Atlassian's AI Team Wrote 180 Tickets Per Engineer. The Other Team Wrote 43.

A principal engineer at Atlassian ran the numbers on his own team on August 9 and found something nobody plans for. Working with AI agents, five engineers wrote roughly 180 Jira tickets each. A comparable team building a comparable product wrote about 43.

The agents did the building. The humans wrote four times more instructions.

Anand Shastri published the comparison on Atlassian’s engineering blog, and the framing is his:

“our per-engineer output was roughly 5x higher across net lines of code, cyclomatic complexity, database schema size, and external integrations”

Same multiple on both sides of the ledger. Five times the output, and close to five times the written specification needed to get it.

Why a marketing team should care about Jira tickets

Swap the nouns and this is the next two years of marketing work.

A ticket is a brief. It says what to make, for whom, within what constraints, and how you’ll know it’s right. When a person does the making, a rough brief still works, because the person fills the gaps from context they already carry. When an agent does the making, every gap becomes either a wrong output or a round trip.

So the volume of briefing goes up, and the standard goes up with it. Shastri measured both. His team’s tickets scored 4.47 out of 5 on writing quality against 2.72 for the traditionally written ones, and 83% were what he calls agent-ready, against 6%.

6%. That’s the share of normal, human-written work instructions clear enough to hand to a machine.

If your marketing team’s creative briefs are of ordinary quality, that’s the number to sit with. Ordinary is not a criticism here. Ordinary briefs work fine, because a designer who has been at the company two years knows what you meant.

The part that gets missed in headcount planning

The tempting conclusion is that agents let you cut the team. The data points somewhere more awkward.

Atlassian’s CTO Rajeev Rajan described the pattern across the company in The Pragmatic Engineer in February:

“Teams are not necessarily getting smaller, but they’re producing a lot more, sometimes 2-5x more, and creativity is up.”

Not smaller. Producing more. That’s a throughput story. Throughput only becomes savings when demand is capped, and marketing demand never is.

What changes instead is the shape of the job. The work moves upstream, into deciding and specifying, which is the senior half of the role. That’s the same squeeze we described in the experience debt building up under AI juniors, arriving from the other direction: if the junior tasks are automated and the remaining work is specification, the entry point to the career closes while the workload at the top rises.

It also complicates the efficiency case. Agencies and in-house teams have been sold on profit per employee as the AI scoreboard. A team writing 4x more briefs to get 5x more output is more productive, and it is not less busy.

What Shastri says goes wrong

The honest limitation in his write-up is the most transferable part:

“Agents frequently surface real problems, latent bugs, shortcuts, tech debt, while working on something else. But with no frictionless way to capture them in the moment, those findings evaporate when the session ends.”

He calls it a kind of amnesia. The same thing happens in marketing. An agent drafting landing page copy notices the pricing page contradicts the positioning, says so mid-output, and nobody writes it down. Next week it notices again.

Read the method before you quote the number

Four things temper all of this, and they should travel with the statistic.

The sample is 300 tickets, 150 per product, drawn at random from the set referenced by shipped commits. The team is five engineers at one company over a few months. The quality scores came from an AI judge that Shastri built himself, scoring a source-blinded sample on clarity, scope, acceptance criteria, context and actionability. That’s a reasonable method and it’s still an AI grading the output of AI-assisted work.

And Atlassian sells Jira. A finding that better tickets are now business-critical is a finding that its product is business-critical.

The steelman

The strongest counter-evidence comes from Atlassian itself. In a July post on its AI-native tooling, the company reported that AI usage across teams rose 65% while overall developer velocity did not follow, topping out around 15%, with many teams seeing 10%.

That sits badly next to a 5x claim, and it should. One is a single team that reorganized around agents. The other is the average across everyone who turned the tools on. The distance between those two numbers is the actual story: the gain is available and almost nobody is capturing it, because capturing it requires rewriting how work gets specified rather than adding a tool.

What to do on Monday

Pull ten creative briefs your team wrote last month. Score each one on whether a competent stranger could execute it with no follow-up questions. That’s roughly what agent-ready means.

If more than one in ten clears it, you’re ahead of the engineers. If it’s closer to Shastri’s 6%, you’ve found the real limit on your AI rollout. The model, the vendor and the budget are all fine.

Nobody has ever had to write the brief down properly before.