Ask Dawiso AI: security overview

Ask Dawiso AI is the conversational AI assistant in Dawiso. Users ask questions in plain language in the panel. Ask Dawiso AI answers from the metadata and documentation managed in the platform, with references to the objects it used.

Key principles

Ask Dawiso AI’s security model rests on a few principles that hold across every feature described below.

• Answers follow the user’s permissions. Ask Dawiso AI answers from content the asking user is allowed to see in Dawiso, and nothing else.
• Nothing changes without explicit confirmation. Ask Dawiso AI reads on its own. Any action that is not a read (creating, updating, or deleting anything in Dawiso) runs only after the user explicitly confirms it.
• Answers cite their sources. Answers carry citations linking to the Dawiso objects they are based on, so users can verify them.
• Models hosted in the EU. Ask Dawiso AI uses only Microsoft Azure OpenAI models hosted in EU data centers.
• Customer data is kept apart. Each customer instance has its own search index, and every stored conversation is bound to a single instance.
• Your data does not train the models. Dawiso does not train or fine-tune language models on customer content.

Architecture and trust boundary

Ask Dawiso AI runs in Dawiso cloud as a service behind the Dawiso application. Users never call it directly. Dawiso signs in the user, decides what that user may see, and passes the question on together with that scope.

flowchart LR
    User["User<br/>(Dawiso AI chat)"]

    subgraph Dawiso["Dawiso cloud"]
        App["Dawiso application<br/>(sign-in, permissions)"]
        Ask["Ask Dawiso AI service"]
        Index[("Search index<br/>(per instance)")]
        History[("Conversation history<br/>(scoped by instance)")]
        App -->|"Question +<br/>user's permission scope"| Ask
        Ask -->|"Tool calls as<br/>the signed-in user"| App
        Ask --> Index
        Ask --> History
    end

    subgraph Provider["Language model provider"]
        LLM["Microsoft Azure<br/>OpenAI Service"]
    end

    Help["Dawiso Help<br/>(public documentation)"]

    User --> App
    Ask -->|"Prompt with<br/>permitted context"| LLM
    Ask -->|"Product questions"| Help

Component	Runs in	Operated by
Dawiso application	Dawiso cloud	Dawiso
Ask Dawiso AI service	Dawiso cloud	Dawiso
Search index and conversation history	Dawiso cloud	Dawiso
Language model	Microsoft Azure OpenAI Service, EU data centers, under Dawiso’s account	Dawiso (service provided by Microsoft)

Language models

Ask Dawiso AI uses large language models provided through Microsoft Azure OpenAI Service. The models are managed by Dawiso. Customers do not need to set up or connect their own model account.

• EU hosting only. Every model Ask Dawiso AI uses, for answering questions and for building the search index, is an Azure OpenAI deployment hosted in EU data centers. Questions and context sent to the model are processed in the EU.
• No training on your data. Dawiso does not train or fine-tune models on customer content. Under Microsoft’s terms for Azure OpenAI Service, prompts and responses are not used to train the models. See Data, privacy, and security for Azure OpenAI Service.
• Stateless model calls. Each request carries the context needed for that answer. The model provider does not hold a memory of your conversations; conversation history is stored by Dawiso (see Data storage and retention).
• Provider safety controls. Requests pass through the content filtering of Azure OpenAI Service in addition to Dawiso’s own safeguards. Automated abuse monitoring screens every request without storing it, but if that check flags a request, Microsoft may store a sample of the prompt and completion for authorized human review.

Data flow

1. Question: The user asks a question in the panel. Dawiso authenticates the user and sends the question to Ask Dawiso AI together with the user’s permission scope.
2. Retrieval: Ask Dawiso AI searches the instance’s search index for relevant objects, limited to the spaces the user can access. When the question needs more detail, Ask Dawiso AI calls Dawiso tools on the user’s behalf, using the user’s own sign-in. Dawiso then applies that user’s permissions to every result. For questions about the product itself, it can read the public Dawiso Help documentation.
3. Generation: Ask Dawiso AI sends the question, the recent conversation, and the permitted context to the language model, which writes the answer.
4. Answer: The answer is returned to the user with citations to the Dawiso objects it is based on. Citations that do not match a retrieved source are removed.

Answers with citations

Ask Dawiso AI grounds its answers in Dawiso content rather than in the model’s general knowledge:

• Answers come from retrieved content. Ask Dawiso AI is instructed to answer only from the Dawiso objects and Help articles it retrieved for the question, not from the model’s general knowledge.
• Claims are cited. Statements in the answer carry citations that link to the Dawiso object they come from, so users can open the source and verify them.
• Citations are checked. Before the answer reaches the user, Dawiso checks every citation against the sources actually retrieved for that question. A citation the model invented is removed, so an answer never points to an object that was not used.

