Reimagining Shared Services Through AI-Native Operating Models
Huron's approach to operationalize AI-enabled shared services transformation, with working demonstrations, a detailed solution walkthrough, case studies, artifacts, and perspectives on value creation that turns vision into reality.
Shared services of the future: The operating model is the unit of change.
Technology is the accelerant.
Most organizations fail to unlock the full value of advances in AI because they treat it as a technology project rather than an operational transformation. AI and technology implementations stall — or fail entirely — when they are not integrated into operations, clinical workflow, and the daily processes that drive performance. The organizations that realize the step-change are the ones that redesign how work is done, not just what tools are used to do it.
Labor and non-labor costs benchmarked against peers, spending efficiency, and value range by function.
•Spending efficiency analysis
•Value range across functions
Function-specific KPIs, end-user satisfaction, and output metrics across revenue cycle, supply chain, and operations.
•End-user satisfaction & perception
•Revenue cycle yield & output metrics
Structural readiness, organizational capabilities, and alignment across talent, operations, technology, and change readiness.
•Talent & digital enablement
•Change readiness & cultural adaptability
Every assessment now has an AI lens. The question is not whether to use AI. It is where AI creates the highest-value change with the least risk. That answer is different for every function, every workflow, and every organization.
Redesigning shared services through AI and automation
Montefiore's hypothesis is right: agentic AI, workflow automation, and selectively managed services can fundamentally redesign shared services.
Shared Services of the Future
Humans Handle Judgment.
Shared services becomes a performance platform, not a back-office cost center. Staff are exception specialists, policy stewards, and process engineers, not transaction processors. Every AI agent is governed. Every decision is auditable. Every dollar of improvement is reinvested in patient access, quality, and care.
The Pathway to the Future
Most health systems blend these three approaches, the question isn't which one, it's which is the core operating model and where the other two are used selectively in support. We recommend the third as the anchor.
Build a Montefiore-owned orchestration platform as the core. Integrate point solutions via APIs for enterprise value, not isolated tools. Use managed services and offshore selectively for overflow, subject-matter expertise, and low-risk populations. You decide what stays in-house, what gets automated, and what gets outsourced, based on your operations and enterprise goals.
Governed agentic AI on your AWS, anchored on your host systems, with the data platform and governance tools to build, measure, and scale.
This is not a technology project. This is a transformation.
Montefiore asked for a firm to support its initiative to "fundamentally reimagine how back-office and shared services functions are delivered at scale." Our answer: build a Montefiore-owned AI operating model that compounds value across all four shared services functions, Revenue Cycle, Corporate Finance, Human Resources, and Supply Chain.
Four layers, each with clear ownership, accountability, and substitution rules. Components are portable. Outcomes are measurable.
Montefiore retains control of the data, the IP, and the optionality at every stage.
Target operating model: how work is structured across AI, automation, and human roles
The operating model is the unit of change. Technology is the accelerant. The result is a fundamental shift in how work is performed at the decision level.
All shared services work organizes into three categories, each with a distinct role, defined accountability, and a clear share of transaction volume.
Your back office does not process transactions. It makes decisions. Thousands of them every day, across Revenue Cycle, Finance, HR, Supply Chain, Coding/CDI, and call centers. Most are made manually, slowly, and inconsistently, even at organizations with significant Epic and automation investment. Real cost takeout comes from classifying those decisions and moving each one to the optimal level of autonomy. The operating model shift below is the result of that decision-level redesign.
The question at Montefiore is not "where are the KPI gaps?" You have dashboards for that. The question is "which decisions are at the wrong level, why, and what is the financial cost of that misalignment?" That answer exists at the decision level, not at the KPI level. This is where enterprise-wide cost takeout is built, and it is exactly what the target operating model below is designed to deliver.
The Decision-Maker Spectrum applies to every back-office function where Montefiore has labor cost exposure. Each function holds hundreds of decisions that can be mapped, scored, and moved to the optimal autonomy level. The table below illustrates representative decisions across in-scope functions.
| Function | Representative Decision | Typical Current | Target with AI |
|---|---|---|---|
| Revenue Cycle - Denials | Which denial to appeal first and how to draft the appeal | L0 / L1 | L3 |
| Coding / CDI | Which codes apply and what CDI query to raise | L1 | L2 / L3 |
| Revenue Cycle - Prior Authorization | Which service requires PA, how to submit, and when to appeal | L1 | L3 |
| Finance | Which invoices to approve, accruals, variance explanations | L0 / L1 | L2 |
| HR | Candidate screening, scheduling, benefits question routing | L1 | L2 / L3 |
| Supply Chain | Reorder triggers, contract compliance, vendor selection | L1 | L3 |
| IT | Ticket classification, routing, Tier-1 resolution | L0 / L1 | L3 / L4 |
| Dimension | Today, 2026 | → | Target, Future State |
|---|---|---|---|
| Strategy | Cost reduction & yield optimization | → | Performance growth & defect prevention (shift-left) |
| System Role | Manual task engine, humans executing rules at scale | → | Exception Management Control Tower, humans governing autonomous systems |
| Workforce | Bulk transaction processors, volume-driven, rules-based | → | Policy stewards, exception managers, AI governors, patient navigators |
| Payer Dynamic | Adversarial, portals, PDFs, phone queues, manual re-entry | → | Collaborative, API-driven, shared-state workflows |
| Close Cycle (Finance) | Day 10, manual reconciliation, rework loops | → | Day 3, automated matching, AI-validated journal entries |
| Value | Back-office cost center, reactive, siloed | → | Enterprise performance platform, proactive, connected |
The operating model categories described above only hold if accountability is designed in from the start, not added as an afterthought. These principles are the operational expression of the Governance & Data Foundation introduced in B-0: the immutable, audit-ready layer that runs beneath every agent, every decision, and every category of work. Without these principles, Category 3 autonomy is a liability. With them, it is a competitive advantage.
