Letting the AI Search, Read and Create in Jira and Confluence

If the administration has enabled AI access to Jira and Confluence (also called AI tools, Jira/Confluence lookups or write actions), the AI can search issues and pages on its own, read them, and — only after your confirmation — create issues, comments and pages through suggestion cards. This page explains what the AI can look up, how the confirmation works, what the optional web search and image generation add, and where the limits are. Page context for the issue or page you are currently on is a separate feature — see Prompts, Skills, Memory and Model Parameters (＋ Menu).

What this does

The most important principle: the AI works with your permissions. It finds nothing that you could not find yourself, and it creates nothing that you would not be allowed to create yourself. Nothing is ever written to Jira or Confluence without your click on a suggestion card.

While the AI works, every lookup appears as a row with a tool icon in the chat, so it stays traceable what happened. These rows are always in English, regardless of your interface language:

Who can do this

All users — within their own Atlassian rights. How far the AI may go is set by the administration under AI access to Jira & Confluence: Off, Read only or Read + write. An agent's Jira/Confluence tools level can narrow this further, never widen it (see Creating Your Own Agents and Skills).

Before you start

• The chat is set up and AI access is not Off — otherwise the AI answers from general knowledge only.
• For creating and updating: the administration has enabled Read + write, and the selected agent does not narrow that.
• Know the project key (e.g. ABC) or space key (e.g. DOCS) you mean — or let the AI list the projects and spaces visible to you.
• The suggestion cards need an active AnyLLM subscription; with an inactive subscription ✓ Execute is blocked.

Step by step

Letting the AI search and read

1. Ask your question and say where to look, e.g. "Search for open bugs in project ABC and summarize them", "What does our onboarding page in DOCS say?", "Which of my issues are due this week?", "Which spaces can I see?"
2. Watch the tool rows while the answer is generated — they show the JQL or CQL used and which issue or page was read.
3. Read the answer and ask follow-up questions in the same conversation; you can also ask the AI which query it used or to search again with different terms.

Expected result — an answer based on the issues and pages the AI actually read, typically with issue keys or page titles you can check.

Having something created — with confirmation

1. State what you want, e.g. "Create a bug in project ABC from this", "Add a comment to ABC-12 with the summary", "Rewrite the description of ABC-7 with a numbered list", "Create a page 'Meeting notes 25 August' in DOCS with a table".

2. The AI shows a suggestion card with the title of the planned action and all fields:

   Card title	What it creates or changes	Fields on the card
   Create Jira issue	a new issue	project key, issue type, summary; optionally description
   Update Jira issue	the summary and/or description of an existing issue	issue key; new summary and/or new description
   Add Jira comment	a comment on an existing issue	issue key, comment text
   Create Confluence page	a new page	space key, title, content; optionally parent page ID

3. Check the fields. Descriptions, comments and page content are shown in Markdown notation (e.g. **bold**, ## Heading) — they are created natively formatted. If something is wrong, tell the AI what to change; it then shows a new card.

4. Click ✓ Execute (the button reads Executing … while it runs) — or Dismiss to drop the card.

Expected result — the card is replaced by a success message with a link to the new or changed item, exactly as the server reports it (partly German even in the English interface): "✅ Vorgang ABC-1 erstellt.", "✅ Vorgang ABC-1 aktualisiert.", "✅ Comment added to ABC-1." or "✅ Page "…" created in space DOCS." The message stays in the history. If the AI's turn consisted of the card alone, its text reads "I proposed an action — please confirm or dismiss it above."

Notes and effects

• Only the click triggers it. Typing "yes, do that" produces at most a new suggestion card; the AI is instructed never to claim it executed anything itself.
• What is created is a normal issue, comment or page — with you as the author. You undo it the same way you would otherwise in Jira or Confluence.
• Formatting included. Headings, bold/italics, nested and numbered lists, links, code blocks, tables and coloured info panels arrive as real Jira/Confluence formatting — just say what you want ("with a table", "steps as a numbered list", "with a warning panel"). Checklists (- [ ]) become real tasks on Confluence pages and ☐/☑ bullets in Jira. Tables and panels work at the top level of the text; inside lists or quotes they are flattened to plain lines.
• Update Jira issue replaces the whole description. The AI rewrites the description from what it read as Markdown: headings, lists, tables, links and code blocks survive, but embedded images and attachments, @-mentions, status lozenges and smart links become plain text or links, and descriptions longer than about 8,000 characters are cut off. The AI is instructed to warn you only about images and truncation — when in doubt, check the issue after the update and re-insert images by hand.
• Dismiss only removes the card; nothing is sent anywhere, and you can carry on in the chat immediately.
• ↻ Regenerate answer discards open suggestion cards. Cards are not exported and are not sent back to the AI.
• The administration's setting is checked again at the moment you click ✓ Execute: if Read + write has been switched off in the meantime, the card fails with "Write actions are not enabled by the admin."

Expected result

A question that needs facts from Jira or Confluence shows tool rows during the answer and an answer grounded in what was read. A request to create or change something ends in a suggestion card, and after ✓ Execute in a success message with a link — the item exists in Jira or Confluence under your name.

Web search and image generation

These two services are optional and only available if the administration has switched them on. Like the Jira/Confluence lookups, they appear as tool rows.

• Web search — the AI searches the internet when it needs current information (row Web search: …). Up to five results with short excerpts feed into the answer, and the AI is instructed to cite the sources it used as links. Only the first 400 characters of the search query are sent to the search service.
• Image generation — the AI can produce an image from a description (row Generating image: …) and embeds it in the answer. The description is limited to 2,000 characters; the app tells the AI that English descriptions often give better results. With many providers the link to a generated image is valid for a short time only — download important images promptly.

Limits

• Per answer the AI may look things up in at most 5 rounds (search, read, list, propose); afterwards it must answer with what it has. Split very long multi-step tasks into smaller questions.
• Each round may take up to 120 seconds, the whole answer at most 5 minutes. A search returns at most 10 issues or pages; project and space lists at most 25 entries; each lookup result is cut at 8,000 characters.
• Only the most recent 16 messages (about 32,000 characters) of a conversation are sent to the AI — in long chats it no longer remembers earlier turns even though they stay in the history.
• In the rare case that streaming is unavailable and the app falls back to a simple request, the AI has no Jira/Confluence access, web search or image generation for that answer.
• Every round counts as one request in the administration's usage statistics. Against your hourly limit (if the administration has set one) each message you send counts once, however many rounds it takes.

Common problems

Related topics

• Chatting with the AI — asking questions, streaming, regenerate
• Prompts, Skills, Memory and Model Parameters (＋ Menu) — including the current page or issue as context
• Creating Your Own Agents and Skills — tool level per agent
• Permissions, Roles and Data Visibility — the principles behind "your permissions"
• General Settings: Language, Default Model, AI Access, Rate Limit, Site Knowledge — the administration's AI access setting
• Enabling Web Search and Image Generation — the administration's counterpart for the optional services
• Troubleshooting · FAQ