Boundaries in the workflow
An operating boundary determines what may reach the AI model and on what terms. We show one recorded run: the user asks for a reply to a complaint, and before sending, AiTrust shows which data will be replaced with tags and which version the model will receive.
- Intent
- Input control (applies)
- Authorization (applies)
- AI
- Result validation
- Human
- Evidence
Highlighted: the points of the flow that this description refers to. Orange marker: a human decision at this point.
What a boundary in the workflow means
Queries in configured paths go through AiTrust. The model receives the context after sanitization of designated data categories. The organization decides which categories are designated and what happens to them.
- Check before sending
- A check of the query content before it reaches the AI model. The user sees the result of the check and the version that will be sent.
- Sanitization
- Replacing detected data with tags, e.g. an e-mail address with the [EMAIL] tag. It limits data exposure, but on its own it is not irreversible anonymization.
- Zone
- A risk class marked with a letter and assigned according to the organization’s rules. The rest of the workflow depends on the zone, as described in more detail in the Operation risk assessment example.
- Configured path
- The route of a query from the user to the model, set up in AiTrust, e.g. a conversation with a selected AI digital worker. Data sent from private accounts, browsers or other applications requires controls on the organization’s devices and network.
The full flow from input to record
A technical version of the diagram from the home page: the same stages with the mechanisms expanded, without rules or configuration parameters.
An explanatory diagram, not a live feed from a production system.
- 01 · IntentInstruction and contextStage 1 of 7 · Intent
Instruction and context
The user's message enters the run together with the context of the project, the task and the related documents.
- Task and project context
- This user's conversation history
- Project documents
- AI digital worker assigned to the project
- 02 · Input controlAccess, document status and contentStage 2 of 7 · Input control
Access, document status and content
The system checks permissions to the conversation, documents and models, whether a document is ready for use, and the content of the instruction.
- Permissions and access level
- Quarantine and document status
- Cleaned version of the document
- Protection against prompt injection
- Data sanitization
Before controlPrepare a reply to a customer complaint. Contact details from the report: e-mail jan.kowalski@example.invalid, phone +48 600 000 000.Version sent to AIPrepare a reply to a customer complaint. Contact details from the report: e-mail [EMAIL], phone [PHONE].An example on synthetic data. Detected identifiers are replaced with tags; the scope of detection depends on the configuration.
- 03 · AuthorizationZone and decisionStage 3 of 7 · Authorization
Zone and decision
The H.A.D.A. zone and the organization's policy determine the next step. It can be execution, limiting the content, a request for approval or a stop. In the panel before sending, the user decides whether to send a restricted version. Operations of designated classes wait before execution for the decision of an authorized person, who does not approve their own request. The approval covers the operation before it is executed.
- H.A.D.A. zone (A to D)
- Organization policy
- Access to models by role
- Tool permissions
- Control panel before sending
- Approval requests
Execute (scope permitted for the task)Approve operation (consent of an authorized person before execution)Stop (the operation is not executed)The H.A.D.A. zone and the organization's policy determine the next step. The policy can also limit the content by sending a masked version.
- 04 · AIExecution within the permitted scopeStage 4 of 7 · AI
Execution within the permitted scope
The model and tools work within the scope permitted for the role, the project and the zone; tool use leaves a confirmation.
- Local model
- Approved external model
- Project tools
- Tool execution confirmation
Local modelApproved external modelThe model receives the context after the specified data categories have been sanitized. The choice of model follows from the configuration.
- 05 · Result validationResponse checkStage 5 of 7 · Result validation
Response check
The response is checked before display, and restricted content is not disclosed to people without the required access.
- Response check
- Content restriction by access level
- Verification against sources, if enabled
- Check of content returned by tools
- Response check before display
- Verification against sources, if enabled
Limit of the assessment: The assessment does not determine whether the result is correct in substance; a person assesses it.
- 06 · HumanAcceptance of the resultStage 6 of 7 · Human
Acceptance of the result
After result validation, a person accepts the result or sends it back for correction. The acceptance covers the prepared result, while consent to the operation is given earlier, at authorization.
- Accept result
- Send back for correction
- Accept result
- Send back for correction
Decisions in the run diagram. The acceptance covers the prepared result after its validation, before it is used further in the process. Consent to the operation itself is given earlier, at authorization.
- 07 · EvidenceA record of successive stagesStage 7 of 7 · Evidence
A record of successive stages
Successive stages leave a record, and the send report with the zone, decision and redactions is shown with the response.
- Event record at successive stages
- Send report with the response
- Log view with filters and export
The scope and order of the mechanisms depend on the specific deployment and process; some controls apply conditionally, according to the configuration.
Conditions, transitions and decisions
Diagram of the controlled run
- Step 1: InstructionThe user types a task in a conversation held in a configured path.
- Step 2: CheckBefore sending, the content is checked for the data categories designated by the organization.
- Step 3: Policy decisionIf designated data is detected, the policy determines the next step. In this example: restricting the content (Redact).
- Step 4: User’s choiceThe user sees the version to be sent and can cancel, edit the instruction, check again or send the restricted version.
- Step 5: Sending and reportThe model receives the sanitized version. The send report is shown with the response.
One example in three screenshots
The three screenshots come from one recorded run in AiTrust. The data is synthetic: the user, the complaint and the contact details were made up for this example.
Instruction in a conversation with the AI digital worker
The user writes to the AI digital worker, which operates in the AiTrust application as the “PM Orchestrator” agent. The user asks for a reply to a customer complaint and includes the complainant’s contact details in the instruction. The model is set to “Auto (by policy)”, which means it is selected according to the organization’s rules.
The caption at the top, with the word “INTENT” and the description “the user requests an action”, was added to the recording and is not part of the interface.
Check before sending
The check panel assigns zone D to the query and shows the version that will be sent to AI: the e-mail address, phone number and account number were replaced with the [EMAIL], [PHONE] and [IBAN] tags. In total there are 3 redactions, that is, masked fragments of the content. The policy decision is Redact, which means restricting the content.
The user decides whether to send: they can cancel, edit the instruction, check it again or send the restricted version.
Result and send report
After the restricted version was sent, the model prepared a draft reply. The tags from the version sent to the model appear in place of the contact details. The “AI send report” bar summarizes the check: zone D, redact, 3 redactions.
The “Message sent” notice refers to sending the query to the model, not the reply to the customer. The draft requires review and completion by a human.
What the screenshots show
- The AiTrust interface in this one example and this configuration.
- The version of the query prepared for sending, the choices available to the user and the report attached to the response.
What the screenshots do not prove
- The effectiveness of data detection in other queries or other data categories.
- Stopping a send in other scenarios.
- How the server-side mechanisms work (the screenshots show the interface).
- Later approvals, sending the reply to the customer or preparing offers.
Related areas
Methodologies and compliance
In the platform description
See the check before sending in a demo
We will show the workflow on synthetic data and discuss which data categories and decisions are worth setting for your processes.