devguard

AI Assistant

Ask questions about your compliance data in plain language, and let the assistant create risks, policies and other records for you to check.

The AI assistant is a chat that reads your organization's data in devguard. Ask it which risks are high, which assets need a review or what a control asks for, and it answers from your own records. It can also do work for you. It creates risks, policies, tasks and other records, and it maps controls to a policy. It has your permissions, so it can only see and change what you can.

The tasks below follow the order of the work:

Scripts and integrations: see For developers below.

The assistant panel opened from the task "Review access control policy": a suggested question, the Explore and Create starting points, the task chip with its list of shared fields, and the question filled in.
Opened from a task, the panel attaches the task and suggests a first question.

Open the assistant

There are two places to use it. Both show the same conversation.

  • The full page: choose AI Assistant in the sidebar.
  • The panel: on any page, choose AI Assistant in the top bar (the sparkle button), or press Cmd+J (Ctrl+J on Windows). The panel opens on the right, next to the page you are working on. On a task, Ask AI opens it too.

In the panel, Open full view moves the conversation to the full page, and Close hides the panel. Drag the left edge of the panel to make it wider. Double-click the edge to reset the width. On a narrow screen, the panel opens over the page.

Ask a question

  1. Type your question in the box at the bottom, marked Ask anything about your policies, assets, risks, and more. Or choose a starting point under Explore or Create. It fills in the box for you.
  2. Choose Submit (the arrow button). The assistant looks up your data and writes the answer.
  3. Under an answer, choose Copy to copy it, or Retry to get a new answer.

Questions that work well name the records and the condition you want, for example:

  • "Which risks have no owner?"
  • "Which evidence records are overdue for review?"
  • "Which incidents are still open, and how severe are they?"

To dictate instead of typing, choose Voice input (the microphone). Your browser turns the speech into text. If your browser has no speech recognition, the button is grayed out.

If the answer is Error processing request, the request did not go through. Wait a moment and send it again.

The conversation stays open while you move between pages. devguard does not save it: when you reload the page or sign out, the next conversation starts empty.

Ask about one record

When you open the panel on a record, such as a risk, a policy or a task, the record is attached to the conversation. It shows as a chip above the box where you type, with its type and title. You can then ask "Summarize this" or "What is missing here?" without naming the record.

The chip also shows what is sent: only the fields in its tooltip go to the model with your question. For a person, these fields include the email address.

  1. Open the record and then the panel. On a task, choose Ask AI. It also fills in a first question.
  2. Hold the pointer over the chip to see which fields are shared with the assistant.
  3. Ask your question. To ask without the record, choose Remove context (the cross on the chip).

When you open another record, the chip changes to that record. The list of records you can attach is in the reference below.

How the assistant works

The assistant is a language model with a set of tools. A tool is a fixed function in devguard, such as "list the risks" or "create a policy". For each message, the model chooses which tools to call, reads their results and writes the answer.

The assistant has your permissions. It sees and changes only what you can, and only in your active organization. A member can ask questions, but a request to create a risk or a policy fails. An auditor cannot use the assistant.

The model runs on AWS Bedrock. The trust page says where requests are processed and lists the subprocessors.

The model can misread a question or a result. Check numbers and new records before you rely on them.

Let it create records

A question only reads your data. A request to create something writes into your register at once. The assistant can:

  • create risks, policies, assets, incidents, vendors, processes, tasks, records and business roles;
  • add risks and policies from the templates;
  • map a control you name to a policy, or map all matching controls automatically.

The full list of what it can create and change is in the reference below.

  1. In the box at the bottom, write the request in plain words. Name the owner, the category or the framework where it matters. For example: "Create a risk 'Loss of a laptop in transit', owned by the Head of IT."
  2. Choose Submit, then read the answer. The assistant says what it created.
  3. Open the new record and check it. Change what is wrong and add what is missing.

Watch out: nothing waits for your confirmation. Check each record the assistant creates, and delete the ones you do not want.

The audit log records each change under your name. It does not mark the change as made by the assistant.

Check your usage

Each message you send counts as one AI request. A request that fails with an error does not count.

There is one exception. With automatic mapping, the assistant picks matching controls for a policy and links them. If it ran automatic mapping before the error, the request counts. That step uses the model itself, even when it links no control.

