Skip to content
JiraVision
All articles

AI Scrum Master

The complete guide to AI Scrum Masters

What an AI Scrum Master actually does: where the draft comes from, what the rationale contains, and why nothing reaches your board until a person accepts it.

4 min read24 October 2026JiraVision team

A team working through a plan together at a shared table

An AI Scrum Master is not a chatbot with a sprint calendar. In JiraVision it is a recommendation engine with an approval gate: it reads the same task rows you can see on the board, produces a draft plan with a written reason for every item, and then waits. Nothing is written to your sprint until a person accepts it.

This guide covers what the draft reads, what the rationale contains, how acceptance works, and where the feature stops on purpose.

What the draft reads

The draft plan is built from four inputs, every one of which is visible somewhere in the workspace:

  • Open load per member — each task assigned to them that is not done, with its story points.
  • History — points completed in recent sprints, used as a capacity ceiling rather than a target.
  • Calendar pressure — meetings and time off in the next seven days, which reduce available hours.
  • Skills — the skill tags recorded on each member, so work is not handed to the only person who has never done it.

Nothing here is inferred from chat content or from activity telemetry. If a signal is not already a row in your workspace, the model does not use it.

The rationale is the point

Every suggested task carries a plain-language reason: why this person, why this sprint, what capacity assumption it rests on. That is a deliberate choice, for two reasons.

  1. You can disagree with the reasoning instead of guessing at it. A recommendation that reads "assigned to Sam because Sam holds the payments skill tag and has 8 open points" can be checked against reality in a few seconds.
  2. The reason is stored with the record. Looking back at a finished sprint, the accepted draft still explains itself rather than presenting as a decision nobody made.

A score without a reason invites two bad responses — ignoring it entirely, or deferring to it uncritically. Both are worse than not having the feature.

Accept, edit or dismiss

Accepting a draft writes through the ordinary task and sprint endpoints. That matters more than it sounds: the same notifications fire, the same activity entries are written and the same validation runs as if you had made the change by hand. There is no privileged write path for the model.

Dismissing works the same way in reverse. A dismissed recommendation moves to an archive that keeps the reason you gave, so a pattern of dismissals stays visible later instead of disappearing.

Re-drafting mid-sprint is allowed, and it names what slips. Accept a revised plan in week two and the draft tells you which committed items no longer fit and why, rather than quietly re-planning around them.

What it watches

Three views sit alongside the draft, all computed from task rows:

  • Sprint health — a gauge over completion, scope change and blocker age.
  • Burndown — the actual line, the ideal line, and a projection derived from the current completion rate.
  • Velocity forecast — the next sprint's likely capacity, from recent completed points.

Each one prints its own formula in the ethical metrics documentation. If a number cannot show you how it was produced, it does not belong on the page.

What it refuses to do

These refusals are product decisions, not gaps in a roadmap:

  • No per-person productivity ranking. Load is shown so work can be moved, not so people can be ordered.
  • No keyboard time, no idle detection, no commit-count scoring. Motion is not value.
  • No automatic acceptance. There is no setting that lets a draft write to a sprint unattended.
  • No sentiment inference from chat used as a performance input.

Rolling it out

  1. Turn it on for one team, for one sprint. Read the draft before accepting and note where the reasoning is wrong — that is the useful signal, not a bug report to file and forget.
  2. Check the capacity ceiling against reality. If history says 40 points and the team finishes 28, the draft will keep over-committing until the input is corrected.
  3. Dismiss out loud. Use the archive reason field; the pattern is what tells you whether the skill tags or the load data need fixing.
  4. Only then widen it. A second team is a different capacity profile and a different history, and the draft should be re-read there too.

Where to go next

The AI Scrum Master documentation lists every endpoint and field. The metrics guide explains the formulas behind the health gauge and the forecast. And if you would rather see it move, the dashboard page runs against your own tasks.

Written by the JiraVision team. Corrections and disagreements welcome via the contact page.

Send feedback

Keep reading

AI Scrum Master

How AI changes the agile routine

The ceremonies an AI assistant genuinely improves, the ones it leaves alone, and the five numbers worth measuring in your own workspace instead of trusting ours.

Read article

Read the next one in your inbox

Double opt-in, no tracking pixels, and an unsubscribe link in every message.

Create an account