Controlled AI

AI can help summarize information, transform unstructured data, or interpret an unusual request. Sovereign's approach is to make those capabilities available with explicit limits on what the model can access and do.

AI is optional. Orbit serves work management and collaboration; Workflows serves process automation. Neither depends on introducing a model into the team's work.

This page explains the approach to controlled AI, including optional AI steps in Workflows. For the automation application itself, start with Workflows.

Even if inference were free

Assume tokens cost nothing. The hardware is free, capacity is unlimited, and running inference competes with no other work.

You would still need to know which operations a process may perform, what happens when a step fails, and which person or service authorized its actions. You would still want repeatable rules to remain repeatable.

An AI model can produce different answers to the same request. Making the entire process depend on that judgment creates uncertainty even when cost is irrelevant.

Define the process explicitly

A workflow can describe a known sequence: receive a request, validate required fields, retrieve an authorized record, obtain approval, update a system, and record the outcome.

Those steps do not all need inference. Their rules, branching conditions, and failure handling can be defined directly.

Deterministic orchestration means the process follows defined rules. It does not mean an external API can never fail, source data never changes, or an AI step will always return the same answer.

Put AI where interpretation helps

AI can be valuable for summarizing a long request, extracting structure from free text, or suggesting a route based on its contents.

Consider an intake process:

1. Validate the request using explicit rules.
2. Let an AI step summarize the request and propose a category.
3. Validate that category against the allowed choices.
4. Route using defined conditions, with human review for exceptions.
5. Perform the authorized operation and record its outcome.

The model contributes to the step that needs interpretation. The surrounding process determines how that contribution is used.

Bound both tools and permissions

A tool list controls which operations an AI step can request. Permission checks control which resources and actions the execution identity is authorized to use.

The Workflows design uses both. An available search tool should not imply access to every document. An available banking tool should not imply permission to perform every banking operation.

For actions taken on behalf of a person, that person's permissions remain part of the boundary. For separately authorized operations, the workflow must use the explicitly configured service authority.

Preserve human oversight and attribution

Some processes need approval before an operation proceeds. Others need an exception routed to a person rather than another model attempt.

The design also calls for actions to be represented in the application audit trail, beyond a model's inference history. Operators should be able to distinguish the initiating person, the executing identity, and an AI step acting on that person's behalf.

Use inference where it earns its place

Once real costs and capacity return to the picture, explicit automation has an additional advantage: steps that follow rules do not require model inference.

That leaves AI capacity for the work that benefits from it. The savings depend on the process and model; reliability, bounded authority, and accountability remain reasons to use the approach regardless of price.

Explore delegated authority