The Montefiore-owned control plane, deployed on your AWS
No proprietary platform. No shadow systems. All orchestration within Montefiore's AWS, integrated with host systems via supported API patterns.
Four layers, each with clear ownership and substitution rules. Click any layer to highlight its architecture components in the diagram below.
How each layer maps to the technology stack. Select a layer above — the corresponding architecture components highlight below.
What makes agents reliable is not the model; it is the context. A general-purpose AI knows nothing about Montefiore's payer mix, escalation paths, or denial logic. We build and maintain a governed knowledge layer specific to Montefiore's operations, encoding the business rules, SOPs, and decision logic each agent needs to act correctly.
At runtime, each agent receives only the relevant slice, keeping outputs focused and grounded. The system compounds with use: every new SOP, every exception documented, every business rule clarified makes every downstream agent smarter. And when context is missing or low-confidence, agents escalate to humans rather than guessing.
How services are delivered: a Montefiore-owned operating model
Montefiore's key stakeholders receive improved shared service delivery through a model Montefiore owns and governs, not one rented from a vendor. Three distinct lanes of work. One shared platform. Improved outcomes across every function.
Every shared services transaction routes to one of three lanes based on complexity, judgment required, and risk profile. The lanes are not siloed, they share the same platform, data, and governance. What changes is who or what handles the work.
Selectively deployed to support Lanes 2 & 3 where human-directed work benefits from specialist depth, targeted expertise, or surge capacity. Not the operating model, a time-bounded tool within it. Outsourced resources work under Montefiore's governance and direction.
- ▸Denial backlog clearance sprints
- ▸Transition period AR support
- ▸Project-specific compliance work
- ▸Complex specialty coding augmentation
- ▸Category-specific sourcing expertise
- ▸Tax & regulatory filing support
- ▸Overflow volume at peak periods
- ▸Interim support during system transitions
- ▸Surge staffing & interim leadership
Regardless of which lane handles the work, every function runs on the same platform, same orchestration layer, same data fabric, same compliance and governance framework, same workflow rules. This is what makes the model scalable and owned by Montefiore.
AI handles routine work autonomously. AI-assisted work surfaces recommendations for human review. Human-led work requires judgment no agent can provide.
- ▸Eligibility & benefit verification
- ▸Claim scrubbing & submission
- ▸Standard denial categorization
- ▸ERA posting & simple resubmissions
- ▸Auth status queries via agent chat
- ▸AI-drafted denial appeal letters, human approves
- ▸AI-scored coding recommendations, coder validates
- ▸Prior auth exceptions flagged for human review
- ▸Underpayment patterns surfaced for decision
- ▸Complex clinical denial appeals & ALJ hearings
- ▸Payer contract negotiation & interpretation
- ▸CDI physician query escalations
- ▸Reorder trigger & PO generation
- ▸Invoice matching & 3-way match
- ▸Contract compliance monitoring
- ▸Delivery confirmation & receipt
- ▸Item availability queries via agent
- ▸AI-flagged spend variances, human reviews
- ▸Non-formulary requests, AI-scored options
- ▸Contract exceptions flagged for approval
- ▸AI-recommended vendors, human selects
- ▸Strategic vendor negotiation & contracting
- ▸Formulary policy & governance decisions
- ▸Capital equipment & supply strategy
- ▸Standard journal entries & accruals
- ▸Bank reconciliation & matching
- ▸Invoice processing & AP routing
- ▸Budget vs. actuals monitoring
- ▸Invoice status queries via agent chat
- ▸AI-flagged variances, human explains
- ▸Non-standard reconciliation review
- ▸AI-modeled forecasts, CFO decides
- ▸Exception invoices flagged for approval
- ▸Capital planning & strategic financial decisions
- ▸Board-level financial reporting & narrative
- ▸Audit response & regulatory filings
- ▸Benefits FAQs & enrollment queries
- ▸System access requests & provisioning
- ▸PTO balance, paycheck, policy lookups
- ▸Standard onboarding task routing
- ▸Scheduling confirmations & swaps
- ▸AI-flagged scheduling conflicts, human resolves
- ▸Accommodation requests, AI-summarizes, HR decides
- ▸Candidate screening scored by AI, manager selects
- ▸Policy exceptions flagged for HRBP review
- ▸Union contract negotiation & interpretation
- ▸Grievance resolution & formal disciplinary action
- ▸Workforce strategy & org design decisions
Every stakeholder group sees measurable improvement: faster responses, better information, more consistent outcomes.
- ▸Faster auth and scheduling confirmations
- ▸Cleaner billing with fewer surprise errors
- ▸Consistent, timely responses to inquiries
- ▸Instant answers to HR, benefits, and access questions via agent chat
- ▸No more navigating multiple portals for routine requests
- ▸Faster scheduling resolution and shift support
- ▸Real-time dashboards replacing weekly reports
- ▸Issues and trends surfaced same-day, not end-of-month
- ▸Clear accountability and auditability for every process
- ▸Prior auth and scheduling friction removed from clinical workflow
- ▸Coding and documentation support delivered in-encounter
- ▸Performance data available without requesting reports
- ▸Supply availability and reorder status surfaced proactively
- ▸More consistent, predictable service from shared functions
- ▸Fewer escalation loops for routine requests
- ▸Enterprise performance visibility in real time, not next quarter
- ▸ROI and performance measured at the decision and workflow level
- ▸Shared services as a strategic asset, not a cost center
An AI-native operating model changes not just speed but the quality of insight, governance, and foresight delivered to the organization.
Upskilling at scale: growing capacity without growing headcount
AI absorbs the volume. The workforce moves to higher-value work. Financial results follow when every individual makes that shift, not just when the technology goes live.
AI surfaces information and recommendations; humans retain full decision authority. Staff become faster, more accurate, less fatigued by routine work. Baseline metrics established.