  1. Open Settings and the Billing tab.
  2. In the Usage section, read the AI Usage card. It shows the requests used this month, the limit and the date the count resets.
The AI Usage card: AI requests 0 / 1000, an empty bar and Resets: 01.11.2026.
Acme AG is on the Business plan: 1000 requests a month, none used yet.

Each plan also has a monthly budget for model costs. The app does not show the budget itself. When 90% of it is used, a meter appears next to Submit.

Once 80% of the requests are used, a banner above the chat shows how many are left. When the requests or the budget run out, the chat shows You have reached your monthly AI limit with the reset date. On the Free plan it also offers Upgrade to Business. If the AI Usage card still shows requests left, the budget is what ran out.

Answer suggestions in questionnaires and the check builder in Actions use the same monthly requests. For questionnaires, one request covers a batch of up to 10 questions.

Turn AI on or off

Owners and admins can turn AI off for the whole organization. The switch covers every AI feature: the assistant, the check builder, the PDF import of policies and questionnaires, questionnaire auto-fill, the AI matching of questionnaire answers and the knowledge base search of connected AI clients. Other roles cannot change the switch.

  1. Open Settings. Stay on the first tab, Settings.
  2. On the AI features card, turn the switch off. A message confirms the change.
The switch goes off, a message confirms it, and the sidebar item now shows Unavailable.
The Settings tab after the switch was turned off: the AI features card with the switch off, and in the sidebar the AI Assistant item with an Unavailable badge.
With the switch off, the sidebar item shows Unavailable.

When it is off, the sidebar item shows Unavailable, the panel and the Ask AI button are hidden, and the full page sends you to the dashboard. In questionnaires, Auto-fill N unanswered and the PDF import are hidden, and library matches use the wording only. Connected AI clients lose the knowledge base search, and their other read-only queries keep working. Nothing of your organization's content goes to a model. Turn the switch on again to bring AI back.

Reference

Limits

PlanRequests per monthRequests per minute
Free1010
Business100020

Each plan also has a monthly budget for model costs, which the app does not show. The request count and the budget reset on the first day of each month.

Records you can attach

Tasks, policies, risks, assets, vendors, controls, evidence, audits, findings, processes, process submissions, people, devices, incidents, records, reviews (risk, asset, vendor and access reviews), trainings and integration checks.

What it can create and change

  • Create: assets, risks, incidents, vendors, vendor assessments, policies, policy categories, processes, records, record categories, tasks and business roles. Also collection entries: labels, locations, schedules, variables, acronyms and definitions.
  • From templates: risks and policies.
  • Change: the title, category, assignee, due date and status of a document in Records, but not its text. Also a task's status.
  • Controls: map a control to a policy or remove the mapping. Map controls automatically, and undo that.
  • Automations: edit the draft of an integration check. The assistant never publishes the draft.
  • Knowledge base: save a note that the assistant can find in later conversations. Only owners and admins can save one, and each note is recorded in the audit log with its author. Any role that can use the assistant can rebuild the search index, which leaves no audit log entry.

Permissions

RoleAsk and readCreate and changeTurn on or off
Owner, adminYesYesYes
MemberYesNo, the request failsNo
Auditor, employeeNoNoNo

An auditor does not see the assistant in the app. Employees use the portal only.

For developers

The chat has no public API. Its endpoint, POST /api/ai/, works only with a signed-in browser session, and it is not part of the API reference.

  • Your own AI client: to let Claude, Cursor or another client read your devguard data, connect it to the MCP server. See Connect your AI. That access is read-only, and your own model writes the answers.
  • A link with a question: /ai/?prompt=<text> opens the full page with the box filled in. The user still chooses Submit. Encode the text, for example /ai/?prompt=Which%20risks%20are%20high%3F.
  • The switch also applies to direct requests: with AI turned off, POST /api/ai/ returns 403.
  • Who may turn it off: the switch saves through the organization update in the sign-in service, which only owners and admins may call.

The endpoint answers with these status codes:

StatusWhen
200The answer streams. If the model fails after the stream started, the status stays 200 and the stream ends with the error text "An error occurred while generating the response."
401No signed-in session, or no active organization.
403The role may not use the assistant, or the assistant is off.
429A limit is reached. The JSON body has reason (rate, quota_requests or quota_cost), resetAt and plan. For rate, it also has retryAfter in seconds and a Retry-After header.
500The request failed before the answer started.

How is this guide?

On this page