A Configuration-Driven Physician Compensation Engine, Built as a Vertical Slice

Challenge

This orthopedic MSO had scaled through acquisition without scaling the back office. Physician compensation ran across eight division-specific models as a manual, one-to-two-week effort each cycle; month-end close took about 26 days against a 10-day goal; and data sat siloed across the EHR, the general ledger, and HRIS, with only a partial warehouse to tie them together. The team also carried a prior failed AI-vendor experience (a rev-cycle vendor that promised 90% autonomous coding and delivered about 15%), so any new approach had to be defensible and verifiable, not another opaque automation that the finance team could not audit. The decision that had to hold up: compensation math that physicians and finance leadership could trace, line by line, back to source.

Approach

We ran structured discovery first, then built a working vertical slice rather than delivering a recommendations deck. The method kept the AI confined to extraction and language at build time; the compensation decisions themselves run as deterministic, versioned, auditable code against a configuration framework. The LLM helped read the eight legacy models and translate them into structured rules. The rules, rates, and tiers then execute as configuration, not as model output, so every dollar is reproducible and reviewable.

• Ran kickoff, infrastructure and access sessions, four compensation-model walkthroughs, and month-end-close scoping, and produced high-fidelity current-state process diagrams as the engineering source of truth.
• Designed and built an automated, configuration-driven physician-compensation engine for the flagship division, with the configuration framework structured to extend to five more divisions.
• Confined AI to reading the legacy models and drafting language; the compensation logic runs as deterministic rules and lookup tables under version control, so results are auditable and repeatable.
• Validated the engine against the client’s own SQL Server warehouse so the numbers reconciled to the systems of record.
• Stood up the secure delivery environment inside the client’s own walled network (managed vendor identity, Azure Virtual Desktop, and Azure AI Foundry as the Claude access path), so PHI never touched the consultant’s laptop.
• Surfaced operational single-point-of-failure and checklist-hygiene risks found during discovery, and translated them into the engineering record rather than leaving them as verbal notes.

Impact

• Replaced a manual, one-to-two-week-per-cycle compensation process for the flagship division with a configuration-driven engine whose outputs reconcile to the client’s warehouse.
• Built the configuration framework to extend from the flagship division to five more, so the same engine can absorb additional models through configuration rather than rework.
• Delivered the engine inside the client’s walled network with PHI kept off consultant hardware, addressing the verifiability concern left by the prior AI vendor.
• Scoped a quantified close-cycle improvement target (Day 26 toward a Day 10 goal). This is a target identified during scoping, not a realized reduction; the close acceleration work sits in the Phase 2 menu below.
• Produced a ranked Phase 2 ROI menu (close acceleration, physician scorecard, legacy-model migration, and a pro forma engine) so leadership could sequence next investments against expected return.
• Flagged operational single-point-of-failure and checklist-hygiene risks for remediation, giving leadership visibility into back-office fragility that the acquisition-driven growth had masked.

Capabilities demonstrated

• Deterministic, auditable compensation logic, with AI confined to reading legacy models and drafting language.
• Discovery that produces an engineering source of truth (high-fidelity current-state process diagrams), not just a slideware summary.
• Vertical-slice delivery: a working engine for one division, with a configuration framework built to scale.
• Secure delivery inside a client-owned walled network (managed identity, Azure Virtual Desktop, Azure AI Foundry) with PHI kept off consultant hardware.
• Validation against the client’s system of record (SQL Server warehouse) for reconcilable results.
• Ranked, ROI-sequenced roadmap planning for back-office automation.

Anonymized by design: client names stay off the narrative per our reference policy. Figures that are modeled, small-sample, or targets are identified as such above.