Australian energy retailers operating in the National Electricity Market make forecasting calls every day that move real money. Hedge sizing, retail pricing, and churn-aware retention all rest on demand and price forecasts. When those forecasts are off by a few percentage points, the cost runs into the millions. Claude-assisted forecasting in 2026 is mature enough to sit in this stack as a default tool, provided it is applied with the right governance.
For a tier-2 Australian retailer with 250,000 customers, the gap between competent and incompetent forecasting is on the order of $8M to $20M AUD a year in hedge cost, retention cost, and unrecovered margin. That figure is why this belongs in a board conversation, not a data-science side project. The retailers that treat forecasting as core infrastructure tend to outprice and outlast the ones that treat it as a quarterly spreadsheet exercise.
Demand forecasting
Demand forecasting for an Australian retailer combines the aggregate load forecasts AEMO publishes with the retailer's own customer mix and weather-sensitive load. The part that used to cost an analyst a week of spreadsheet assembly is the part Claude handles well, freeing the analyst to interrogate the result rather than build it.
Where a well-governed model earns its place:
Customer-segmented demand forecasts the retailer can act on, not just an aggregate number
Weather-conditioned load forecasts built on BOM forecast data
Solar and storage export forecasts for the retailer's rooftop-solar base
Demand-response opportunity identification for commercial and industrial accounts
Aggregate accuracy is not enough. Hedge decisions and tariff design happen at segment level, so the segmentation is what turns a forecast into an action. A forecast that is right on average but wrong on the segments that drive the hedge book is worse than useless, because it reads as precise.
Wholesale price forecasting
Wholesale prices in the NEM are volatile and genuinely hard to predict. No model removes that volatility. What a Claude-assisted approach adds is faster identification of regime shifts and structural breaks than a hand-maintained model can manage, and a clearer audit trail of why the forecast moved.
Useful applications include:
Short-term price-path forecasting over 24 to 72 hours with explicit uncertainty intervals
Medium-term regime classification across a one-to-six-month horizon
Counterfactual analysis for what-if scenarios on hedge-book composition
Stress testing of the hedge book against adverse price scenarios
These feed the hedge committee's weekly decisions. The committee still owns the call; the model shortens the path from raw data to a defensible recommendation, and keeps the assumptions behind each number visible.
Retail pricing
Retail tariff design is the slowest-moving and highest-impact part of the stack. Australian retailers can refresh standing offers only a limited number of times a year, and each refresh shifts both margin and churn. Getting one annual reprice wrong can cost more than a full year of model operating spend.
Pricing analytics support:
Customer-segment elasticity estimates drawn from historical pricing experiments
Competitor tariff monitoring and positioning analysis
Customer-level churn-risk forecasting given a proposed tariff change
Margin-impact modelling across proposed tariff scenarios
The pricing committee makes the decision. Claude provides faster and more granular inputs so the committee spends its time on judgment rather than data assembly.
AEMC and AER compliance
Australian energy retail is regulated by the AEMC and the AER, and any forecasting or pricing system has to respect the rules covering hardship customers, market participation, and customer protections. A model that ignores them creates regulatory exposure, not advantage.
The compliance checklist that has to sit around the model:
Hardship-customer identification and treatment under the National Energy Customer Framework
Best-offer messaging obligations under the AER Better Bills Guideline
Privacy Act compliance for any customer-level forecasting or modelling
Algorithmic transparency, so a pricing decision can be explained if the regulator asks
Claude's auditability helps here. Because the inputs and the reasoning behind a recommendation can be captured and reviewed, the retailer keeps a defensible record of how a number was reached, which is exactly what an AER query or an internal audit will ask for.
What stays a human decision
Forecasting with Claude is an input layer, not an autopilot. The retailer's people stay accountable for every consequential decision, and the operating model should make that explicit from day one.
Hedge execution stays with the trading desk, which signs off on every position
Final standing-offer pricing stays with the pricing committee under delegated authority
Hardship and customer-protection calls stay with the customer team and are never automated
Model assumptions get reviewed each quarter against actuals so drift is caught early
Drawing that line early is what makes the build defensible to an internal audit team and to the regulator. The retailers that succeed treat the model as a faster analyst, not a replacement for the committee.
Cost and timeline
A working AI forecasting stack for a tier-2 Australian energy retailer typically costs $400,000 to $1.2M AUD to build and $150,000 to $400,000 a year to run. Build time runs 16 to 28 weeks, and payback usually lands inside the first hedge cycle.
The sequencing that works: start with demand forecasting where the data is cleanest, prove the segment-level lift, then extend into price forecasting and pricing analytics once the governance and the hedge committee's trust are in place.
If your retail business is sizing a forecasting build, you can book an energy pilot scoping with our team.
