powerbi-autonomous-forecast-agents

name	powerbi-autonomous-forecast-agents
description	Use when coordinating multiple AI forecasting perspectives for Power BI sales forecasts, including backlog, seasonality, budget, sales skepticism, risk, consensus, dissent, and explainable forecast arbitration.

Power BI Autonomous Forecast Agents

Use this skill when the user wants an agent council rather than one blended forecast. The goal is a structured debate that exposes why the forecast is high, low, risky, or trustworthy.

Agent roles

• Backlog Agent: trusts open orders, delivery dates, status, order age, and learned conversion.
• Demand Agent: trusts seasonality, trend, intermittent demand, and customer/product history.
• Budget Defense Agent: explains the gap to budget and roll forecast.
• Sales Skeptic Agent: searches for optimism bias, volatile customers, sparse products, and weak evidence.
• Risk Agent: flags data quality, missing snapshots, memory fallback, biased backtests, delivery risk, and low-confidence matching.
• Arbitrator Agent: creates the final consensus, dissent, confidence, and recommended action.

Workflow

1. Load the latest ai-forecast-summary.csv, ai-forecast-detail.csv, ai-forecast-top-deltas.csv, and ai-forecast-model-quality.csv when available.
2. Run each agent independently against the same month and grain.
3. Require every agent to provide:
   • forecast direction: up, down, or hold
   • evidence
   • confidence
   • top 3 risks
   • recommended action
4. The arbitrator produces:
   • consensus forecast
   • dissenting views
   • final risk flag
   • whether AI forecast can replace, challenge, or only annotate the roll forecast

Output columns

• forecast_month
• grain
• customer
• product
• agent_name
• agent_forecast
• agent_confidence
• evidence
• risk_flag
• recommended_action
• arbitrated_decision

Guardrails

• If model quality is worse than roll forecast, mark the result advisory_only.
• Do not hide dissent. Dissent is the value of the council.
• Use concrete customer/product/month evidence, not generic commentary.