Professional Services forecasting
How do you build a useful forecast when trustworthy historical snapshots don’t exist?
- Pipeline snapshot
- Historical rates
- Calculated baseline
- Context & human review
- Published forecast
A change in approach
The initial plan was to train a model on historical weekly pipeline snapshots. Available stage history could not reliably reconstruct those snapshots, so I shifted toward deterministic forecasting using historical category close rates, hierarchical fallback, and smoothing.
The application
The project evolved toward a Salesforce-native application using Lightning Web Components and Apex. The workbench separates the system-calculated forecast from the human-editable forecast. Weekly snapshots preserve what was known when a forecast was made.
Where the LLM belongs
Language models help interpret qualitative sales context, organize evidence, and surface risks. The numerical baseline belongs to the deterministic calculation layer; the final published forecast belongs to the human reviewer.
State at handoff
Substantial backend work and an older wired frontend existed. The preferred new interface still required complete backend wiring, verification of the weekly lifecycle and formulas, and production promotion. The handoff records those gaps rather than treating the intended architecture as fully shipped.