Controlled AI orchestration · Reference implementation
Human-controlled AI workflows that keep execution outside the model
MAWI turns business signals into structured decisions, approval-gated actions and auditable outcomes, with deterministic fallback when model output fails.
Functional portfolio/reference implementation
The business problem
Recommendations are not enough for operational work
Many AI systems stop after proposing a next step. An operational system must also control state, apply policy, collect approval, execute through bounded tools, handle failures and record evaluation evidence.
MAWI separates those responsibilities so model output remains a proposal—not an external action or the source of workflow truth.
Model-only behavior
Business signal → model recommendation
The recommendation does not establish durable state, approval authority, safe execution or an inspectable outcome.
Controlled-system behavior
Signal → structured decision → policy → approval → simulated execution → evaluation record
The orchestrator preserves workflow state while policy and people determine whether a proposed action may proceed.
Current-state workflow
A visible path from signal to recorded outcome
Three proof points
Control is part of the workflow contract
Structured workflow contracts
Typed inputs and outputs give validation, orchestration and tools a defined boundary instead of passing unbounded model text through the system.
Human-controlled execution
Approval gates keep proposed external actions pending until a person makes the required decision.
Deterministic fallback and auditability
Deterministic mode remains the default and preserves workflow continuity when model output is unavailable or invalid, while audit history records the path taken.
Product walkthrough
Six views of the controlled workflow
The approved product screenshots are not published on this site yet. These reserved positions document the intended walkthrough without linking to missing assets.
01 — Signal intake
Screenshot reserved for the business-signal intake view.
02 — Structured proposal
Screenshot reserved for the typed decision output.
03 — Validation and policy
Screenshot reserved for validation and policy results.
04 — Approval queue
Screenshot reserved for the human approval decision.
05 — Simulated execution
Screenshot reserved for the bounded tool result.
06 — Audit and evaluation
Screenshot reserved for persisted history and evaluation evidence.
AI versus controlled-system behavior
The model proposes. The system decides what can happen.
Proposal and control
- The model proposes a structured decision from the available signal.
- The orchestrator controls workflow state, transitions and sequencing.
- Policy controls whether a proposal may proceed or needs review.
Approval and evidence
- Humans approve gated actions before execution.
- Tools execute only bounded, simulated actions in this implementation.
- Evaluation records workflow outcomes as reusable evidence.
Architecture
Execution remains behind orchestration, policy and approval boundaries
This page translates the current architecture into the same components and responsibilities exposed by the reference implementation; it does not add integrations or workflow services.
Decision path
Evidence path
Reliability controls
Failures return to a controlled path
Before a decision
- Typed outputs
- Validation
- Retry behavior
- Deterministic fallback
Before an action
- Policy checks
- Human approval
- Bounded simulated tools
After every transition
- Audit history
- Persistence
- Evaluation evidence
Expected value
Value categories for controlled workflow operations
- Reduce manual coordination
- Make decisions inspectable
- Control external actions
- Preserve workflow continuity during model failure
- Turn workflow outcomes into reusable evidence
Maturity and limitations
A functional reference implementation—not a production deployment
Current maturity: Functional portfolio/reference implementation with local SQL persistence.
Simulation boundary: Simulated email and CRM tools only. The project does not connect to real email, CRM or customer systems.
Claims not made: No production customer deployment claimed. No scale or multi-tenant claim. No real revenue-impact claim.
Roles and projects
Interested in a controlled AI workflow?
Let’s discuss the operating decisions, approval boundaries and evidence your workflow needs.