Ishan Pandey: Welcome to Behind the Startup, Ori. Tell us about your journey and what made you build LinearB in the first place?
Ori Keren: Thanks for having me, Ishan. I started writing code as a kid, programming in BASIC on a Sinclair Spectrum in 1987, and I got my first developer job in 2000. From there I moved into management. I ran an engineering group at Interwise, which AT&T acquired, and then I was VP of R&D at CloudLock, which Cisco acquired in 2016. That's where I met Dan Lines, my co-founder.
What stayed with me from those years is how blind software development is. You run a team of a hundred engineers, and every week the business asks when something will ship or why it's late. You answer from gut feel and status meetings. Meanwhile, every answer is sitting in Git, in Jira and in your CI system, just never connected.
So in 2019, Dan and I started LinearB to connect it. We wanted to give engineering leaders real data on how their teams deliver, instead of gut feel, and then help them act on it. Helping teams act on the data is still what drives everything we build, and today those teams include agents.
Ishan Pandey: LinearB started in 2019 by giving engineering leaders delivery metrics, yet you have now rebuilt the platform around a knowledge graph of the whole software delivery lifecycle. What did you see in customer data over the past year that told you this called for a rebuild of the company's core?
Ori Keren: Two things, and they showed up in the data at the same time.
The first was that AI made writing code cheap but didn't make delivery faster. Our benchmarks cover 8.1 million pull requests from 4,800 teams. AI-assisted pull requests are about 2.6 times larger than unassisted ones. Pull requests opened by agents wait 17.6 hours for a first review, against 3.4 hours for human ones. The bottleneck simply moved downstream, into review, merge and ownership.
The second was why agents struggle. An agent starts every task cold. It has a good idea of the code, but it doesn't have the business context. It doesn't know that this service had two incidents last month, that this file has been rewritten five times this quarter, or who owns it. We had all of that, linked from ticket to release, for every customer.
That's when it clicked. Everyone else is focused on writing code. We hold a knowledge graph of everything happening across delivery, and that's exactly what agents and teams need. A dashboard can't hand that context to an agent, so we rebuilt LinearB around the graph itself. Context is the new king.

Ishan Pandey: When you raised your Series B in 2022, LinearB served more than 5,000 development teams. How has the customer base grown since then and which kind of company is now expanding fastest on the platform?
Ori Keren: We've moved up market. In 2022, a lot of our growth came bottom up, from engineering managers at fast-moving tech companies. Today our core is the enterprise, meaning organizations with hundreds or thousands of engineers, several AI coding tools running at once, and a board asking what it's getting for the spend. We're also seeing strong growth among healthcare organizations.

The pattern is consistent. A company adopts Copilot here, Cursor there and Claude Code somewhere else. Within a quarter or two, leadership realizes it can't answer basic questions, like whether the code holds up, where it's working and what it's costing. That's when they come to us, and that's when they expand.
Kraken is a good example of what that looks like. They used our MCP with an AI assistant to find out what was slowing their code reviews. It turned out to be time zones and reviewer availability, not change size. Within four months, they cut review time at the 90th percentile by more than half.
Ishan Pandey: Boards and CFOs now ask directly whether AI coding spend is paying off. How has that changed who signs a LinearB contract, from the VP of engineering to the CFO or CIO, along with how a deal is won?
Ori Keren: The SVP of engineering or CTO is still our executive buyer.
Two things changed:
-
Who they answer to. Ken Ringdahl, the CTO at Emburse, puts our charts straight into his board decks because his board asks specifically about AI use and ROI. That's now normal.
-
There are a lot of company-wide AI initiatives now, driven by the CFO or CEO, and in the companies we work with, engineering is usually where AI spend is highest.
So finance is in the room far more often. Engineering leaders bring us in because they need numbers that hold up in front of a CFO, not a vanity metric from a vendor's own dashboard.

