Administration
AI assistant (Advanced edition)
What the assistant proposes on each screen, who can ask for it, what goes to the model and what never does, the monthly budget, and how you keep the decision.
Edition: Advanced The assistant drafts and reviews: it proposes, you decide. No proposal is saved until you have taken it in and then saved it with the screen's own button. This page says, screen by screen, what it proposes, who can ask for it and what goes to the model. The full list of what is sent is also in Security and privacy.
Screenshot to come
What is the same on every screen
- The model. A model of the Claude family, called through Forge LLMs, an Atlassian service. The app calls no other provider.
- The mention. A proposal carries the mention "Suggested by the assistant". When it comes from the model, a badge adds "Generated by AI". When the model could not answer, the badge says "Standard answer": a fixed answer of the app, with no AI, and the reason (see below).
- The decision is yours. Under the proposal: "Insert", "Replace" or "Ignore" (or, for a form, you tick the fields to take). The text you take stays a draft in the screen's field; it is saved only by that screen's save button, with that screen's rights.
- The log. Each request and each decision (taken, replaced, fields taken) is a line of the log, with the function, the object and the number of tokens. The text sent and the text received are not in it.
- Edition. Outside Advanced, the button stays visible with the "Advanced edition" label and makes no call. Without an active licence the assistant does not answer either. The 30-day trial includes a preview of the assistant, limited to about 30 proposals (see "The monthly budget" below). See Editions and licensing.
- Rights. The assistant reads the object again with your rights before calling the model. If you cannot edit or read the object, the request is refused ("You are not allowed to use the assistant here").
- Time. The model has 20 seconds to answer. After that: "The assistant did not answer in time (20 seconds)" and a standard answer.
What the assistant proposes, screen by screen
| Screen | Button | What it proposes | Who can ask |
|---|---|---|---|
| Policies, section editor | "Ask for a proposal" (Write, Complete, Shorten, Make more formal, Adapt to an active framework) | The text of one section | Whoever edits policies |
| Policies, editor | "Suggest controls" | Controls of the catalog that the text supports, each with the quoted passage | Whoever edits policies |
| Policies, submitted policy | "Summarise for the approver" | A summary of what changed, section by section | Whoever edits or approves policies |
| Policies, editor | "Suggest a flowchart" | A flowchart of the section, to take into the editor (Flowcharts) | Whoever edits policies |
| Risks | "Propose a draft" | Title, threat, vulnerability, consequence, assets and controls, from one sentence | Whoever edits risks |
| Risks | "Propose an assessment" | Inherent likelihood and impact, treatment (mitigate, transfer or avoid), controls, residual likelihood and impact, each with a reason | Whoever edits risks |
| Controls, a control's drawer | "Propose a draft" | The "How we implement it" text | Whoever edits the statement of applicability, or the person the control is assigned to |
| Statement of applicability, a row's panel | "Propose a justification" | The justification for applying or excluding the requirement | Same |
| Controls, a control's evidence | "Review with the assistant" | An opinion ("Covers what is expected", "Partly covers it", "Does not cover it"), what is missing, notes | Whoever reads the controls |
| Tests and continuous monitoring, a failing or overdue test | "Explain the failure" | The probable cause and the steps to take in Jira | Whoever reads the tests |
| Action plan, a task's detail | "Explain this task", "Suggest evidence" | An explanation, a list of evidence ideas | The assigned person, or whoever reads the controls in full |
Rights are meant as in the rights matrix: "edits policies" is the edit right of the Policies domain, and so on.
What the assistant does not do:
- It never accepts a risk. The assessment proposes mitigate, transfer or avoid. Accepting a risk is a person's decision, recorded in their name, under the separation of duties rule.
- It approves nothing, saves nothing, creates no ticket or task, changes no status.
- It does not read the content of an evidence file. "Review with the assistant" judges from the name, the type, the date and, for a frozen Jira query, the query and the number of issues. The screen says so: "It does not read the content of a file".
- It does not invent identifiers. A control, an asset or a risk level the installation does not know is removed from the answer before the screen. A level outside the risk matrix is brought back to the matrix's top level, and a residual level never exceeds the inherent one.
What goes to the model, and what does not
The cleaning, for the ten recent functions
Before sending, the text goes through a common cleaner:
- e-mail addresses, web addresses (Atlassian addresses included), Atlassian account identifiers and
@namementions are removed; - the names of the people the call knows are replaced by their role ("the person in charge", "the person asking"). A person is known when the object names them: the owner and the approver of a policy, the owner of a risk or of a control, the people assigned to a control's tasks, and you;
- the fields that carry a person (owner, author, approver, address, display name) are removed from what is sent, whatever they hold.
Limit. The cleaner knows only those people. The name of a colleague, a customer or a third party written in a free text, without being the owner of the object, is not removed. A first or last name of fewer than four letters, alone, is not removed either. Do not type names in the fields that go to the model.
The two older functions ("Explain this task" and "Suggest evidence") are cleaned the same way: the app does not send the task's owner, and removes addresses, links and account identifiers from the task's title, description and guide. A name typed in those texts is sent as written.