Routine, high-confidence decisions move to governed autonomy. Human role shifts to exception management, policy stewardship, and AI governance. Role redesign begins.
Roles formally redesigned. New job families emerge. Workforce right-sized through attrition management and redeployment — not layoffs. Training pathways complete for transitioning staff.
Today's workforce skews toward routine, high-volume transaction roles. As AI absorbs that volume, the pyramid inverts. Most staff move into judgment-intensive and strategic roles. Reductions occur through attrition. Roles are redesigned before positions change.
the base
Change management is the financial multiplier
With structured change management, employees are 7x more likely to achieve transformation objectives. Without it, gains stall and the financial case disappears.
The Consumer Engagement Center centralization succeeded here using this approach, augmentation, structured retraining, and a burden-reduction narrative that required no union escalation. The same model now scales across Revenue Cycle, Finance, HR, and Supply Chain.
Path to productivity gains: how we achieve step-change improvement
Step-change gains approaching 50% come from three levers: AI-native process redesign, workforce redeployment, and a platform that compounds value as it scales. This is how Huron achieves it, and what the trajectory looks like from Year 1 to Year 3.
Value is measured across four dimensions: productivity, revenue improvement, quality, and function-level performance. When AI owns the routine, people who remain are doing work that matters: catching the denials that would otherwise slip through, closing the financial cycle faster, and governing systems that compound value over time. Better outcomes are not a byproduct. They are the point.
A directional model based on public benchmarks for a large academic health system at safety-net scale. Inputs can be calibrated to Montefiore's actual claims volume, denial rates, and operating costs in our first working session.
Your Inputs
Modeled impact, year 1
Huron's financial modeling translates decision-level autonomy change into specific labor cost impact, covering FTE reduction through non-backfill, attrition optimization, and role restructuring, with confidence ranges and validation gates at every step.
Decision inventory: current state
13 decisions mapped across the denials function. Click any decision to drill into current state, target model, blockers, and the transformation plan. This is one functional area. The same methodology applies across all ten revenue cycle functions.
Denials management: from decisions to value
The complete transformation from current state to autonomous target model, sequenced across three horizons.
Current state themes: denials function
Autonomy level distribution: current vs. target
Workforce transformation: denials function (5-year projection)
| Role | Current | Year 1 | Year 2 | Year 3 | Year 4 | Year 5 |
|---|---|---|---|---|---|---|
| Denial Specialists (HB) | 12 | 10 | 7 | 5 | 3 | 2 |
| Denial Specialists (PB) | 8 | 7 | 5 | 3 | 2 | 1 |
| Denial Nurses (Clinical) | 6 | 6 | 5 | 4 | 3 | 3 |
| Follow-up / Status Staff | 4 | 3 | 1 | 0 | 0 | 0 |
| Supervisors | 3 | 3 | 2 | 2 | 1 | 1 |
| Denial Manager | 1 | 1 | 1 | 1 | 1 | 1 |
| Denial Prevention Strategists | 0 | 1 | 2 | 3 | 3 | 3 |
| AI Exception Managers | 0 | 0 | 2 | 3 | 3 | 3 |
| Complex Case Specialists | 0 | 1 | 2 | 3 | 3 | 3 |
| Payer Intelligence Analyst | 0 | 0 | 1 | 1 | 1 | 1 |
| AI/Automation Operations | 0 | 0 | 1 | 2 | 2 | 2 |
| Denial Analytics Lead | 0 | 0 | 0 | 1 | 1 | 1 |
| Total FTE | 34 | 32 | 29 | 28 | 23 | 21 |
How we implement: a phased, controlled transformation
Montefiore-led, Huron-partnered. We prove value in a defined first release, refine and scale across similar workflows, then expand across functions, with governance gates, human-in-the-loop controls, and union engagement embedded from day one.
These principles govern every phase. They are structural requirements, not aspirations, that determine sequencing, autonomy expansion, and how we protect Montefiore throughout.
Define scope, set a baseline, demonstrate results before expanding. No broad rollout without evidence.
Augment first. Validate. Then automate. Then restructure. Autonomy is earned, not assumed.
Design and rollout overlap once scale is achieved. Parallel workstreams accelerate the timeline.
Governance approval is required before increasing AI autonomy at any stage. The gate is never waived.
Each phase has defined gates: readiness, first release go-live, autonomy expansion, scale. Advancing requires governance approval against six metrics: quality, turnaround time, exception rate, human override, staff adoption, and drift.
Union engagement begins in Phase 1 as a foundational design input, not a reaction to change. Transparency protects the transformation.
The first release scope maximizes proof of value while minimizing risk: process-driven and labor-intensive enough to demonstrate AI at scale, bounded enough to control and learn from.
One Region · Epic-Based · Defined FTE Scope
Revenue cycle is the right starting point: process-driven, labor-intensive, and complex enough to demonstrate real value at scale. Managed Medicaid follow-up combines known denial patterns, rule-based workflows, high claim volume, and work that already lives in Epic, making it ideal for a controlled autonomous first release with clear measurement.
Phases overlap deliberately. Design for Phase 2 begins before Phase 1 closes. Five workstreams run concurrently throughout.
Huron will help Montefiore design the model, build the platform, operationalize the workflows, transition the workforce, and stabilize the run-state, while keeping Montefiore in control of policy, workflow, data, and long-term capability. We are human-led, AI-enabled: we will help you redesign your operations while enabling AI to support them.
One room until scale is achieved.
A joint command center runs from first release through enterprise scale. It is the operational nerve center, not a reporting mechanism.
The orchestration layer, governance framework, agent architecture, and decision methodology built for the AR/denials first release are directly reusable across every shared service function. Each new function is a configuration, not a new contract, platform, or vendor relationship.
Same classify → route → act → learn pattern proven in the AR/denials first release walkthrough. Every sub-function runs on the same orchestration layer already built for the first release.
