Atlassian’s 2026 State of Developer Experience report found that 25% of developers now save 10 or more hours a week with AI, down from 68% in 2025. At the same time, 87% say AI helps them ship work faster. For a buying committee, the two numbers describe a tool that still speeds up the individual and has stopped returning the same hours to the company.
What the Atlassian report measured
Atlassian published the findings on October 20, 2026. Individual sentiment about output remains strong: 87% of developers say AI helps them ship work faster.
Friction has not fallen with it. According to Atlassian, 72% of developers experience the same or more friction than before AI arrived. The top sources are the same ones the company has recorded for years: finding information, meetings, context switching, technical debt and unclear requirements. Finding information ranked first again in 2026.
The report also counts the new work AI creates. 65% of developers say they regularly spend time correcting AI output.
The 68% to 25% drop
The comparison with 2025 is the figure budget owners should read twice. Atlassian reports that 68% of developers saved 10 or more hours a week in 2025 and 25% did in 2026. The company calls this surprising given how quickly models have improved, and asks how much of the time gained from faster coding is now consumed elsewhere.
Atlassian’s own explanation is a systems argument. In the report’s accompanying blog post, its author writes that “optimising for one part of software delivery, coding, creates new problems downstream” because more work has to pass through the same bottlenecks. The post gives an example from a conversation with an engineering leader whose team was swamped by output from upstream teams that were using AI. Using AI to process that output would only pass the bottleneck to the next team.
One caution applies to reading the figures. This is a survey of self-reported time savings from a vendor that sells developer collaboration software. Both years’ figures are Atlassian’s own, so the year-over-year comparison is more useful than either number alone.
Usage is being measured, outcomes are not
The finding with the clearest budget consequence concerns strategy. Atlassian reports that only 17% of developers and 23% of leaders say their organization has a clearly communicated and implemented AI strategy across the software life cycle. Meanwhile 60% of developers say AI usage is treated as a productivity signal, and 41% say they sometimes use AI more than necessary because they feel expected to.
The strategy gap has a measurable effect inside the survey. Atlassian says developers at organizations with a clear, implemented AI strategy are about twice as likely to report less friction than before AI. The comparison group is developers working under inconsistent or unclear policies.
In our reading, those three numbers describe a measurement problem that finance teams will recognize. If a licence dashboard shows seat activity and token consumption, and managers read that activity as productivity, developers have a reason to generate usage whether or not it helps. The spend then grows with the metric.
A second data point from the cost side
Harness, a software delivery vendor, published a related finding on October 20 when it opened access to AI Budgets inside its Cost Management Agent. According to the company, its 2026 State of AI in FinOps Report found that 73% of organizations have an AI cost policy and 13% can tell whether that policy is working. This is a vendor-reported figure from a company selling the tool that addresses it, so it is best treated as directional.
The product mechanics show what a usage control looks like in practice. Harness says budgets can be scoped by team, developer, model or project, and run on a weekly cycle so a team cannot spend a month’s forecast in week one. A warning fires at 90% of budget, and access pauses once weekly spend passes 100% until the next reset or an approved increase. The agent also reports cost per million tokens, which Harness argues says more about whether a team chose the right model than total spend does.
Harness reports that on its own AI tooling bill, average weekly AI spend per developer fell from $184 to $115, about 37%, and cost per million tokens dropped from near $1.80 to $0.80 in the week governance went live. Those are one vendor’s internal results with no control group.
The two sources measure different things, hours saved and dollars spent, and they point at the same gap. Developers are using AI heavily, and few organizations can show what the usage returns or whether it stays inside a policy. Earlier coverage on this site reached a similar conclusion from the finance side: AWS’s own business-case guidance warns that a saved hour is not saved money, and Google’s Gemini agent launch added real-time spend caps.
What it means for the buying committee
For a CIO or CAIO approving developer AI renewals this quarter, the Atlassian figures support three changes to how the purchase is evaluated.
- Ask for the 2025 baseline. If a business case rests on hours saved per developer, check whether the figure was measured this year. Atlassian’s data shows the share reporting 10 or more hours saved fell by more than half in twelve months.
- Replace activity metrics with delivery metrics. Seat utilisation and token counts show adoption. Lead time, review queue length and rework rates show whether the extra code reaches production faster.
- Write the strategy before the renewal. Only 17% of developers report a clear AI strategy at their company. A one-page statement of which problems AI is meant to solve, owned by an executive, is the cheapest item on the list and the one the survey ties to lower friction.
Review and verification capacity belongs in the same budget conversation. With 65% of developers correcting AI output regularly, the cost of the tool includes the reviewer hours it creates, and those do not appear on the vendor invoice.
What to watch
Two things will show whether the 2026 drop is a one-year adjustment or a trend. The first is whether the next annual survey from Atlassian, or a comparable study from another publisher, shows the 10-hour group shrinking again. The second is whether vendors of coding assistants start publishing delivery-side measurements, such as cycle time, instead of satisfaction scores. Until they do, buyers have the developers’ own estimates, and those have just fallen sharply.
Source: Atlassian
