AI-Native FinanceOps Control Plane for Small Business

Inspiration

Small businesses use separate tools for accounting, banking, payments, payroll, CRM, and documents. Their finance teams spend too much time collecting data and checking spreadsheets before making decisions.

We built FinanceOS to explore a simple idea: instead of adding a chatbot to finance software, what if a company could operate with a coordinated AI-native finance department?

What it does

FinanceOS is a finance operations control plane for small businesses.

A company can:

Create separate workspaces for each business.
Connect financial accounts through secure provider authorization.
Upload transaction CSV files.
Monitor transactions, close readiness, and material exposure.
Run a financial-close review.
Investigate duplicate invoices, missing evidence, unmatched bank transactions, vendor changes, and intercompany differences.
Review agent activity and recommendations.
Approve, reject, or defer material decisions.
Ask the FinanceOS Close Manager questions about the active workspace.

FinanceOS uses five specialist agents:

Close Manager coordinates the workflow.
Data Assurance Analyst checks source completeness and freshness.
Reconciliation Analyst reviews financial records.
Audit Reviewer investigates evidence and policies.
Case Controller maintains decisions and audit history.

How we built it

We built FinanceOS with:

Next.js, React, and TypeScript for the application.
Google Agent Development Kit for TypeScript for agent coordination.
Gemini 2.5 Flash for the FinanceOS assistant and agent reasoning.
MongoDB Atlas integration for finance memory and future evidence search.
Secure account authorization for live financial connections.
Server-Sent Events for live dashboard updates.
Vercel for deployment.

We also created deterministic finance checks for calculations that should not depend only on an AI model. These checks detect duplicate invoices, missing evidence, unmatched transactions, and other financial risks.

Challenges we ran into

One major challenge was designing agents that perform useful financial work instead of only producing text.

We addressed this by giving each agent a specific responsibility, tools, and limits. Agent actions are displayed in the workbench so users can understand what happened.

Another challenge was account connectivity. Financial credentials must remain secure, so FinanceOS uses protected provider authorization instead of collecting passwords directly.

We also needed to make failures honest. FinanceOS does not show an account as connected until the connection and sync status are verified.

Finally, deployment required different configurations for local containers and Vercel's Next.js runtime.

Accomplishments that we're proud of

Built a working multi-agent finance workflow.
Created isolated workspaces for multiple companies.
Implemented secure financial-account connection flows.
Added transaction CSV uploads and normalized data views.
Built deterministic financial-risk checks.
Added controller approval gates for material decisions.
Created a live agent activity workbench.
Built a complete six-stage close workflow.
Passed all automated tests and production builds.
Deployed the application on Vercel.

What we learned

We learned that an AI-native finance product needs more than a language model.

It needs reliable data, clear workflows, deterministic controls, secure account access, and understandable decision history.

We also learned that full AI-native operation does not mean hiding what the system is doing. Users need to see each agent's responsibility, current task, evidence, and result.

What's next for AI-Native FinanceOps Control Plane for Small Business

Our long-term goal is a fully AI-native finance department for small businesses.

The next version will include:

Persistent multi-company workspaces in MongoDB Atlas.
More accounting, banking, payroll, and payment connections.
Continuous transaction monitoring.
Automated reconciliation and close preparation.
Custom company policies and approval rules.
Evidence retrieval using vector search.
Cash-flow forecasting and treasury workflows.
Tax, compliance, payroll, and reporting agents.
Role-based access and complete audit history.

The system will continuously operate the finance workflow, detect problems, prepare actions, and explain its decisions. Human review will remain available for businesses that require it, but the product is being designed for increasingly autonomous finance operations.