AI-validated journal entries and exception routing. Close cycle compresses from Day 10 to Day 3.
High-volume, rule-based decisions with strong automation potential. Employees interact with agents directly for benefits questions, access, and PTO on Workday.
Invoice matching and PO auto-approval on Infor, integrated into the shared platform. Contract compliance monitoring runs continuously.
Prior auth automation via FHIR standards. Eligibility checks upstream reduce downstream rework. Future exploration area as the platform matures.
Tier-1 resolution automation on ServiceNow. Ticket classification mirrors the deny-classify-route pattern. Future exploration once the shared services core is stable.
- Orchestration layer, routing, observability, guardrails, model registry
- Agent framework, how agents are built, deployed, and governed
- Context engineering, payer/vendor/counterparty data models, policy libraries
- Governance framework, AI Trust Model, audit trail architecture, board reporting
- Decision methodology, the inventory, autonomy ladder, financial case approach
- Change methodology, role transformation playbooks, union consultation approach
Point solution vendors build for a specific workflow. Each new workflow = new vendor, new contract, new integration. Our model, the orchestration layer, the governance, the methodology, is the investment. Each new workflow is a configuration, not a contract.
AR/Denials Walkthrough
The following solution walkthrough demonstrates the Huron solution applied to the AR Follow-Up & Denials Management workflow. This is our show-and-tell, with a live workflow demo, HIGH/MED/LOW routing logic, and detailed orchestration of AI-native models applied to real world examples.
We chose the AR/Denials workflow as our walkthrough vehicle because it carries the highest decision complexity and volume, the clearest financial value (reworking a denied claim costs $25 to $181 and many are never appealed), the most testable Epic integration surface, and the most direct reusable framework demonstration: the orchestration layer, agent architecture, and governance model shown here extend directly to coding/HIM, Finance close, HR compliance, and supply chain AP.
Six orchestrated stages: from claim status to resolution and learning
AI agents handle 70%+ of all actions autonomously. Human specialists focus on high-value exceptions. The principle in action: agents handle volume, humans handle judgment. Click any step for detail.
AI Agent, Fully Automated
- Agent polls payer portals and EDI 277 feeds every 4 hours, covering all contracted payers
- Auto-classifies status: paid, pending, denied, additional info requested, zero-pay, and retroactive eligibility changes
- Matches each response back to Epic Resolute encounter using claim ID, patient MRN, and DOS
- Flags aging claims with payer-specific escalation rules and tracks timely filing deadlines by payer
- Prioritizes denial queue by dollar value, payer type, overturn probability, and timely filing risk
Technology Stack
AI Agent, Fully Automated
- NLP engine parses CARC/RARC codes, EOB narratives, and payer correspondence, with payer-specific edit code parsing and MCO-specific denial logic
- Classifies into denial taxonomy: eligibility/enrollment lapses, medical necessity, prior authorization, coding/DRG, timely filing, and clinical documentation
- Separately flags retroactive eligibility changes for re-verification workflow, a high-volume, high-recovery pattern specific to safety-net systems
- Assigns overturn probability using models trained on Montefiore's historical payer outcomes, with payer-specific models for each contracted plan
- Routes to auto-resolve, AI-assisted, or human specialist queue based on confidence threshold and dollar value
Technology Stack
AI-Assisted, Human Review for Complex Cases
- AI agent pulls full claim context: clinical documentation, coding, authorization status, payer contract terms
- Cross-references against payer-specific denial patterns and contract provisions
- Generates root cause hypothesis with evidence chain (e.g., "Medical necessity not documented, clinical note lacks severity indicators required by Healthfirst UM Policy 2024.114")
- Simple root causes (eligibility, duplicate, timely filing) resolved autonomously
- Complex clinical denials surfaced to specialist with pre-populated analysis
Human Specialist, Escalation Triggers
- Clinical denials where AI confidence <75%
- Novel denial patterns not in training data
- High-dollar claims (>$50K) regardless of confidence
- Payer disputes requiring judgment on contract interpretation
- Cases requiring peer-to-peer clinical review
AI Agent, Autonomous Execution
- Selects optimal action from payer-specific playbook: resubmit, corrected claim, appeal, eligibility re-verification, or write-off
- For corrected claims: auto-updates coding, modifier, or authorization data in Epic Resolute
- For eligibility denials: auto-triggers re-verification through enrollment systems, resolving a significant share of eligibility denials without human touch
- For standard appeals: drafts appeal letter with clinical evidence, payer policy citations, and applicable regulatory references
- Tracks each action with full audit trail and regulatory-compliant documentation
Technology Stack
AI-Assisted, Appeal Execution
- AI drafts payer-specific appeals in the correct format for each contracted plan, each with appropriate regulatory language and evidence structure
- Human specialist reviews complex clinical appeals and high-value escalations before submission
- Agent monitors appeal status and auto-escalates through each payer's appeal levels per contractual and regulatory timelines
- For peer-to-peer reviews: prepares clinical summary with payer-specific criteria alignment using InterQual/MCG and applicable coverage policy
- Tracks all resolution outcomes and updates Epic Resolute with payment/adjustment posting
Human Specialist, High-Value Resolution
- Payer phone escalations with plan provider relations and contractor representatives
- Peer-to-peer clinical reviews with health plan medical directors
- External appeal preparation and representation for high-value cases
- High-dollar case escalation preparation for cases surviving initial appeal levels
- Pattern identification for systemic payer behavior changes, critical intelligence for contract negotiations
AI Agent, Autonomous Learning Loop
- Every resolution outcome feeds back into classification and overturn prediction models, with payer-specific model tuning for each contracted plan
- Agent identifies emerging patterns: new payer edit logic, coverage policy updates, and seasonal eligibility fluctuations that drive denial volume
- Auto-updates payer-specific playbook rules when new high-success resolution paths are validated
- Generates weekly denial trend reports by payer type, denial category, department, and provider, with payer-level breakdowns for contract negotiation intelligence
- Surfaces upstream prevention opportunities (e.g., "42% of plan denials trace to eligibility gaps at registration, recommend real-time eligibility verification at scheduling")
Technology Stack
Where AI acts. Where humans decide. What the interfaces look like.
