Atlassian AI chief: firms still aren’t translating AI to organisation-wide gains

TL;DR:

• Atlassian Chief AI Officer Tamar Yehoshua told City AM that “individual productivity is increasing, but not the overall productivity of the organisation” — the structural gap exposing why corporate AI spend has not yet shown up in published productivity statistics.
• Atlassian, used by 85% of Fortune 500 companies, has 154 billion mapped connections across its workplace data layer spanning Jira, Confluence and external Microsoft, Google and Figma surfaces — the scale of “company memory” needed for agentic systems to actually take useful action.
• Mercedes-Benz uses Atlassian agents to process bug reports from vehicle test fleets, improving report quality by 90% and cutting duplicate detection time by 85% — concrete case study with measurable improvements.

The comments come as enterprises face mounting pressure to justify AI software spend, with growing scrutiny over whether the technology is genuinely delivering measurable returns. Yehoshua described internal Atlassian use cases where agents analyse customer meeting transcripts, identify recurring complaints, create Jira tickets, assign tasks and draft follow-up emails — “you’re not doing just a one-off query anymore, you’re doing something that has an action associated with it”.

The productivity gap is now the operating question

Yehoshua’s framing is the most candid statement yet from a major workplace-AI vendor that the productivity dividend has not appeared at the organisational level. UK Morgan Stanley AlphaWise data released this week showed UK firms posting 6% net job losses from AI alongside 10.3% productivity gains — but those gains are individual and team-level, not yet measurable in aggregate. The disconnect Yehoshua describes — individual productivity rising while organisational output does not — is exactly what produces the labour-market profile the UK is now showing.

Why workplace agents are different from copilots

Yehoshua distinguished agents from earlier AI assistants by their ability to take action: “The agents are more powerful because they take action. You’re not doing just a one-off query anymore. You’re doing something that has an action associated with it.” That distinction lands in the same week NatWest unveiled its 2026 fintech cohort, where at least five of eight selected firms describe themselves as “agentic” — and HMRC began rolling out an agentic complaint-handling pilot. The shift from copilots to agents is happening across UK enterprise simultaneously.

The Jevons paradox warning

The interview surfaces a counter-argument quietly: businesses have been warned AI productivity gains could fuel more work rather than less, a modern Jevons paradox. Yehoshua’s answer is essentially “yes but it frees you for higher-value work” — the same answer vendors have given for every previous productivity wave, and the same answer that has typically led to expanding scope rather than reducing headcount.

Data-handling question

The piece flags growing scrutiny over how enterprise software vendors use customer data to train AI. Atlassian recently disclosed plans to collect certain metadata and in-app data from cloud customers unless enterprise customers opt into stricter controls. Yehoshua said Atlassian does not build its own large language models but routes tasks through an AI gateway connecting to “dozens of third-party models” based on cost, performance and task. For UK enterprise procurement, the model-routing approach is helpful — but the metadata-collection default is the more interesting governance question.

Looking forward

Watch whether Atlassian discloses any aggregate organisation-level productivity numbers from its enterprise customers in upcoming reporting. Yehoshua’s candour is unusual; the next set of vendor earnings calls is likely to feature more honest framing of the individual-to-organisational productivity gap than the past two years of bullishness. For UK boards reviewing AI spend against ROI, “we cannot yet measure it at the organisational level” is a more credible answer than the previous default.