Clear boundaries.
Considered control.
For that reason, access, data handling and human oversight belong in the design conversation from the start. Consequently, the controls should fit the work and the consequences of an error. For example, a low-risk recommendation needs different safeguards from an automated action that changes a customer record.
Start with what the workflow needs.
Specifically, we identify the information involved and the people or systems that need to use it. Therefore, those requirements inform the access model, integrations and choice of platform.
- First, limit data collection and sharing to the purpose of the workflow.
- Next, define user roles and system permissions around the tasks they perform.
- Also consider how development, testing and production environments are separated.
- Finally, review vendor capabilities, data terms and integration constraints before selection.
Match the approval to the impact.
For example, an AI-generated suggestion and an automated action can have very different consequences. So, we distinguish between them when defining what a system can do and when a person must review it.
- At this stage, scope the knowledge and tools available to an AI-assisted workflow.
- Then, validate outputs against the rules the process needs to satisfy.
- Afterwards, define approval points for uncertain results and sensitive actions.
- As a final safeguard, provide an exception route when the workflow cannot proceed reliably.
In practice, the appropriate controls depend on the data, jurisdiction, business impact and client-approved risk position. Consequently, standard automation or manual review may be the better choice for some workflows.
Make important actions traceable.
As a result, the team operating a system needs to understand what happened, what failed and who should respond. Therefore, we consider those needs alongside the feature requirements.
- Initially, identify the actions and exceptions that need a reviewable record.
- In addition, agree who monitors integrations and responds to failures.
- Also, define maintenance, access ownership and support responsibilities.
- Over time, assess changes to vendors, data or business rules.