That changes how a deal is won. Five years ago, we won by showing a better cycle-time chart. Today we win by answering the finance question in the evaluation itself. We connect the data, and within days we show which teams are getting value from AI, which aren't, and what each delivered change costs. Token spend makes this sharper. In our 2026 AI benchmarks, which cover 253 engineering organizations, developers at the 90th percentile now spend $35.20 per coding day on tokens, up nearly 75% in three months. When a CFO sees that line, the conversation moves fast.
Ishan Pandey: LinearB built an audience through the Dev Interrupted podcast and free developer tools long before Gartner gave the category a Magic Quadrant. How does that community-led motion now work alongside enterprise sales and which channel brings in your largest customers today?
Ori Keren: Dev Interrupted was never meant as a lead generator. We started it because engineering leaders had nowhere to talk about the craft of running engineering. That trust turned out to be our moat. When a category gets a Magic Quadrant, every buyer suddenly has a shortlist. Our edge is that many of the people on the buying side have listened to the show for years, used our free tools, or borrowed our benchmarks for their own planning.
Today the two motions feed each other. Community and free tools create familiarity, often with a director or a staff engineer. Benchmarks and research give the champion something credible to take upstairs. Enterprise sales then turns that into an organization-wide rollout with security review, procurement and an ROI case.
Our largest deals close through our enterprise sales team, and when we trace them back, someone inside almost always knew LinearB before the first sales call. That's the compounding effect we're after.

Ishan Pandey: Your effort skill sends simple changes through cheaper agent runs while giving risky changes more care and more tokens. How do you price a product whose value is partly measured in tokens saved and what will the next skills for review and merge add for customers?
Ori Keren: First, a clarification. The effort skill isn't a cost-cutting tool that happens to touch quality. It's a judgment tool. It looks at a change's complexity and risk, using the full knowledge graph, and decides how hard the agent should work. Sometimes that means fewer tokens. On a fragile payments service, it means more, with careful, tested edits. The goal is the right effort on each change.
That's why we don't price on tokens saved. If we took a cut of savings, we'd be rewarded for making agents cheap rather than for making delivery better. Customers judge us on the cost of each delivered change, not the raw token bill.
The next skills go where the bottleneck now sits, in review and merge. They'll use the same context to help teams decide where human attention is needed and what can move faster. After that comes governance of AI-written code and spend optimization.
Ishan Pandey: Gartner sizes the developer productivity insight market at about $400 million while enterprises already spend billions on AI coding tools. How large can this category become once it measures and guides agents as well as people over the next three years?
Ori Keren: The $400 million figure describes the developer productivity insight category, platforms that give engineering leaders visibility into how their teams deliver software. We've always operated in that category, and we've also always gone beyond it. Our philosophy has been to explore, measure, act, and the act part now shows up in our agent skills.
Once a company spends that much on agents, a layer that decides how those agents should work, where they should spend effort and whether their output is safe to ship isn't a reporting tool anymore. It's part of the delivery system, like CI/CD or observability. So the right comparison isn't to other dashboards. It's a share of total spend on software delivery, human and agent combined.
What I'm confident about is the direction. We think it's time for the category to stop being something leaders look at once a month and become something agents consult on every change.
Ishan Pandey: For CTOs and VPs of engineering whose boards will ask them this year to prove AI's return, what is the first thing you would tell them to measure and which older metric should they stop reporting?
Ori Keren: Measure the cost per effective delivered change, not code generated, pull requests opened or lines accepted from an assistant. An effective change is one that merged, shipped and didn't come back as rework. Take what you spend on people and tokens, and divide it by those changes. You want effective changes trending up and the cost of each one trending down. That's the number that tells a board whether AI is paying off, and it holds up when finance pushes on it.
Emburse tracks exactly this. Over seven months, most of its code moved to AI, and its cost per effective pull request fell about 35%.
Right next to it, track what each model costs against the quality of the code it produces. The differences are bigger than most teams expect. Kraken found that two engineers can ship the same work while one burns thousands of dollars in tokens and the other spends hundreds, depending on which models they use and how.
Then stop reporting raw throughput, like pull requests or commits per developer, and above all, stop reporting AI adoption rates. Telling a board that 90% of your engineers use AI says nothing about return.
Inside engineering, though, keep measuring how work flows through your software factory. The bottlenecks haven't gone away. They've moved to review and release.

Why AI Code Waits 5x Longer for Review and How Context Fixes It
For CTOs and VPs of engineering heading into a board meeting, the practical takeaway is his metric. Cost per effective delivered change puts people and token spend over work that merged, shipped and stayed fixed, which gives a CFO a number to test. With AI coding spend compounding and the measurement category growing more than 40% a year, I expect the teams that adopt this kind of accounting early to have the easiest budget conversations of 2027. Leaders who want to compare their own teams can start with LinearB's free benchmarks report or the LinearB platform.
Don’t forget to like and share the story!
Vested Interest Disclosure: HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR.

