Engineering an agentic finance operations platform.
From fragmented workflows to an intelligent operational platform that reduced manual administrative effort by at least fourfold.
- Industry
- Medical market research
- Client
- Medpanel
- Services
- Product engineering · Agentic AI · Workflow automation · Enterprise integration
The manual-effort reduction is based on Medpanel’s reported operational experience. The platform has been in production use at Medpanel.
Financial operations had become a coordination problem
For a medical market research company managing hundreds of active projects, financial operations extend well beyond traditional accounting. Research participants must be paid. Vendor invoices must be collected, validated, and allocated. Client invoices must be raised against project milestones. Revenue must be recognized and reconciled.
Each process involves different people, supporting documents, approval steps, and external systems. The administrative burden lies not only in performing these tasks, but in coordinating everything around them: finding missing information, chasing documents, tracking approvals, following up with colleagues, and determining what needs attention next.
The initial requirement was a focused system for tracking research participant honoraria. The broader opportunity became clear: build a unified operational platform that could bring these financial processes together, automate routine coordination, and eventually make application capabilities accessible through AI.
Four problems had to be solved together
Six domains, different rules
Six interconnected financial domains, each with its own rules and workflows. Participant payments, client invoicing, vendor bills and revenue recognition follow different approval paths and had to share one data model.
Coordination with accountability
Documents, approvals, tasks and external systems had to be coordinated without losing accountability. Every handoff needed a clear owner and a recorded history, including steps that depend on outside systems.
AI access with controls intact
AI needed access to business capabilities without bypassing authorization, validation or confirmation. Anything it could propose had to stay within what the application already permits for that user.
A platform that can change
The platform had to stay maintainable as operational requirements change. New processes and integrations needed to be added without rewriting what already worked.
Three layers, each with a distinct role
Founding Minds engineered a business application, a layer of deterministic automation, and an agentic orchestration layer above them.
- Layer 1Business applicationRecords, work inbox, business domains, reporting and integrations.
- Layer 2Deterministic automationScheduled bill intake, document extraction, duplicate checks, workflow routing, notifications and transaction preparation.
- Layer 3Agentic orchestrationNatural-language intent, selection of registered application capabilities, schema-constrained plans and governed actions.
Scheduled automations run as application services and are not agent-initiated.
One operational platform
The business application is a unified finance operations platform spanning six interconnected business domains.
- Research participant payments
- Client invoicing
- Vendor bill processing
- Revenue recognition
- Master data and project management
- Operational reporting and controls
The platform replaces fragmented task handling with a single operational experience. At its center is a unified work inbox. Instead of navigating separate modules to discover what needs attention, users see a consolidated, role-specific view of pending work across the business.
Behind the interface, the same platform services hold the records, enforce role-based access and workflow rules, and connect to external systems.
Work progresses without constant human intervention
The platform’s value is not limited to conversational interaction. A significant part comes from proactively preparing work, automating predictable steps, and reducing the need for people to coordinate routine activities manually. These are scheduled, deterministic automations. They do not depend on the AI assistant.
- AutomatedVendor bill receivedFrom a shared Microsoft 365 mailbox
- AutomatedInformation extractedRelevant fields read from the bill
- AutomatedDuplicate validationChecked against existing records
- AutomatedReview task createdRouted into the review workflow
- PeopleApprovalDecision by the responsible user
- PeopleBookkeepingFinal stage of the workflow
Vendor bill processing. Scheduled automations run as application services, separate from the AI planner.
Vendor bill processing. The platform collects vendor bills from a shared Microsoft 365 mailbox, extracts relevant information, checks for duplicates, and routes records into the appropriate review workflow.
Payment coordination. Research participant payments move through defined request, approval, finance, and bookkeeping stages with automated notifications and a complete transition history.
Client invoicing. Invoice requests are linked to projects and billing milestones, with controlled approvals, invoice-document handling, and financial reconciliation.
Revenue recognition. Quarterly recognition periods are prepared from billing records and financial data, supporting review, reconciliation, and journal-entry preparation.
Unified task management. Regardless of which process produces the work, the next required action appears in the responsible user’s inbox.