Data used for the answer

The language model receives only what is needed to answer the current question:

• The user’s question and the earlier messages of the same conversation.
• Excerpts from the search index, scoped to what the user can read (see Permissions and access control).
• Results of Dawiso tools called on the user’s behalf, scoped the same way.
• Public Dawiso Help articles, for product questions.

The search index is built from catalog metadata: object names, descriptions, attributes, and relations managed in Dawiso. It does not contain the underlying data of the cataloged systems. When the index is built, metadata text is sent to the language model provider to create the search representations (embeddings).

Permissions and access control

Ask Dawiso AI enforces the same permissions a user already has in Dawiso, at every step of answering a question.

• Permission-scoped answers. Dawiso tells Ask Dawiso AI which spaces the user can read. Search results outside those spaces are excluded. If no scope is provided, Ask Dawiso AI returns nothing rather than everything.
• Tools run as the user. When Ask Dawiso AI calls Dawiso tools, it uses the asking user’s own credentials, not a shared service account. The user cannot obtain information through the chat that they could not see in Dawiso directly. Ask Dawiso AI additionally filters tool results against the user’s scope before the model sees them.
• Private conversations. A user’s conversation history is visible only to that user, within their own instance.

Actions and changes

Ask Dawiso AI only reads on its own. Besides answering questions, it can help users make changes in Dawiso, such as updating an object. Any action that is not a read requires the user’s explicit confirmation, every time. The AI model cannot perform a change by itself.

• Confirm before anything changes. Every create, update, or delete is shown to the user as a confirmation with the details of what will happen. The change runs only after the user accepts it; rejecting it, or not answering, leaves Dawiso unchanged.
• The confirmation shows the real change. It is prepared by Dawiso from the actual change, not from the AI model’s description of it, so what the user approves is what runs.
• Only with a person present. Changes are offered only in a live chat, where a user is there to confirm them. Background and automated processing never performs changes.
• Never beyond the user’s own rights. The change runs with the user’s own permissions. Ask Dawiso AI cannot do anything the user could not do in Dawiso themselves.

Isolation between customers

Each customer instance of Ask Dawiso AI is isolated:

• Every instance has its own search index and its own document storage.
• Conversation history, feedback, and configuration are stored with the instance they belong to, and every read is limited to the instance of the request.
• Each request is bound to a single instance by its authentication. A request for one instance cannot read another instance’s index or conversations.

Data storage and retention

Ask Dawiso AI stores the following in Dawiso cloud:

Data	Purpose
Search index	Finding relevant catalog metadata for answers
Conversation history	Showing users their earlier chats and continuing a conversation
Feedback	Ratings and comments users give on answers
Usage metrics	Response times and usage volumes, without message content
Instance configuration	Settings and instructions for the instance’s assistant

• The search index is rebuilt from Dawiso on a regular schedule. Objects removed from the catalog drop out of the index on the next rebuild.
• Credentials Ask Dawiso AI uses to connect to the customer instance are stored encrypted.
• When Ask Dawiso AI is decommissioned for an instance, its search index and stored conversations are removed.

For data deletion requests or questions about retention, contact Dawiso support.

Safeguards against misuse

Ask Dawiso AI combines several layers of protection. Each one addresses a specific risk of AI assistants:

Safeguard	What it protects against
Input checks reject oversized questions and known prompt-manipulation patterns before they reach the model	Prompt injection: attempts to override the assistant’s instructions, for example by posing as a system message
Provider content filtering and jailbreak detection from Azure OpenAI Service	Harmful content, and jailbreak attempts that try to make the model ignore its rules
Permission-scoped retrieval and tools running as the user	Sensitive information disclosure: obtaining content through the chat that the user is not allowed to see in Dawiso
Explicit confirmation of every non-read action	Excessive agency: the model changing data on its own, or being manipulated into doing so
Grounded answers with checked citations	Hallucination and misinformation: answers that are not backed by Dawiso content, or references to objects that were never used
Bounded processing limits the number of steps and the time an answer may take	Runaway or looping behavior and excessive resource consumption from a single request
Separate search index per instance, and instance-scoped access to conversations	Cross-customer data leakage
Secure development requires every change to pass automated security scanning (static analysis and secret detection) and code review before release	Vulnerabilities and leaked credentials in the service itself