Last updated 2026-11-08

People, automation and AI activity

For teams preparing a DORA or NIS2 review, reviewers typically ask which changes in a Jira history were made by automation rather than by a person. Jira records every change under an account and gives each account a type. MetaFrazo reads that type and shows it across the dashboard. This article explains what the type is, which visuals use it, and where it stops, including what it cannot tell you about AI.

Actor Classification

Actor Classification is MetaFrazo's name for this capability: reading the account type Jira reports for every actor in your history, and carrying it through the dashboard so that a figure about people is not quietly counting apps. Everything below is part of it: the four classes, the per-visual treatments, and the limits at the end.

Two words that sound alike and are not the same thing:

  • Class is a fact Jira reports: person, app, customer or unknown. You cannot change it, and MetaFrazo never guesses it.
  • Role is something you assign yourself on the Workflow Discovery page, such as engineer, product owner or bot. MetaFrazo seeds it from the class as a starting point, and a role you set always wins.

Where this article says a visual separates people from apps, it means the class.

The four account classes

Jira reports a type for each account. MetaFrazo shows it as one of four classes:

ClassWhat Jira reportsExamples
PersonAn Atlassian account (Jira's own value is atlassian)The people in your organization
AppAn app accountAutomation for Jira, or an app that acts in Jira through its own app account
CustomerA Jira Service Management customer accountSomeone raising a request through the customer portal
UnknownNo typeAn account Jira returned no type for, or one that has not been looked up yet

Unknown is shown as unknown. It is never counted as a person. Two other labels appear on some rows and are not account types: erased, for an account whose personal data has been erased, and system, for an event that records no actor. Neither counts as a person.

If you are looking for "bots", MetaFrazo's word is app account. An account counts as an app because Jira types it as one, never because of its name or how busy it is.

MetaFrazo looks up each account's type in the background and applies it to all of that account's events, including those from before the lookup. A newly seen account can read unknown until its lookup finishes. From each lookup, MetaFrazo keeps the account ID and the type, and for an app account its name, so a panel can read "Slack (app)" instead of a short ID. It never keeps the name of a person or a service-desk customer; see What data we hold, and what we do with it in Security and hosting: where your data lives and how it is protected.

Which visuals separate people from apps

Each column of the matrix below is one way a visual treats app accounts:

  • Shown separately. The figure that leads counts people's accounts. Apps, customers and accounts of unknown type are counted beside it and never inside it. Where an app or customer account acted, the people figure differs from an all-accounts figure.
  • People only. The measure is about people, such as workload, throughput or lead time, so app and customer accounts are left out. Accounts of unknown type stay in and are labeled.
  • Labeled. Nothing is removed. Each row shows the account type of the actor, so a reviewer can see which rows an app made.
  • By account type. The account-change visuals count accounts being created, updated and deleted. They split by the type of the account the event is about, so an app being installed or a customer signing up is not counted as a person being added.

What each treatment looks like in practice

  • Shown separately, as on Segregation of Duties Violations 5.6.6. A rule that creates an issue and then closes it is a single-actor chain, but it is not a person skipping a review. The headline counts people's chains and the rule's are counted beside it, so the number you report is the one the question was about.
  • People only, as on Weekly Team Throughput by Actor 4.5.5. A nightly rule that transitions two hundred issues would otherwise be the most productive member of the team. It is left out.
  • Labeled, as on Configuration Change Event Audit Log 2.2.11. Nothing is removed, because for a configuration change you want every row. Each row carries the actor's class, so a block of permission changes made by an app is visible as such without filtering anything away.
  • By account type, as on User Lifecycle Heatmap 2.6.7. Installing a Marketplace app creates an app account. Without the split that reads as somebody joining, and an account-growth chart gains a joiner nobody hired.

Each visual appears once, by its dotted ID; a KPI follows the visual it goes with, or stands on its own. Names such as Segregation of Duties Violations 5.6.6 and SoD violation rate 5.6.2 are MetaFrazo's own measurement names, defined in each visual's description. They are not findings about your organization.

x.y.z visual · KPIx.y.z KPI · Each number links to its page; hover for the name.
People only: app accounts left out · By account type: account changes split by account type

¹ A pair that includes an app or customer account is left out.

² Accounts Jira reports as apps start with the bot role. A role you set yourself wins, and every other unmapped account shows as unknown.

Apps only. Automation Activity Over Time 2.1.18 and Quiet App Account Detector 1.19 count only the changes app accounts make, so you can see when an automation such as Automation for Jira last changed a work item and in which project it went quiet. Jira records every Automation for Jira rule under the same app account, so neither can tell one rule from another.

Visuals not named here do not look at the account type. They count every account alike. Which edition includes a visual is shown on its page in the Visuals catalog.

Five figures that read differently

  • Off-hours actors flagged 5.6.3. Apps and customers have no working hours of their own, so they are not flagged, and the panel shows how many such accounts had off-hours activity. Accounts of unknown type are still rated, because they may be people.
  • Missing Approval Heatmap 5.2.9. The rate covers closures made by people. A comment from an app or a customer does not clear a closure; a comment from any other account does, including an account of unknown type or an erased account, and so does a comment with no recorded author.
  • First Response Time 2.7.11. A response from an app or a customer account does not count as the first response, and the panel shows beside each project how many issues received one. A response from an account of unknown type still counts.
  • Self-Assignment Rate by Project 5.3.8. The rate counts assignee changes made by people. Changes made by apps, customers and accounts of unknown type are shown beside the bar, never inside the rate.
  • New Actor Burst Detection 5.3.6. An account debuts once, in the week of its first recorded event anywhere in your Jira site, so it does not debut again when it first works in another project. The burst threshold is the mean plus two standard deviations of people's first-week activity. Apps, customers, accounts of unknown type and erased accounts are compared with it and can be flagged, but do not move it, and the KPI Burst actors 5.3.2 counts the people among those flagged. With fewer than three people's debuts in the range, or when they are all the same size, there is no threshold: nothing is flagged and the KPI shows a dash rather than 0.