The division of labor by step, measured, not asserted, followed by illustrative views of the agent dashboard, denial analysis, and auto-generated appeal that bring the pipeline to life.
Role distribution across the AR & denials pipeline
Task volume by role type, per workflow step. Illustrative steady-state (Year 2+).
Representative views of the platform for AR specialists, denial analysts, and managers.
| Claim ID | Patient | Payer | Amount | Denial Type | Status | Confidence | Action |
|---|---|---|---|---|---|---|---|
| CLM-2847291 | █████, R. | Healthfirst (Managed Care) | $12,340 | Medical Necessity | AI Analyzing | 87% | Appeal Draft |
| CLM-2847156 | █████, M. | Emblem Health (Commercial) | $8,420 | Auth Non-Compliance | AI Drafting | 91% | Appeal Draft |
| CLM-2846998 | █████, J. | MetroPlus (Managed Care) | $3,210 | Eligibility Lapse | Auto-Resolved | 96% | Re-verified Eligibility |
| CLM-2846877 | █████, A. | Fidelis Care (Managed Care) | $67,230 | Auth, Retro Denial | Escalated | 52% | Human Review |
| CLM-2846801 | █████, T. | UnitedHealth (Commercial) | $4,567 | Duplicate Claim | Auto-Resolved | 99% | Voided Dup |
| CLM-2846790 | █████, S. | Healthfirst (Managed Care) | $22,100 | Clinical, LOS | Escalated | 44% | Human + Legal |
The same model applied to finance: two walkthroughs, one platform
This is our second domain walkthrough, the RFS Option C framed through the same decision-level approach we showed for AR/Denials. Below: two parallel finance walkthroughs, Procure-to-Pay (Exception-Based AP) and Treasury (Rolling Cash Flow Forecast). This is the clearest demonstration of platform reusability and something that other vendors cannot produce effectively.
The walkthroughs below mirror the structure of the AR/Denials walkthrough in C-1. Five orchestrated stages, each with a clear AI-vs-human role assignment, exception routing, and a closing self-improvement loop. Click any step to expand.
Exception-based accounts payable: 80 to 90% straight-through
The gold standard for AP efficiency: automation handles 80 to 90% of clean transactions; AP staff focus exclusively on high-risk or high-value discrepancies. To make this work in a healthcare P2P environment with high-volume service billings and non-PO spend, you have to control the variables, through five orchestrated stages.
The foundation of exception-based AP is ensuring data enters the system without human touch. For health systems, this means maximizing EDI (Electronic Data Interchange) and pushing every reasonable invoice flow into a 3-way auto-match.
- Goal: Transition every eligible vendor from manual entry to EDI 810 (Invoices)
- Auto-Approve Logic: Leveraging EDI 810 (Electronic Invoices) to transition as many high-volume vendors (distributors, lab supplies) to EDI. This allows the ERP to perform a Three-Way Match (PO, Receipt, Invoice) in milliseconds.
- Coverage Target: Implement a "No PO, No Pay" policy within the system configuration (with industry exceptions). By controlling spend through Purchase Orders, you promote measured cash controls while ensuring that the accounting strings and approvals are captured upfront.
- Tech Stack: Infor ERP · EDI translator · vendor onboarding workflow · auto-match rules engine
Not every vendor can use EDI, especially local suppliers and specialized medical providers. Intelligent OCR with machine learning eliminates manual data entry for the long tail.
- AI-Driven Capture: ML-trained OCR extracts header and line-level data from PDFs, faxes, and paper invoices
- Role Shift: AP staff no longer type, they validate. The system only flags an invoice if AI confidence score is low or extracted data doesn't find a matching PO
- Confidence Routing: High confidence + PO match = auto-post. Low confidence or unmatched = exception queue (Stage 04)
- Continuous Improvement: Each human correction trains the OCR model, accuracy compounds as volume processes
In a health system, tolerance and receipt-sequencing rules are calibrated to contract net payment terms, so the right invoices auto-pay and the right ones get scrutiny.
- Configurable Tolerances: ERP auto-matches if variance is under a defined dollar amount or percentage, calibrated per vendor, category, and contract
- Header-First Logic: Match on header total first; if it ties, line-level review is bypassed entirely
- Header-Failure Path: Only when header doesn't match does the system "explode" the invoice into lines for exception handling
- Governance: All tolerance thresholds owned and signed off by Montefiore Finance, not by Huron, not by a vendor
In an exception-based world, an AP clerk's morning starts with a categorized error queue, not a stack of invoices. Each exception is routed to the role that owns the resolution.
| Exception Category | Automated Resolution Path |
|---|---|
| Unit of Measure (UOM) Mismatch | System flags "Case vs. Each" discrepancy; routes to the Buyer to update the Item Master |
| Price Variance | Automatically pings the Contract Manager if invoice price differs from the GPO contract |
| Missing Receipt | ERP sends an automated "nudge" to the loading dock or department head to confirm delivery |
True automation doesn't just flag errors, it helps prevent them. The closed loop is what turns exception-based AP from a tool into a continuously improving operation.
