Back
Control example

Resilience and failure handling

Resilience does not mean that failures never happen, but that it is known in advance what happens when an element of the workflow fails. This page is a list of decisions to make during deployment, not a promise of specific operating parameters.

The points of the flow this example concernsSee the whole flow
  1. Intent
  2. Input control
  3. Authorization (applies)
  4. AI (applies)
  5. Result validation
  6. Human
  7. Evidence

Highlighted: the points of the flow that this description refers to. Orange marker: a human decision at this point.

Explanation

Control does not disappear when something fails

The most important decision concerns the order of things: if the check cannot be performed, the operation should not be performed by bypassing it. Technically, this is called the fail-closed principle. Its opposite, fail-open, means continuing work without the check.

Failure situations and decisions to make during deployment
SituationWhat to decide during deploymentMore
The model or its provider is unavailableWhether the task waits or switches to a fallback model, and whether the fallback model is approved for the same data zone.Model selection
The check before sending cannot be performedWhether the query is held until the check is restored (fail-closed). For data designated as sensitive, continuing without the check increases the risk of disclosure.Boundaries in the workflow
An AI digital worker acts incorrectly or approaches its limitsWho can suspend or revoke its agent passport, and which operation and cost limits to adopt.Accountability for decisions
No Internet accessWhether the process is to work offline. This requires local models and all the services needed in the process, including search and OCR if they are used.Working without the Internet
Server failure in your own infrastructureBackups, the way of restoring and updating, and the person responsible for maintenance.On-premise deployment
You need to find out what happenedWho has access to the evidence trail and how long it is kept. The organization sets the evidence retention period.Documenting the workflow
Diagram

Conditions, transitions and decisions

Diagram of the controlled run

  1. Step 1: DetectionAn element of the workflow reports an error, unavailability or a limit being exceeded.
  2. Step 2: ConditionCan the operation be completed without skipping the check and within the zone’s boundaries?
  3. Step 3: DecisionIf yes, the workflow continues according to the agreed procedure. If no, the operation is held.
  4. Step 4: Limiting the impactIf needed, an authorized person suspends or revokes the agent passport under which the AI digital worker operates.
  5. Step 5: InvestigationThe workflow is reconstructed from the evidence trail, and the procedure is corrected.
Illustrative diagram of the failure handling rules to be agreed during deployment, not a view of a running system or a description of how a specific configuration behaves.
Separate example

Suspending and revoking an AI digital worker

Separate example

Actions in the agent passport

The view of the agent passport under which the AI digital worker operates has the “Suspend passport” and “Revoke passport” buttons. They are used to limit the impact when the AI digital worker acts incorrectly.

The screenshot shows AiTrust on synthetic data and is a separate configuration example, unrelated to the recorded complaint run.

A separate configuration example: an agent passport with the “Suspend passport” and “Revoke passport” actions (synthetic data).

A separate configuration example: an agent passport with the “Suspend passport” and “Revoke passport” actions (synthetic data).

The “Agent passport” view with the settings of an AI digital worker, which operates in the AiTrust application as the agent “Contract Analysis Agent”. It shows the actor type ai_agent, the issue and expiry dates, maximum zone B and the approval required above zone B. It also shows the limits on operations per hour and cost per day and the Suspend passport and Revoke passport buttons.

What the screenshots show

  • The suspend and revoke actions available in the passport view in this example.

What the screenshots do not prove

  • The effects of suspension during an operation in progress.
  • How the system behaves when a model, the network or a server fails.
  • Availability parameters (maintenance terms are agreed for each specific deployment).

On-premise and resilience

An on-premise deployment in the customer’s infrastructure means that continuity of operation also depends on the organization’s infrastructure: servers, backups and the way updates are applied.

An on-premise deployment alone does not determine whether work without the Internet is possible. That depends on the configuration and the dependencies of the given process. Details: Working without Internet access.

Let’s discuss failure scenarios for your deployment

In the demo, we will go through the list of failure situations and identify the decisions that need to be made for your infrastructure and processes.

Request an AiTrust demo

Leave your contact details. We’ll agree on the scope of the demo and show how AiTrust can support your organization’s processes.

The controller of the data provided in this form is Polski Instytut AI Sp. z o.o. We will use the data to answer your request and arrange the demo. More information, including your rights, is in the documents GDPR information notice and Privacy policy.

* required field

Want to add your company, phone or a description of the task? Open the full form