Deep Analysis and the account type

Deep Analysis works from the account type that a panel's data carries. It does not infer automation from an account's name, ID or activity pattern, and it says so when the data does not record the type.

What Jira records when AI works on a person's behalf

Jira records an action taken on a person's behalf under that person's account. Whether AI activity can be told apart from a person's own work therefore depends on how the AI reached the work item, not on whether AI was involved at all.

MetaFrazo counts a change made under a person's account as that person's, whatever tool acted through it. MetaFrazo does not detect AI, Rovo or AI-generated changes. It reads the account type Jira reports and nothing more.

Four ways AI reaches a work item, and they are recorded differently

Four ways AI reaches a work item, and the account each one is recorded under. An Automation for Jira rule is recorded under an app account and is labeled as app activity and shown separately, never counted inside a people figure. Rovo chat and a tool over the Rovo MCP server are each recorded under the signed-in person's own account, with nothing in the record marking the change as AI, so each is counted as that person's work. A Rovo agent is the case you choose: your Atlassian administrator picks the identity it acts under. Set to its own account it is recorded as an app account and shown separately; set to the person who runs it, it is recorded under that person and counted as their work. MetaFrazo reads the account type Jira reports for each actor. It does not detect AI.

How it actedWhat Jira recordsHow MetaFrazo counts it
An Automation for Jira rule calling Rovothe Automation for Jira app accountShown separately, as app activity like any other app
Rovo chatthe signed-in person's account, with no other markerCounted as that person's work. Nothing in the record tells it apart
A tool through the Rovo MCP serverthe signed-in person's account, with no other markerCounted as that person's work. Nothing in the record tells it apart
A Rovo agent running under its own identitythe agent's own account, which Jira types as an appShown separately, as app activity like any other app

A fifth case exists and is none of these: a Rovo agent assigned to a work item creates an agent session. Atlassian documents it as a screen in Jira, and the work item also stores it in a field called Agent Sessions, which apps can read: who started the session, the agent's account, when it started and what state it is in. It records that an agent worked on the item, not what the agent changed, so an edit made under a person's account still reads as that person's. Atlassian does not document the field, and MetaFrazo does not read it today.

Those four routes are all about Rovo. An installed app acting under its own app account is the ordinary app case, whether or not AI is involved. A Marketplace app holding permissions in your site acts as itself, and MetaFrazo shows it separately like any other app, the same treatment Automation for Jira gets in the first row above.

What the record cannot say is whether that app used AI to decide what to do. An app account tells you an app acted. Jira reports the account type, not what happened inside the app, so this is the same limit that applies to a change made under a person's account, one level up.

What the record carries, and what it does not

The issue, its change history and the event payloads MetaFrazo receives all carry the account that acted. None of them carries how the change was produced. No field marks a change as coming from Rovo, from an MCP client or from a person typing, so no product reading those payloads can tell you, and that includes MetaFrazo.

Your organization's audit log is a different record, and it does hold some of this. It needs an organization admin, and Atlassian documents a retention of roughly 180 days.

Each route in more detail, with worked examples: Three ways AI reaches a Jira work item. Only one stays separable. on the MetaFrazo blog. It covers the JQL that does and does not help, and where the evidence actually lives.

Making AI activity easy to tell apart, as a configuration choice

If you need AI work to be easy to tell apart in the record rather than reconstructed afterwards, Atlassian exposes a per-agent identity setting: a Rovo agent can act as the person who runs it, or under its own account. An agent acting under its own account is a distinct actor in the history rather than one more action under somebody's name.

Atlassian says such an account is managed by organization admins "just like users", and that changes from automated runs then appear under the agent's name.

Such an account reads as an app. Jira reports accountType: "app" for it, so the visuals above treat an agent running under its own identity exactly as they treat Automation for Jira: shown separately, labeled, and never counted inside a people figure. Nothing needs setting at the MetaFrazo end, because the rule is the one applied to every actor, which is the type Jira reports.

This is what the Jira API reports today, and what the visuals show follows it. Agent identity is an area Atlassian is still developing, so treat it as current behavior rather than a guarantee.

It is a decision for your Atlassian administrator, it applies to future activity only, and it does not reclassify history already recorded under a person's account.

Questions a reviewer could ask

QuestionWhere to look
Which changes were made by automation rather than a person?The visuals above that label or split by account type, for example Configuration Change Event Audit Log 2.2.11 and Audit Compass 5.8.1
Were closures, skips or batch changes made by people or by apps?The visuals under Control and approval, for example Segregation of Duties Violations 5.6.6 and Direct Open-to-Done Skip 5.2.7
Which accounts are apps?The account type shown on a labeled row, as Jira reports it
Did an AI agent work this item?Not answerable from the event record. If your admins use Rovo agents, Jira keeps an agent session on each work item an agent was assigned to, visible on the For You page and findable with the undocumented search agentSessions is not EMPTY. MetaFrazo does not read it today
Was AI involved in this change?The events MetaFrazo receives carry the account that acted, not how the change was produced. See What the record carries, and what it does not above for where to look and what the answer costs

Where to go from here