What each function sends
| Function | Sent to the model | Never sent |
|---|---|---|
| Policy section | The policy title, its language, the section title, the section's text as it is in the editor (up to 20,000 characters, without its flowcharts), the action chosen, the names of the active frameworks | The other sections, the other policies, the owner and the approver |
| Suggested controls | The policy title, the whole text of the policy, the list of catalog controls (identifier and title) | The owner and the approver |
| What changed | The policy title and, for each changed section (30 at most), its text before and after (20,000 characters at most each) | The sections that did not change |
| Flowchart | The title and the text of the section, the role names read in the sections of the saved policy whose title speaks of roles or responsibilities (twelve at most, five words each), the policy's language | The flowcharts already in the section, the rest of the text |
| Draft a risk | The sentence you type (1,000 characters at most), the assets of the register (identifier and name, 200 at most), the catalog controls (identifier and title) | The owner, the description of the assets, their classification |
| Assess a risk | The title, the description (4,000 characters at most), the levels and appetite of the matrix, the assets linked to the saved risk (identifier and name), the catalog controls | The owner |
| Implementation of a control | The title and description of the control, the requirements it covers, the name and type of the attached evidence (20 at most; for a frozen Jira query, the start of the query, 120 characters), the title and status of the tasks (20 at most), the answers to the assessment and their notes (20 at most, 300 characters of notes each), the declared state, the current text | The content of files, the owners of the tasks |
| Statement justification | The code and title of the requirement, the framework, applicable or not, the description of the control, the title and state of the controls that cover it, the current text of the field | The owner of the control |
| Review a piece of evidence | What the control expects, the control's title; for a file or a reference: its name (the file's original name), the date it was added, its type; for a frozen Jira query: the query (400 characters at most), the period, the date of the latest capture, the number of issues and the first ten issue keys | The content of the file, its address, the person who added it |
| Explain a failing test | The test name, its source type, its last result (status, reason code, measure), the date since when it fails, the titles of the controls it checks (five at most) | People, tickets |
| Explain a task, Suggest evidence | The task's title, description and implementation guide, the evidence hint written by us, the requirements covered, the status, the priority, the review cycle | The task's owner, the evidence, their files |
In every case, nothing else is sent: not the audit log, not the access reviews, not the people, not the vendors, not the incidents.
The text sent is presented to the model as data, with the instruction not to follow any instruction found in it. The answer must have a precise shape; an answer that does not is never shown: you get the standard answer.
The monthly budget
- The model consumes tokens. By default the app allows 750,000 a month per installation (calendar month, UTC). One person can use only 40% of it: they cannot drain the others' share.
- During the 30-day trial the assistant is open with a preview limited to about 30 proposals: a budget of 100,000 tokens for the whole trial, not per month. It does not go back to zero at the change of month (a 30-day trial often spans two). The personal share of 40% applies to it too. Once it is used up you get the standard answer and the screen says "The assistant trial is over: the AI assistant is included in the paid Advanced subscription. The rest of the app works normally."; no reset date is announced, because there is none. The other Advanced features stay open.
- A proposal given by the model shows "N tokens left this month". When the number is not known, or when the request was refused (standard answer), the line is not shown: there is nothing to count.
- This budget is set by the app's publisher for all sites. There is no setting in the app, neither to change it nor to switch the assistant off.
When the limit is reached
There are two limits, and the screen says which one:
- The site's budget is used up: "The site's AI budget is used up this month. The counter starts again on the 1st of next month (UTC); until then the answer is standard, without a model. The rest of the app works normally."
- Your personal share is used up: "Your share of the AI budget is used up this month. The counter starts again on the 1st of next month (UTC); until then the answer is standard, without a model. Your colleagues keep their own share; the rest of the app works normally."
In both cases:
- the counter goes back to zero on the 1st of the next month, at midnight UTC (the month is the universal-time one, not your time zone's);
- until then, every request gets the standard answer, computed by the app without a model (see below);
- the rest of the app works normally: nothing else is limited;
- there is no top-up: tokens cannot be bought nor the limit raised from the app, and no administrator can do it. You have to wait for the 1st of the next month.
When you get a standard answer
The badge says "Standard answer" and a sentence gives the reason:
| Reason shown | What happened |
|---|---|
| "The AI assistant is part of the Advanced edition." ("This feature is part of the Advanced edition." on the button) | The installation does not have the Advanced edition, or the licence is not active |
| "The assistant trial is over: the AI assistant is included in the paid Advanced subscription. The rest of the app works normally." | The licence is a trial and its budget is used up |
| "The site's AI budget is used up this month. The counter starts again on the 1st of next month…" | See "When the limit is reached" |
| "Your share of the AI budget is used up this month. The counter starts again on the 1st of next month…" | See "When the limit is reached" |
| "The assistant is not available on this site." | The model is not available for this installation |
| "The AI counter is unavailable." | The token count does not answer: nothing is sent |
| "The assistant could not answer." | The model failed |
| "The assistant's answer could not be used." | The model answered, but not in the expected shape |
The standard answer is computed by the app, with no model: a skeleton to complete for a section ("To write: …"), a word-matching reading for suggested controls, the list of gaps for "what changed", a chain of the listed steps for a flowchart, the matrix's middle values for an assessment ("to be assessed by a person"), and never better than "Partly covers it" for a piece of evidence. It is not an AI proposal: the screen does not present it as one.
Where the language is set
Proposals are written in the installation's language (Language). For a policy section and a flowchart, the policy's language applies.
Known limits
- The assistant judges from what it is sent. An opinion on a piece of evidence that rests only on a title is called "partial"; it is not an auditor's opinion.
- It does not replace review: the text it proposes is a draft, and a proposal can be wrong.
- Where the model processes the request and how long Atlassian keeps it has not been verified for this app: see the Forge LLMs documentation.
- The result never says that an organisation complies with a standard.
Was this page helpful?