OpenAI review source
Primary production path
When review_source is openai, GPT-5.4 returns category,
automation type, impact, feasibility, risk, projected monthly hours saved,
confidence, recommendation, and priority.
Guidance used for those scores
Impact rises with higher task frequency, more manual time, and more rework.
Feasibility rises when the steps are clearer, prompt evidence exists, and the
constraints are lighter. Risk is based on sensitivity and compliance exposure.
Priority is a weighted blend of impact and feasibility with penalties for risk and
duplicates.
Heuristic fallback
Exact deterministic formula
Impact score
Monthly task hours times 2.4, capped at 70, plus rework
hours times 1.2, capped at 20, plus a current-steps bonus
of 1.5 per step, capped at 10. Final range:
5-100.
Feasibility score
Base 45 plus: current steps +15, sample prompt
+10, tools bonus up to +15, notes +5, then
risk penalty -20 for high risk or -8 for medium risk.
Final range: 10-100.
Priority score
impact * 0.55 + feasibility * 0.35 - risk penalty - duplicate penalty.
Risk penalty is 30 for high risk, 12 for medium,
0 for low. Duplicate ideas get -10. Final range:
5-100.
Derived metrics
Hours saved, duplicates, confidence
Estimated monthly hours saved
Monthly task hours are multiplied by a feasibility ratio between 0.2
and 0.85.
Duplicate detection
Duplicate checks compare ideas only within the same company. The heuristic duplicate
threshold is Jaccard token overlap of 0.45 or higher.
Confidence
The fallback starts at 0.52 and increases when steps, prompts, tools,
or notes are present, then drops slightly for duplicates. Final range:
0.30-0.95.