- Root-Cause Analysis: If the same vendor causes a Price Mismatch five times, the ERP triggers a workflow to Procurement to update the Item Master or Contract Price, preventing the next 50 exceptions
- Evaluated Receipt Settlement (ERS): For trusted, high-volume suppliers (Med-Surg distributor), the system settles payment based on receipt quantity and PO price, eliminating the invoice entirely
- Pattern Detection: Recurring exception clusters surface as systemic findings to Procurement leadership, vendor performance, contract gaps, item master quality
- Compounding Effect: Each resolved root cause permanently removes a category of future exceptions from the queue
Automated rolling cash flow forecast: active liquidity management
An automated, digitally governed rolling cash flow forecast (RCFF) is the gold standard for cash and liquidity management in a large health system. Data ingestion, transformation, and model refresh are handled systematically, Treasury focuses on variance analysis, working-capital decisions, and strategic insight. Five stages, multiple ERP and EHR sources, complex revenue cycle timing, and a cross-functional stakeholder base spanning Revenue Cycle, AP, Payroll, Supply Chain, and IT.
The foundation of an automated RCFF is ensuring source data reaches the model without manual intervention, on a defined cadence, with built-in error detection before any downstream processing begins. This eliminates the manual file-pull and upload steps that introduce lag and key-person dependency.
- Sources: EHR charges and payments · ERP general ledger balances · AP and payroll postings · bank transaction files
- Ingestion Routine: Python-based collector pulls files from a shared staging environment, validates each on arrival for completeness and structural integrity (flagging duplicates, missing fields, format anomalies), loads to SQL Server
- Cadence: Oracle GL, payroll, and AP extracts are fully scheduled. Bank files ingest as they post from the banking partner.
- Quality Gates: No data passes downstream until validation completes; missing or late files are flagged to IT before model run
Not every bank transaction arrives pre-labeled. Categorization of posted transactions into meaningful forecast line items is a critical transformation step that must be systematic and auditable.
- Rule-Based Classification: SQL scripts process each transaction against a library of business rules, combining BAI codes, transaction descriptions, and counterparty identifiers
- Forecast Categories: Patient receipts · supplemental funding · AP disbursements · payroll · intercompany sweeps · financing activity
- Role Shift: Treasury staff act as rule stewards, not transaction reviewers. The system surfaces a transaction only if no matching rule exists
- Self-Improving: Each manual review triggers a permanent rule addition, classification accuracy compounds over time
With actuals staged and classified, the model executes two parallel forecasting tracks, receipts and disbursements, each using methodology appropriate to the underlying cash flow behavior.
Patient collections forecast from 12 months of charges and payments, segmented by payor, region, and billing type. Forecasted charges netted against contractual adjustments.
Near-term AP driven by open invoices in ERP. Outer-week disbursements use rolling averages. Payroll aligned to the payment schedule with manual override for incentive pay.
The Cash Office meeting starts with a structured variance and exception queue — not with data assembly. The reporting package surfaces what matters and routes each issue to the right owner. The package includes a 13-week hybrid summary (trailing actuals + forward forecast), entity-level variance reporting, collections performance, payables analysis, and trailing 4-week line-item detail trends.
| Forecast Category | Automated Action / Resolution Path |
|---|---|
| Actuals Data Gap | Pipeline flags missing/late source file; IT notified automatically before model runs |
| Transaction Rule Miss | Unmatched bank transaction routed to Treasury for manual classification and rule update |
| Forecast-to-Actual Variance > 5% | Variance flagged in reporting package; Cash Office owner assigned root-cause analysis before next meeting |
| One-Time Item Not Reflected | System prompts functional owner (Revenue Cycle, AP, Payroll) to confirm or update prior to weekly distribution |
| Stale Manual Input | Automated check detects inputs not refreshed within window; workflow triggers reminder to responsible party |
True automation doesn't just produce outputs — it creates a feedback loop that makes the forecast more accurate and the process more resilient over time.
- Continuous Calibration: Forecast-to-actual variance analysis embedded in the weekly governance cadence. Cash Office identifies whether variances are timing-related or permanent, assigns root-cause explanations, feeds corrections into forward weeks. Persistent variances trigger structural methodology adjustments — not one-time overrides.
- Digital Target State: Platform-native architecture — Microsoft Fabric data landing zone with governed semantic model and Power BI delivery — eliminates Excel dependency, automates model refresh, makes forecast available on demand
- Cash Office Shifts: From assembling information to acting on it — variance analysis available the moment actuals post; scenario modeling is self-service; system of record is governed, auditable, scalable across entities
- Enterprise Integration: Same data architecture extends to capital planning, strategic financial planning, operational performance reporting — a unified financial intelligence layer supporting liquidity, covenant compliance, and working capital optimization from a single governed source
Both finance walkthroughs, and the AR/Denials walkthrough, resolve to the same three architectural primitives. This is what makes the platform reusable: the patterns are domain-agnostic; only the rules, data sources, and tolerances change.
AI agents match transactions across sub-ledgers, sub-systems, bank feeds, EDI streams, and payor remits in real time using fuzzy matching rules calibrated to Montefiore's Infor configuration. Variance <= tolerance auto-posts; variance > tolerance routes to reviewer queue. Same engine runs AR remits, AP 3-way match, and bank reconciliation.
Exception queue prioritized by $ impact and aging. Each exception arrives with full AI analysis: probable cause, suggested resolution, supporting evidence. Human confirms, overrides, or escalates. Every action logged for learning. Same queue framework handles denial exceptions, AP variances, and forecast variances.
Standard recurring journal entries validated & posted autonomously. Accrual calculations checked against prior periods & variance thresholds. Close dashboard tracks outstanding items real-time. Audit-ready documentation generated automatically. The same self-learning loop that retires denial root causes retires AP exception categories and forecast classification gaps.
The same orchestration layer that runs denials runs the close, runs exception-based AP, and runs the rolling cash flow forecast. The same agent framework, the same governance, the same monitoring, the same decision-level methodology. This is what "platform, not point solutions" means in practice, and it's why corporate finance transformation compounds rather than starts over with each new use case.
Huron in action: three representative engagements
Healthcare automation at a top pediatric system, enterprise operating model transformation at a leading academic medical institution, and global finance transformation at a hypergrowth technology company. Three contexts. One consistent methodology.
