Methodology
The Sovereign AI Operations Framework.
Five layers that turn AI from a chat window into governed operational infrastructure. The approval queue is the product. Everything else supports it.
- 01
Integration
OAuth, webhooks, and polling into the systems you already run.
- 02
Workflow
Deterministic state machines with narrow, scoped task agents.
- 03
Guarded LLM
Model calls bounded by input, output, topical, and policy rails.
- 04
Approval Queue
the wedgeEvery customer-facing action lands here for human sign-off.
- 05
Audit
Every tool call, decision, rail trip, and cost is logged.
01
Integration
What it does
OAuth, webhooks, and polling into the systems your business already runs — Gmail, Outlook, forms, CRM, Xero, Sheets. Events are normalised into a single consistent stream.
What the client sees
A connections page with scope per integration, status indicators, and a revoke button on every connection.
Why it matters
Least-privilege access by default. We connect read-only first, and request write scopes only when a workflow needs them.
02
Workflow
What it does
A deterministic state machine plus narrow, scoped task agents (LangGraph). No open-ended autonomy. Each workflow has a defined trigger, defined stops, and defined outputs.
What the client sees
Every workflow is a named, documented operation — “Lead intake”, “Quote prep”, “Inbox triage” — with explicit inputs, steps, and outputs.
Why it matters
Predictability is the precondition for trust. Open-ended agents do not survive contact with real businesses.
03
Guarded LLM
What it does
The model call sits inside guardrails, not behind them. Input rails validate the request before the LLM sees it. Output rails check the response before it leaves. Topical, dialog, and policy rails constrain behaviour at every step (NVIDIA NeMo Guardrails rail types). Model calls run via Amazon Bedrock (Sydney region) or Azure (Australia East) where the chosen model is supported.
What the client sees
Per-workflow model and region documented in the policy. Provider can be restricted by client policy.
Why it matters
Data stays onshore. Behaviour stays in scope.
04
Approval Queue
What it does
Every customer-facing or system-changing action lands here for human approve / edit / reject.
What the client sees
A single inbox of decisions. Each item shows the proposed action, the source context that produced it, a diff if it’s an edit, and a one-tap approve / edit / reject control. Audit metadata is captured automatically.
Why it matters
This is the product. Without the approval queue, AI is a liability. With it, AI becomes operational infrastructure.
05
Audit
What it does
Every tool call, every approval decision, every rail trip, every cost is logged. The monthly client report rolls it up into business outcomes.
What the client sees
A searchable activity log and a monthly PDF that lands in the owner’s inbox.
Why it matters
Governance only works if it’s visible. Logs turn AI from a black box into a system you can manage like any other operational function.
Sidebar — Outside the v1 stack
Premium runtime.
Premium runtime — NemoClaw / OpenShell. Reserved for sandboxed autonomous workloads, security-sensitive deployments, and enterprise customers who require it. Not part of the v1 default stack. Available on request as part of an Embedded subscription.
Operating engine room
The skills that prepare your deployment.
Behind every audit, pilot, and monthly report sits a small library of named, narrow, advisory skills. They produce the deliverables. They do not act on client systems.
Client Intake Analyst
Converts onboarding answers into a structured client profile.
Workflow Discovery Analyst
Identifies repetitive workflows worth automating.
Data Sensitivity Classifier
Rates client data and workflow risk before automation is proposed.
Integration Planner
Determines required system connections and the minimum scopes for each.
AI Opportunity Scorer
Ranks workflow candidates by ROI, repeatability, and risk fit.
Approval Policy Designer
Defines where humans must approve, who, and on what timeline.
Guardrail Spec Writer
Produces input, output, topical, and policy rails per workflow.
Pilot Proposal Writer
Turns audit outputs into a fixed-scope pilot proposal.
Monthly Impact Analyst
Converts logs and metrics into the client-facing monthly report.
Expansion Opportunity Analyst
Finds the next workflow to automate from usage logs and feedback.
Skills are advisory and preparatory. They run during discovery, onboarding, workflow design, reporting, and optimisation — never in the live workflow itself. They cannot send emails, modify CRM records, publish reports, or execute client-system actions. Workflow execution remains deterministic and approval-gated.
In closing
Three propositions the framework rests on.
Governance is not theatre. Approval queues, logs, and rails are not paperwork. They are the only things that make AI survivable inside a real operating business.
Narrow workflows beat open autonomy.A scoped Lead Intake agent that runs every day under approval beats a “general AI assistant” that runs once during the demo and never again.
The approval queue is the entire commercial thesis. Everything else in this framework exists to make the approval queue trustworthy enough to use.
Ready to see where this lands in your business?
A 15-minute audit call, no slide deck.