Written by: Mike Cannon-Brookes
Atlassian CEO & Co-founder
What is AI changing about how software teams work?
That’s the question behind the new Atlassian AI SDLC Index, which brings together data from more than 340,000 engineers at over 500 engineering organizations to examine AI’s impact across the software development lifecycle.
CEO and Co-Founder Mike Cannon-Brookes shares his thoughts.
AI has made code cheap, but building software is still hard. Every engineering leader seems to be wrestling with the same two questions: what is all this AI actually costing us, and is the code any good?
For most of this year, the answer to those two questions was more, and not as good as we’d hoped, but October brought some better news.
For the first time in six months, AI spend per unit of output fell, dropping nearly 28%. Code survivorship also rose by 9 percentage points. That’s how much AI-written code is still standing in later commits. It’s one month of data, so it’s too early to call it a trend, but it’s a welcome change.
These are the kinds of shifts we’re tracking in the new Atlassian AI SDLC Index.
Built on real-world data from DX, Atlassian’s engineering intelligence platform, the index brings together two things we don’t often look at side by side:
- What developers are doing: PRs, reviews, deploys, and AI-generated code, measured directly from engineering systems.
- How developers are feeling: what’s getting better, what’s getting worse, and what’s slowing them down.
A lot of what gets said about AI and engineering right now is based on anecdotes, a single hypothesis, or one snapshot in time. What leaders need is real data, grounded in real work and tracked consistently over time.
That’s why we’re building this index, to help separate real shifts from short-term noise. This is the first edition, and we’ll add to it every quarter so you can see where the industry is heading and how your teams compare.
There’s a lot to dig into this quarter, but what we’re seeing is that AI has squeezed the middle of the software lifecycle, which is writing the code. The hard parts have moved to either end, before the code and after the code.
What the data tells us about AI and software development
1. The middle is getting cheaper and better

From April to September 2026, AI spend per unit of output rose 4.5x. We measure output as merged PRs, weighted by size and complexity, so a major feature counts for more than a minor fix. That lets us compare AI spend against the work actually delivered. Not every AI cost gets captured, so the real increase is probably higher.
Then October happened. Spend per unit of output dropped from $40 to $29 – nearly 28%.
My hunch is that three things are going on:
- Organizations are getting more disciplined. In our monthly pulse survey, 40% of knowledge workers said they’d recently been asked to monitor or justify their AI usage, hit a usage cap, lost access, or been moved to a cheaper tool. Early experimentation made sense. Leaderboards that reward burning through budgets don’t.
- Context is doing more of the heavy lifting. Models keep leapfrogging each other, but context compounds. Give an AI tool the right context, and it gets to a good answer in fewer, cheaper steps.
- Developers are learning where AI actually works. After months of experimenting, teams are getting smarter about which tasks to automate, which models to use, and when human judgment matters most.
The teams doing this well are measuring delivery, not usage.
Usage and spend are easy to count, but they’re not outcomes. We look at AI through delivery measures like cycle time, throughput, and quality, because that’s where you see whether it’s helping teams ship.”
— Jose Ramirez, Executive Director of Enterprise Agility & Delivery, Kaplan North America
What leaders can do:
- Measure outcomes, not tokens. Cut costs through better context, smarter model choice, and more targeted use of AI. At Atlassian, every employee gets a monthly AI budget they own, and what we’re finding is that the people who hit their limit are often our most inventive AI users.
2. The bottleneck has moved

AI is writing a lot more of the code. Across DX customers, AI-generated code went from 49% of total merged code in April 2026 to 70% in October.
But more code doesn’t mean better code. That’s why the Index tracks code survivorship – the share of AI-authored lines retained in later commits. From April to September, survivorship fell from 80% to 59%, raising questions about how well AI-generated code was holding up over time.
Then, in October, survivorship rebounded to 68%, a 9-percentage-point jump in a single month.
Teams have spent months experimenting with a cool new technology, and they’re starting to find places where it really works. At Atlassian, I’ve seen our engineers build much sharper judgment about which workflows suit automation and which suit augmentation. AI platforms are also getting better at routing tasks to the right models and applying the most relevant context.
Put that next to falling spend, and it’s a good sign – lower cost, and increasingly higher quality. Customers are seeing the same shift.
Seeing more AI-generated code make it into subsequent commits is encouraging. It tells us that the quality and usefulness of AI-assisted development are improving. But writing code is only one part of the journey. What really matters is what happens next.”
— Sudhanva Ramesh, VP, Office of Agile Tools & Analytics, Synchrony
But cheaper code doesn’t remove the bottleneck; it moves it. In October, human PR review volume also grew faster than in previous months. Code is getting cheaper to write, but someone still has to read it, test it, and stand behind it.
What leaders can do:
- Measure value, not volume. Stop treating the percentage of AI-generated code as a success metric. Track how much of that code survives.
- Watch for the new bottleneck. Monitor change failure rates and review wait times to spot where more code is creating more problems.
- Strengthen quality controls. Invest in review capacity, automated testing, and CI/CD so quality keeps pace with AI-driven code production.
Review time was the first bottleneck we hit, but we were able to manage it differently. We built a classifier for the type of PR, and delegated the simple ones to AI. It was our biggest blocker, and we were able to remove a lot of it.”
— Guillaume Roy, CPTO and Co-founder, Workleap
3. Developers are not spending more time on innovation
Disruption without innovation is a real miss. Developers still spend roughly 58% of their time building new features, with the rest going to maintenance. That innovation ratio has barely moved in nearly two years, despite rapid AI adoption.