The system handles much of the preparation and coordination, leaving people to address approvals, exceptions, and decisions that require judgment.
- Vendor
- Harbor Lane Recruiting LLC
- Bill #
- HL-2231
- Bill date
- 10/05/2026
- Due date
- 11/04/2026
- Project ref
- KLM-2609-02
- Amount
- USD 8,400.00
- Code
- KLM-2609-02
- Name
- CARDIOLOGY PANEL WAVE 2
- Status
- ACTIVE
DATE 10/05/2026
DUE 11/04/2026
PROJECT KLM-2609-02
- ✓Bill received from shared mailboxOct 6, 2026, 9:12 AM
- ✓Fields read from the PDFOct 6, 2026, 9:12 AM
- ✓Duplicate check: no matching billOct 6, 2026, 9:13 AM
- ✓Review task created for FinanceOct 6, 2026, 9:13 AM
The AI reasons. The application governs. People remain in control.
The engineering challenge was not connecting a language model to an application. It was making AI capable of working with meaningful business functions without bypassing the controls that financial operations require.
Founding Minds built an AI orchestration layer around a registry of 35 explicitly defined application capabilities, covering payments, invoice processing, vendor bills, company and project records, search, navigation, and reporting. When a user expresses an intention, the AI planner identifies appropriate capabilities and produces a structured application plan. The platform validates that plan, applies access controls, and presents proposed changes for human confirmation before executing the relevant business services.
The AI proposes plans. Changes to records are made by the application’s own services, after authorization and confirmation.
Explicit capabilities.
The AI works from a registry of 35 application capabilities. Its output follows a defined schema that limits it to supported operations.
Governed execution.
The application applies role-based access and existing workflow rules to each plan. Changes to records are previewed for confirmation and carried out by the application’s own services.
Operational observability.
The platform records AI usage, latency, token consumption and outcomes, with configurable usage limits.
Rather than giving a language model unrestricted control, the platform exposes a deliberate set of safe, reusable capabilities.
The AI cockpit: the user states an intention, the planner selects a registered capability, and the change waits for confirmation. Illustrative reconstruction with fictitious data.
A platform designed for complexity, integration, and growth
Founding Minds built the application using a modular architecture that separates business domains while sharing essential platform services.
Core capabilities include workflow orchestration, role-based access, audit trails, notifications, financial reporting and import tools. Integrations connect the platform with QuickBooks Online, Bill.com and Microsoft 365.
The agentic layer works above these established application services rather than replacing them. Business capabilities remain available through conventional user interfaces and automation, while AI provides another way to discover, organize, and initiate supported operations. The system can therefore evolve without making core financial processing dependent on unpredictable AI-generated logic.
Designed for operational accountability
- Explicit workflow states and approved transitions
- Role-specific permissions
- Human approval for consequential actions
- Auditable changes and workflow histories
- Duplicate detection and validation
The result is a maintainable business application with an AI interaction layer, not a standalone AI demonstration.
What changed
For document follow-up and related administrative coordination, as reported by Medpanel.
Medpanel uses the platform in its financial operations. Through the integrated platform and its automated workflows (unified task management, integrated data and scheduled automation), manual administrative effort for document follow-up and related coordination has fallen by at least a factor of four, as reported by Medpanel.
Work that previously required repeated follow-ups and manual tracking can now be prepared, surfaced, routed, and progressed through integrated workflows. Employees spend less time chasing documents, determining task ownership, and coordinating routine steps.
The agentic layer was introduced after these workflows were in place, as a separate architectural advancement. It gives users another way to find and initiate supported operations through governed actions. The reported fourfold result is not attributed to the agentic layer. What began as a focused participant-payment application is now a production finance operations platform spanning six business domains and integrated external systems.
This is a reported result for manual document follow-up and related coordination, not a measurement of every financial process.
The engineering decision that mattered
Founding Minds designed the platform so AI could work through the same trusted application services used by conventional workflows. The AI interprets intent and proposes actions, while established business rules, authorization, and human approvals determine what gets executed. That separation made agentic functionality a controlled extension of a production business platform, rather than an independent layer of automation.
Building an AI-first application?
We help companies turn complex business processes into intelligent, production-ready software.