# SXF Agent Opportunity Atlas — methodological disclosure
Version 1.0, October 22, 2026

Fifty use cases are original SXF editorial design scenarios. They are NOT verified examples of fifty companies using AI agents. Four mutually exclusive suggested workflow classifications: agent candidate, deterministic workflow, AI-assisted drafting and human decision required. The classification is a provisional starting point for a concrete business pilot, not a general ban or universal engineering solution.

The Opportunity Index is a transparent **heuristic**, not a validated probability, objective measure of AI readiness, certification, safety score or guarantee of ROI. Users select 0–5 planning variables; the score must not be used to authorize high-stakes actions or to rank people.

Decision comparison revision: the page now compares **an incumbent deterministic workflow** (per-task operating cost, monthly platform fees, and full human handling for its exception share) with a **hypothetical agent proposal**. It does not claim the agent is superior when cost alone is lower: real acceptance, quality, privacy, safety and workflow fit must be independently tested. The optional Opportunity Index only screens patterns and cannot override mandatory human authorization.

ROI calculator assumptions: baseline human cost = monthly task volume × time per task / 60 × fully loaded hourly cost. Incumbent workflow cost = task volume × current per-task cost + current platform/maintenance + task volume × manual-exception rate × baseline labor per task. Proposed agent cost = task volume × model/tools cost per attempted task + agent platform/maintenance + manual labor for rejected tasks + review labor for accepted tasks. The monthly delta equals incumbent cost minus agent cost. The first-year delta equals 12 × monthly delta minus one-time setup. Accepted tasks are assumed to require only the specified review fraction of baseline task time. One-time setup is deducted from illustrative first-year savings. Not included: uncertainty, rework, error loss, management, audit, integration, support or migration. This is arithmetic, not observed company ROI.

The offline Python lab uses 120 **fictional** tasks, with predetermined input-invalid, human-approval and response-loss-after-side-effect faults. It contrasts a naive at-least-once side-effect handler with explicit guard checks and idempotency. No cloud AI, model calls or provider reliability testing. Reproduce: `python sxf-agent-workflow-lab.py`.

Source methodology:
- NIST AI RMF / AI 600-1 guidance on risks, governance, human oversight: https://www.nist.gov/itl/ai-risk-management-framework
- Anthropic guidance on workflows versus agents: https://www.anthropic.com/engineering/building-effective-agents
- Google Cloud architecture on agents and orchestration: https://cloud.google.com/architecture/choose-design-pattern-agentic-ai-system
- OpenAI practical agent-building guide: https://openai.com/business/guides-and-resources/a-practical-guide-to-building-ai-agents/
- OWASP LLM and Agentic application security overview: https://genai.owasp.org/

A credible external industry benchmark would require pre-registered tasks, authorized real user data, independently checked acceptance criteria, run traces, actual bills, error rates, and repeat trials. None of those measurements is asserted for the 50 editorial scenarios.