Accountability for decisions
The organization and the people it designates are accountable for the consequences of AI actions, not the model. Control means that each AI digital worker has a record of who is accountable for it, within what boundaries it operates and who approves operations that go beyond those boundaries.
- Intent
- Input control
- Authorizationapproves the operation
- AI
- Result validation
- Humanaccepts the result
- Evidence
Highlighted: the points of the flow that this description refers to. Orange marker: a human decision at this point.
Three questions about accountability
Before an AI digital worker starts working in a process, the organization answers three questions. In the AiTrust application, the AI digital worker operates as an agent, and the answers are recorded in the agent passport.
Who is accountable for the AI digital worker?
The owner is the person who admits the AI digital worker to work and is accountable for its configuration. In the AiTrust application, this is the “Owner (human)” field in the agent passport.
What can the AI digital worker do on its own?
Its operating boundaries are the maximum risk zone, a limit on operations over time, a cost limit and the tools granted. Operations within these boundaries do not require additional consent unless the organization decides otherwise.
Who decides beyond the boundary?
Operations of designated classes require a human decision before execution. In the example passport, operations above zone B require approval.
- Agent passport
- A view in the AiTrust application with the record card of the AI digital worker that operates as an agent. The card contains the name, actor type, issue and expiry dates, owner, limits and permissions.
- Tool grant
- The AI digital worker’s permission to use a specific tool. In the example shown, no tool has been granted.
- Suspension and revocation
- Suspension puts the passport on hold, e.g. while irregularities are investigated. Revocation cancels it.
- Consent versus content approval
- Consent to perform an action and approval of the content of a specific document version differ in scope. In the process, you need to decide where each decision is required.
A human decision also occurs in an ordinary conversation: in the recorded complaint run, it is the user who decides whether to send the restricted version to the model (see Boundaries in the workflow).
Conditions, transitions and decisions
Diagram of the controlled run
- Step 1: AdmissionThe organization issues a passport: owner, validity period, maximum zone and limits.
- Step 2: Working within the boundariesThe AI digital worker performs operations covered by the agent passport within the limits granted.
- Step 3: ConditionAn operation above the designated zone or of a designated class waits for approval.
- Step 4: Human decisionAn authorized person approves or rejects the operation and is accountable for that decision.
- Step 5: Suspension or revocationIf there are irregularities, the passport can be suspended or revoked.
The AI digital worker’s passport
In the AiTrust application, this view is called the “Agent passport”. The screenshot is a separate example of a configuration on synthetic data, unrelated to the recorded complaint run.
“Contract Analysis Agent”
This is the passport of the AI digital worker that operates in the application as the agent “Contract Analysis Agent”. You can see the actor type (ai_agent), the issue and expiry dates, maximum zone B, the requirement for approval above zone B, the limits on operations per hour and cost per day, and the “Suspend passport” and “Revoke passport” buttons.
The limit values are examples. The “Owner (human)” field is empty in this example. It is worth filling in during deployment, because without a named person it is hard to establish accountability for the AI digital worker.
What the screenshots show
- The passport fields and the available actions in this configuration example.
What the screenshots do not prove
- The approval of a specific operation or who approved it.
- Enforcement of the limits while the AI digital worker is working.
- The effects of suspension or revocation in other configurations.
Related areas
Methodologies and compliance
In the platform description
Let’s establish who decides on AI actions
In the demo, we will show the agent passport under which the AI digital worker operates. We will also discuss which operations in your organization should require human approval.