It makes sense. When people free up time, they spend it on the well-scoped tasks already on their list. For decades, the bottleneck in software was execution, so that’s a perfectly rational instinct.
Molly Sands, Head of Atlassian’s Teamwork Lab, shared a perspective I agree with when she said, “If people are able to do more, and the organization hasn’t shifted what activities they reward, the reasonable thing for engineers to do is check off more tasks on the lists.”
Because engineering is output-bound, not input-bound. In 25 years, I’ve never heard an engineering team say, “We’ve reached the end of the roadmap; can we have some more ideas?” So freed-up time will always get filled. The real opportunity is what you fill it with.
A lot of the work around code, like deciding what to build and aligning with other teams, hasn’t become faster with AI.
What leaders can do: Don’t assume time saved by AI will turn into time spent on innovation. Create the conditions for it.
- Make room for experimentation. Use hackathons and dedicated exploration time to make experimentation a normal part of the job, just like time for maintenance.
- Reward new thinking, not just output. Recognize engineers for automating the stuff that doesn’t need a human, instead of simply closing more tickets.
- Rethink how work gets prioritized. When AI frees up capacity, deliberately point some of it at new opportunities instead of letting the backlog swallow it.
4. Better code starts with better planning and documentation
As AI makes generating code easier, the quality of the information used to generate that code becomes even more important.
As part of the Index, developers rated 14 aspects of their work, from documentation to deep work. We then looked at how strongly each factor relates to four outcomes: speed, quality, ease of delivery, and ease of releasing code to production. Comparing each factor’s importance with how developers rate it shows where improvements will have the biggest impact.
Two opportunities to improve quality stand out: planning processes and documentation. Both are strongly linked to higher-quality code, yet only ~62% of organizations have a majority of developers who rate them favorably.

The good news: not everything needs fixing. Incident handling is the most highly rated factor and the most strongly linked to quality. Change confidence and local iteration speed also perform well.
What leaders can do: Quality starts before the first line of code. The teams seeing real gains in speed and quality start with clear specs, well-scoped work, and up-to-date context.
- Invest heavily in inputs. Give people and agents strong starting context through clear, well-scoped briefs and specs. Whoever owns a spec should stand behind every word, because mistakes compound once people and agents build on them.
- Document for agents as well as people. Keep repos clean, files labeled, and decisions current. Use automations to add decisions made in meetings to a single source of truth so briefs don’t go stale.
- Protect what’s working. Keep strengths like incident handling as teams bring in new AI workflows.
So where does this leave us?
AI has made code cheap; building software is still hard, but the hard part has moved.
The good news is that engineering leaders have a lot they can influence. Better planning, clearer specs, documentation that people and agents can trust, enough review capacity for humans to stay close to the work, and deliberate space for teams to experiment, rather than letting the backlog consume the time AI gives back.
None of this is new; it’s what good engineering teams have always done. AI just makes it more important.
I’m optimistic about the future. Every time software has become cheaper to build, we’ve found more things worth building, and I think AI will be the same. But we’re still early, and nobody has this fully figured out – us included. That’s why we’ll keep tracking it.
We’ll be back next quarter with more data and more to learn. I’m keen to see whether October was a blip on the radar, or the start of a real shift.
The Atlassian AI SDLC Index brings together engineering-system and developer-survey data to show how teams are using AI, where work is getting better or worse, and whether AI-generated code holds up over time. We’ll update the Index next quarter to see whether the lower AI spend per unit of output and higher code survivorship observed in October mark a lasting shift.