Working demonstrations & deployed examples
Artifacts from live engagements and internal development will be shared during our interactive visioning session. We look forward to demonstrating the AI-native operating model at work.
Denial Management AI Agent
Working prototype: denial classification, root cause analysis, and auto-generated appeal letters with payer-specific citations.
Enterprise Labor Planning Agent
Deployed for a major entertainment enterprise. Agentic workforce scheduling and demand modeling.
Insights Agent
Deployed for a major entertainment enterprise. Agentic finance and IT insights.
The Huron difference: what makes us the right choice
The solution comes first. We come second. But the solution only works if Huron can actually deliver it in a unionized academic medical center, across all shared services functions, on your tech stack, while navigating intense growth and financial pressure. Here is why that firm is Huron.
Functional breadth: every in-scope domain
Huron unites operational, clinical, digital, AI, and cultural expertise in one team. Our 3,800+ healthcare professionals, including 100+ clinicians with firsthand leadership experience, understand how decisions in one area ripple across clinical care, operations, technology, and financial performance. Domain experts, not generalists.
Digital expertise from the inside
Not one system. Your entire tech stack. 3,200+ digital consultants across EHR, ERP, analytics, and automation. Huron delivers strategy through execution, without rip-and-replace disruption. Specifically: 250+ Epic resources averaging 10+ years and 4 certifications each. 34% are former Epic employees. 500+ Epic projects across 100+ clients.
Operationalized change in unionized environments
Pure-tech competitors do not have the change management muscle. Huron has navigated role transformation in unionized, regulated healthcare repeatedly, including a deep New York client base such as Northwell and NYCH+H and our revenue cycle and call center work at Montefiore.
Strategy + implementation + operation, as one team
Other firms can write a strategy. An outsourcing firm can operate a back office. Huron does both, in one team, with one accountable leader. That means no handoff risk, no reinterpretation of strategy at implementation, and no "not our problem" when something needs to be adjusted in operation.
AI-forward across 190+ engagements
This isn't a new practice area. We have delivered 190+ AI and automation engagements for 70+ healthcare clients. We have shipped L3 to L4 autonomous workflows in live production. We know what works, what doesn't, and what's ready today vs. roadmap.
Functional breadth across every in-scope function + platform depth across every system you run + change management in a unionized environment. Point solution vendors have none of these. Outsourcing firms don't do AI-native or change. Other firms don't implement or operate. AI-native companies can't run a Rev Cycle inside an academic medical center. The combination is what it takes to actually deliver what Montefiore is asking for, and that combination is Huron.
“The organizations that get there first will not be those that deployed the most technology. They will be the organizations that redesigned their operating models and fundamentally changed the way work is performed at an individual level — and built governance infrastructure that allowed each successive deployment to compound intelligence for the next.”
Who we are, the Huron you're working with
Before diving into how Huron would work with Montefiore on shared services transformation, it helps to see the full picture: who we are, what we do in healthcare, and why our capabilities are a natural fit, not a bolted-on addition.
Huron Services & Capabilities
Who we are, what we do in healthcare, and how our AI capabilities fit.
Huron is a preeminent global professional services firm built on deep industry focus and full-service execution. Our healthcare practice is the largest single expression of that model, and our AI capabilities are a natural extension of it, not a separate offering bolted on from the outside.
About Huron, a global professional services firm
Formed in 2002, Huron is a preeminent global professional services firm and technology partner with deep industry focus and growing digital and managed services capabilities. We have served clients in more than 80 countries worldwide, and we are publicly traded on NASDAQ under the ticker HURN.
(before reimbursables)
worldwide
Huron Healthcare Services. Where we transform ideas into action
One integrated team to drive strategic direction, operational performance, and digital transformation, with deep healthcare operating model expertise. The healthcare practice is the largest single expression of the Huron model: full-service, measurable, and built around long-term partnership.
- Strategy, Innovation & Consumer Experience
- Financial Advisory Services
- Care Transformation
- Revenue Cycle & Clinical Documentation
- Operations & Cost Optimization
- Human Capital Management
- Physician Enterprise & Access to Care
- Research Enterprise
- Advisory and Innovation
- Data Management and Governance
- Advanced Analytics
- Intelligent Automation
- Enterprise Resource Planning (ERP)
- Electronic Health Record (EHR)
- Customer Relationship Manager (CRM)
- Revenue Cycle
- Digital
- Interim Leadership
Representative Healthcare Clients Confidential
800+ Healthcare Organizations ServedHuron has worked with more than 800 health systems, hospitals, and physician organizations through transformational change. The following is a representative list of Huron healthcare clients, including Montefiore Einstein Health System.
- Adventist Health
- Alameda Health System
- Allina Health
- Altru Health System
- Ann & Robert H. Lurie Children's Hospital of Chicago
- Archbold Medical Center
- Aria - Jefferson Health
- Atlantic Health System
- Avera Health
- Baptist Memorial Health Care Corporation
- BayCare
- Baylor Scott & White Health
- Beth Israel Deaconess Medical Center
- BJC HealthCare
- Care New England
- Carilion Clinic
- Children's Nebraska
- Children's Hospital of Los Angeles
- Christ Community Health Services
- Christus Health
- Citizens Memorial Hospital
- Cleveland Clinic Health System
- Cook Children's Health Care System
- Cooper University Hospital
- CoxHealth
- Crouse Hospital Health System
- DaVita Inc.
- Dayton Children's Hospital
- DHR Health
- Driscoll Children's Hospital
- Duke University Health System
- Faith Regional Health Services
- Floyd Medical Center
- Froedtert & The Medical College of Wisconsin
- Froedtert ThedaCare Health, Inc.
- Good Samaritan Hospital Los Angeles
- Griffin Hospital
- Hartford HealthCare
- HealthPartners
- Henry Ford Health
- Heywood Healthcare
- Holyoke Medical Center
- Hospital for Special Surgery
- Hunterdon Medical Center
- Huntington Health
- Inova Health System
- Insight Hospital and Medical Center Chicago
- Inspira Health Network
- INTEGRIS Health, Inc.
