How we work with AI
Provenance, authorship, rights, eval discipline, and approval tiers across every Sprint and Factory engagement. Operating notes from real work — not generic compliance copy.
A short overview of Comuvia's governance posture — check the AI on every checkpoint, the five commitments, the five-tier approval model with advisory at Tier 1, and the four artifacts that ship with every deliverable.
The five operating commitments
Every Comuvia engagement runs against the same five governance commitments — these are the questions a buyer should expect every creative shop using AI to answer, and the answers we hold ourselves to.
Authorship
Comuvia claims, markets, and protects only the human-authored parts of a work. AI systems assist with ideation, drafting, generation, and post-production, but are never treated as authors in Comuvia records. Source — `governance/ai_generated_content_and_ip_guidance.md`.
Provenance
Every AI-touched asset carries provenance metadata — model, prompt family, generation timestamp, and rights category. Provenance travels with the file; nothing ships without it.
Rights and reuse
Client briefs name the rights category up front (Comuvia-generated, licensed-source, mixed-source). Deliverables only include reuse rights Comuvia can actually grant — appreciable AI-generated material is identified separately from human-authored material.
Eval discipline
Each engagement format has a written acceptance bar — what "good" looks like for brand fit, taste, and intent. Eval gates sit at brief, mid-production, and final. AI produces volume; humans hold the bar.
Approval tiers
Routine production runs on standing approval. Anything novel — new client, new vertical, new model class, novel risk surface — escalates one tier. The approval matrix is canonical and reviewable on request.
The five-tier approval model
"Approval tiers" above is not a slogan. Comuvia runs a five-tier model for how much autonomy any AI-assisted action is granted, and every engagement names the tier it runs at. The higher the tier, the more an action reaches the outside world without a human checkpoint — so client-facing advisory work stays low.
| Tier | What it means |
|---|---|
| Tier 0 | Observe only — the system surfaces information; a human does everything |
| Tier 1 | Draft and recommend — the system proposes; a human owns every output that ships |
| Tier 2 | Act within a bounded internal policy — reversible, logged, no external effect |
| Tier 3 | Act externally with a human checkpoint before the action lands |
| Tier 4 | Act externally without a per-action checkpoint — reserved, rare, heavily instrumented |
For an advisory engagement, almost everything runs at Tier 1: a human owns every recommendation that reaches you. There is always a named human reviewer for every AI-touched output, with explicit pass / re-run / kill criteria — not "trust the AI," but check the AI on every checkpoint, and document every check.
| Engagement | Runs at |
|---|---|
| Architecture & Decision Review Sprint | Tier 1 |
| Decision Map Workshop | Tier 1 |
| Simulation Decision Pack | Tier 1 |
| Fractional AI & Architecture Advisor | Tier 1 (client work); Tier 2 (internal preparation) |
The discipline is governed by a four-view loop — Govern (owner, policies, allowed models), Map (use cases, data, anticipated risk), Measure (evals, telemetry, thresholds that trigger human review), and Manage (gates, runtime monitoring, incident handling, improvement).
What this means for a client engagement
Concretely, the five commitments translate into four artifacts that ship alongside every Comuvia deliverable.
AI provenance manifest
A per-deliverable manifest listing every AI-touched asset, the model and prompt family used, generation timestamp, and rights category. Filed alongside the project archive.
Rights summary
One paragraph naming what rights Comuvia is granting to the client (Comuvia-generated material, licensed-source segments if any, mixed-source segments and their treatment) and what restrictions apply.
Eval log
Brief / mid-production / final eval notes per deliverable — short bullets capturing the brand-fit and taste calls made, who reviewed, and what was adjusted.
Approval audit
A timestamped record of which approval tier the engagement ran at, who approved, and whether any item escalated.
Want to dig deeper?
Comuvia's full governance canon — AI Management System, AI-generated content and IP guidance, approval matrix, copyright filing — lives in the operating documentation. Open to clients on request.
Request the governance pack →