I am working on a poc where the agent should respond to a natural prompt like: "Give me the forecast for the next 3 months of orders."
And the expected output would look something like this (generic example):
Forecast Summary (Next 3 Months)
Month 1 → $XXM
Product A ($XM)
Product B ($XM)
Product C ($XM)
Month 2 → $XXM
Product D ($XM)
Product E ($XM)
Product F ($XM)
Month 3 → $XXM
Product G ($XM)
Product H ($XM)
Product I ($XM)
Insights
Revenue trend across months
Top contributing products
Inventory planning recommendations
I’ve already set up
OrderItem ,Order and Product data streams in Data Cloud, created relationships.
👉 Has anyone implemented this kind of
forecasting pipeline? Any guidance, best practices, or Salesforce documentation links would be very helpful.
,
Here are my few cents:
- Don't try to make the agent "calculate" the forecast in real time. Build a Calculated Insight (or a data transform feeding a DMO) in Data Cloud that pre-computes monthly forecast numbers by product — trend/moving-average logic.
- If there is a dire need, better to create a predictive model in Data Cloud (see this for how part:
https://www.salesforceben.com/how-to-approach-predictive-ai-in-salesforce-data-cloud-key-steps-and-considerations/).
Reasons for above:
-
Token limits — 3 months of OrderItem-level data across products will blow past what you can stuff into a single prompt/action response, especially at scale.- Reliability — LLMs are decent at spotting a rough directional trend from a summarized table, but bad at actual numeric extrapolation. Ask it to "predict Q on the fly" and you'll get plausible-sounding numbers, not accurate ones — risky for anything decision-driving.
- Consistency
— same prompt twice can give you two different forecasts. Not great when a sales/ops person questions the number.
Best middle approach: pre-aggregate the data via a Data Cloud CI/transform, then hand
that summarized table to the LLM and ask it to interpret trends.