- Intermountain Healthcare
- Jamestown Regional Medical Center
- Jefferson Health
- Kaleida Health
- Kettering Health
- Lake Health
- Loma Linda University Health
- Louisiana Children's Medical Center Health
- Marin Health
- Mary Washington Healthcare
- Mass General Brigham
- McLaren Port Huron Hospital
- Medical University of South Carolina Health System
- Memorial Health
- Memorial Health at Gulfport
- Memorial Hermann Health System
- Mercy Hospital Southeast
- Methodist Dallas Medical Center
- Methodist Healthcare
- Montefiore Einstein Health System
- Mount Sinai Health System
- Munson Healthcare
- MUSC Health - Orangeburg
- Nathan Littauer Hospital
- NewYork-Presbyterian
- Northwell Health
- Northwest Community Hospital
- Northwestern Medicine
- Norton Healthcare
- NYC Health and Hospitals
- Ochsner Health System
- OhioHealth
- Oregon Health and Science University
- Orlando Health
- Palomar Health
- Parkview Health
- Regional One Health / Regional One Medical Center
- Roper St. Francis Health System
- Rush University System for Health
- Saint Tammany Parish Hospital
- San Antonio Regional Hospital
- Shenandoah Medical Center
- Silver Cross Hospital
- South Shore Hospital
- Southwell
- SSM Health
- Summa Health System
- Summit Healthcare
- Sutter Health
- TJ Regional Health
- TriHealth
- Trinity Health
- Tufts Medicine
- UNC Health
- University of California Davis Health
- University of California Health
- University of California Los Angeles Health
- University of California San Diego Health
- University of Connecticut Health Center
- University of Kentucky Health System
- University of Missouri Health System
- University of New Mexico Hospital
- University of Pittsburgh Medical Center
- UNM Health System
- UVA Health System
- UW Medicine
- WellStar Health System
- Wickenburg Community Hospital
- Woman's Hospital
Meet the Huron team
Hand-selected to bring healthcare operations expertise, AI leadership, and direct Montefiore experience. Purpose-built for this engagement from day one.
Context management for Montefiore shared services AI & automation
How we generate, structure, and govern the knowledge layer that makes agentic workflows reliable.
The Core Problem
AI agents are only as good as the context they operate within. A general-purpose LLM knows nothing about Montefiore's specific workflows, business rules, escalation paths, or operational constraints. Without structured context, agents produce generic outputs that require heavy human correction — defeating the purpose of automation.
Context engineering is the discipline of assembling, structuring, and delivering the right knowledge to an agent at the right time, in a form the agent can reliably act on. It is the layer between raw enterprise knowledge and effective agent action — encoding not generic best practices, but the specific rules, exceptions, and decision logic Montefiore's workflows depend on.
The design principle: knowledge that compounds
Our approach is built on the LLM Wiki pattern — an architecture where AI maintains the knowledge layer while humans curate sources and validate quality. The core insight: RAG retrieves and forgets; a wiki accumulates and compounds. Traditional retrieval systems find documents on each query but never build persistent understanding. A maintained context layer gets richer with every source ingested, every workflow deployed, and every exception encountered.
Three layers, distinct ownership: Raw sources (human-owned, immutable) — SOPs, system configs, interview transcripts, operational data. Context layer (AI-maintained, human-reviewed) — structured knowledge objects synthesized from sources. Schema (human + AI co-evolved) — the configuration that makes the AI a disciplined knowledge worker rather than a generic chatbot.
The context layer compounds with use. Every new SOP ingested, every edge case documented, every business rule clarified makes the system smarter for all downstream agents. The knowledge doesn't walk out the door when people leave — it's durable, structured, and continuously governed.
Three-layer context architecture
At runtime, agents receive a composed context from all three layers. The institutional layer provides analytical scaffolding; the operational layer grounds it in Montefiore's reality; the workflow layer defines the agent's specific mission and constraints.
The context generation pipeline
New sources → update objects
Agents consume context
Detect drift & decay
Ingest is where the system grows. A single new source — a revised SOP, a system change notice, an operational exception — may update 10–15 context objects in one pass. The AI reads the source, determines what's new or changed, and updates every affected object while maintaining cross-references. Humans review changes before they go live.
Query is where agents consume context at runtime. Rather than loading everything, the system selectively composes only the most relevant context objects for each agent's specific task — keeping agents focused and outputs grounded.
Lint is where quality is maintained. Automated health checks surface contradictions, stale claims, orphaned references, and objects that have drifted from current operations — making the system self-sustaining rather than slowly decaying.
Each context object addresses one of five knowledge types: definitional (what things mean), procedural (how to do things), interpretive (how to judge results), presentation (how to communicate), and decision (what to do next given a pattern or constraint).
Key principle: Context is a finite resource. We selectively load the most relevant, most current, highest-confidence content for each specific agent task — keeping agents focused and outputs reliable.
Ongoing management framework
Staleness detection, broken reference checks, and freshness scoring flag context objects that may be outdated before agents rely on stale information.
Contradiction detection, consistency scoring, and usage tracking identify which context is actually being consumed and which has drifted from operational reality.
Portfolio-level review surfaces archival candidates, identifies gaps where agents lack adequate context, and promotes proven local patterns into shared institutional knowledge.
Every context object carries six quality dimensions: lift (does it improve agent outputs?), freshness, usage, consistency, provenance (can we trace claims to sources?), and maintainability. These scores drive automated review queues — surfacing problems to humans rather than waiting for manual audits to discover drift.
How this supports agentic workflows
Each automated workflow consumes context through a standard service interface. Agents read specific context objects by type, search for relevant business rules, and receive updates when the context they depend on changes. This means: