Infrastructure for Write-Off & Discount Prediction

Write-off and discount prediction requires documented billing strategy policies — which client types are expected to negotiate, what realization rate is acceptable by service line, and when a write-off requires partner approval. At L3, these policies must be findable and current because the ML model uses them to define what 'normal' realization looks like versus a risk signal. Audit requirements for write-off authorization ensure that approval thresholds and write-off rationale are documented, providing training signal for the prediction model.

Write-off prediction requires automated capture of time entry descriptions, billing codes, pre-bill review decisions, client negotiation outcomes, and final write-off or discount amounts across the billing workflow. At L4, PSA systems automatically log each billing decision — time entries submitted, pre-bill reviewed, amounts adjusted, invoice issued, negotiation recorded — creating a complete audit trail that the ML model uses as training data. Automated capture from the billing workflow is essential because manual logging of negotiation outcomes is too inconsistent to train reliable prediction models.

Write-off prediction requires a consistent schema linking time entries to billing codes, consultant roles, client identifiers, project types, and historical write-off outcomes. At L3, the PSA financial data model provides these fields consistently across engagements, enabling the ML model to compute realization rate features — write-off rate by client, by billing code, by consultant, by service line — that predict future write-off probability. This structured feature set is what distinguishes a predictive model from a simple aging report.

Write-off prediction requires the ML model to query current unbilled time entries, client negotiation history, project performance metrics, and historical realization rates from PSA and ERP systems. At L3, API access enables automated daily scoring of unbilled time, surfacing high write-off risk entries to pre-bill reviewers before invoices are generated. This proactive identification — before revenue is recognized and invoiced — is the core value of the prediction capability.

Write-off prediction models require retraining as client payment and negotiation behaviors evolve, but realization patterns change gradually rather than in discrete events. At L2, scheduled periodic model retraining — aligned with quarterly business reviews that assess realization performance — is sufficient to keep predictions calibrated. The model benefits from accumulating new write-off outcomes over time through the billing cycle rather than requiring event-triggered retraining.

Write-off prediction primarily integrates PSA (time entries, billing codes, historical write-off records) with the pre-bill review workflow. At L2, point-to-point integration between PSA and the billing review interface enables risk scores to appear alongside time entries during pre-bill review — giving reviewers AI-enhanced prioritization without requiring a full integration platform. CRM integration for client relationship context would enrich predictions but isn't available at this baseline.