Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

AI Agent Builds a Team Dashboard from Jira in an Hour Instead of a Month of Analyst Work

Building a team dashboard from Jira used to take an analyst about a month, and most of that time went to reconciling data. Now the work is done by the skills of an AI agent that reads Jira, GitLab, Confluence and meeting recordings, and a board for a new team is ready in an hour. The agent still makes mistakes. For example, it counted open MRs differently in different places. So the author keeps the data separate from the layout and cross-checks snapshots before publishing.

AI-processed from Habr AI; edited by Hamidun News
AI Agent Builds a Team Dashboard from Jira in an Hour Instead of a Month of Analyst Work
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
The agent has read-only accessit cannot change statuses or MRs
Key metric #2The board is a single reference HTML file, with the data stored in it as a separate JSON block
Key metric #3Paginated Jira exports silently drop rows, so the task count has to be verified

A team dashboard built from Jira usually looks like a month-long project: an analyst exports the data, reconciles it, builds the layout, then reconciles it again. The author of a write-up on Habr, a former CTO who ran an engineering organization of more than 50 people and now leads a platform handling 2000 financial transactions per second, turned this process into a set of skills for an AI agent. He now builds a board for a new team in an hour at most. The most useful part of the text is not the speed but the honest list of where the agent goes wrong.

Why it used to take a month

Drawing a page with charts is easy. The time goes into the data and checking it, and the errors in it are quiet: the board looks confident while the numbers are wrong. The author lists typical cases.

The reason a task was carried over is buried in the comments: a task slipped to the next sprint, and why is visible only from the discussion thread. A paginated Jira export can silently return noticeably fewer tasks than actually exist, with no error. The team perimeter can be set incorrectly, so the volume of work comes out several times smaller than the real figure, yet the chart still looks plausible. And a typical engineering BI dashboard often contains no sprint metrics at all. That is why the analyst's month is mostly reconciliation.

How the setup works

The agent reads Jira, meeting recordings, the calendar, GitLab and Confluence, with read-only access everywhere. The board itself reports that the agent wrote nothing. The requirement is strict: an agent that can change a task's status or close an MR has no place on such a board.

The board is a single reference HTML file, with the data stored in it as a separate JSON block. Each sprint, the agent changes only the data and publishes the board at the same address, leaving the layout untouched. This guards against the agent "fixing" a number directly in the page text so that it diverges from the rest.

A skill here is instructions plus scripts: where to get the data, how to calculate each metric, which template to fill in, and what to check before publishing. Everything the analyst used to keep in their head is written down once. For a new team, the author sets the perimeter, checks the first run, and the board is ready. The quarterly review and the backlog forecast work the same way: scripts recalculate them from the Jira export, not a person in a spreadsheet.

The metrics are chosen to match the questions the author asks at a sprint review:

  • how much is closed versus how much of the sprint time has elapsed, with a traffic-light indicator;
  • the trend over several sprints, because one bad sprint means nothing, but a series does;
  • throughput and the number of tasks in progress per person;
  • carry-over depth: how many sprints in a row a task keeps slipping;
  • "rubber-stamp" closing — the share of tasks closed on the last day of the sprint; if it is high, the burnup chart lies;
  • unplanned tasks from meeting recordings, stuck tasks, open MRs without a reviewer;
  • discrepancies between Confluence and Jira.

BI fundamentally cannot show unplanned tasks or discrepancies between documents and the tracker: that takes meeting recordings and document text, not just a table of tasks.

Where the agent goes wrong

The author stresses that the agent makes mistakes regularly. For example, open MRs without a reviewer were counted differently in two snapshots. Each snapshot looked fine on its own, and the discrepancy showed up only when they were placed side by side. In another case, the sprint day in the board header did not match the sprint day in the data. A small thing, but a reader who notices it will stop trusting the other numbers.

Two rules follow. First: every number on the board comes from the JSON block, and a number typed by hand into the HTML counts as an error, even if it matches right now. Second: if a metric is calculated in two places, the snapshots are compared before publishing, and if they diverge, the board does not go out until the cause is found. The agent doesn't get tired, but it doesn't doubt its own numbers either, so a human has to do the doubting, and the rules make that cheap.

What to take into practice

  • Start with read-only access, so the agent physically cannot change anything in the tracker or repositories.
  • Build one reference board by hand and proofread it; from then on the agent fills in only the data.
  • Check the number of tasks in the export against the number in Jira itself.
  • Write down the team perimeter explicitly: which projects and boards are included.
  • Add what the tracker lacks: meeting recordings and documents.

The limitation is obvious: this is one manager's account of personal experience, with no measurements of accuracy or token cost, and the demo board is built on fictional data. But the practical takeaway is clear: a team board stops being a project, and it can be set up for every team and updated every sprint without a dedicated person. The time that went into assembly is freed up for a conversation about what to do with the numbers.

Story

Key Questions Answered

Why it used to take a month?

Drawing a page with charts is easy. The time goes into the data and checking it, and the errors in it are quiet: the board looks confident while the numbers are wrong. The author lists typical cases. The reason a task was carried over is buried in the comments: a task slipped to the next sprint, and why is visible only from the discussion thread. A paginated Jira export can silently return noticeably fewer tasks than actually exist, with no error.

How the setup works?

The agent reads Jira, meeting recordings, the calendar, GitLab and Confluence, with read-only access everywhere. The board itself reports that the agent wrote nothing. The requirement is strict: an agent that can change a task's status or close an MR has no place on such a board. The board is a single reference HTML file, with the data stored in it as a separate JSON block. Each sprint, the agent changes only the data and publishes the board at the same address, leaving the layout untouched.

Where the agent goes wrong?

The author stresses that the agent makes mistakes regularly. For example, open MRs without a reviewer were counted differently in two snapshots. Each snapshot looked fine on its own, and the discrepancy showed up only when they were placed side by side. In another case, the sprint day in the board header did not match the sprint day in the data. A small thing, but a reader who notices it will stop trusting the other numbers. Two rules follow.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…