WorkpaperOS: AttestWork + WPOSBooks

Gemini-powered AI agents for audit and accounting: AttestWork streamlines engagement workflows, while WPOSBooks categorizes transactions, reconciles accounts, and keeps professionals in control.

Comment

Inspiration

What it do## Inspiration

Audit and accounting professionals spend substantial time moving information between disconnected systems, preparing repetitive documentation, categorizing transactions, reconciling accounts, and ensuring that every conclusion remains supported by evidence. Smaller firms often lack affordable tools that combine professional workflows with governed AI.

WorkpaperOS was created to make AI useful inside the actual workflow—not as a separate chatbot. This submission focuses exclusively on two applications: AttestWork and WPOSBooks.

What it does

AttestWork

AttestWork supports structured audit and assurance engagements. It brings engagement setup, planning, procedures, evidence, files, results, cross-references, review, and sign-off into one workspace.

Through the WorkpaperOS Model Context Protocol (MCP) server, a Gemini-powered agent can securely access authorized engagements and perform permitted actions. The demonstration shows the agent working with an audit engagement, connected engagement storage, planning procedures, engagement-letter documentation, and supporting files.

Professional judgment remains with the accountant or auditor. AI-generated work is subject to human review, authorization, and sign-off.

WPOSBooks

WPOSBooks provides AI-assisted bookkeeping and accounting workflows. The application helps users:

  • Review connected bank records
  • Analyze and categorize transactions
  • Prepare accounting entries
  • Identify reconciliation differences
  • Complete and document bank reconciliations
  • Maintain professional review and sign-off

The objective is to reduce repetitive accounting work while keeping the underlying records, adjustments, explanations, and approvals visible to the professional.

How we built it

The applications use a TypeScript and Next.js architecture deployed through Vercel, with PostgreSQL data services, secure authentication, firm-level authorization, and connected cloud-document storage.

Google Gemini and Google Antigravity are integrated through MCP. Instead of giving the AI unrestricted database access, WorkpaperOS exposes defined tools with validation, authorization, firm and engagement boundaries, and human-controlled actions.

The MCP connection uses OAuth authorization, scoped access, protected-resource discovery, and secure token handling. This allows compatible AI agents to discover and use WorkpaperOS capabilities without exposing credentials or unrestricted application data.

Development was led by Benedict E. Mbecha using Google Antigravity, Codex, and Claude. A software engineer provided unpaid technical oversight and review, while two chartered accountants independently reviewed and tested the professional workflows.

Challenges

The most important challenges were not simply generating AI responses. They included:

  • Preventing cross-firm and cross-engagement information exposure
  • Enforcing authorization for every AI operation
  • Limiting and validating Gemini output
  • Maintaining professional review and sign-off
  • Connecting cloud documents without weakening engagement security
  • Supporting OAuth and different MCP client transports
  • Designing accounting and audit tools that produce reviewable results rather than unexplained AI conclusions

These requirements led us to use narrow, validated MCP tools instead of broad database access.

Accomplishments

We completed working AttestWork and WPOSBooks applications with secure Gemini-connected workflows during the May 19–August 17, 2026 competition period.

The submitted demonstration shows an AI agent moving between audit and accounting workflows: accessing an engagement, working with engagement documentation and files, reviewing bank activity, categorizing a transaction, and completing a reconciliation.

What we learned

AI is most valuable in professional services when it operates within clearly defined authority. Reliable professional AI requires context boundaries, evidence, validation, audit trails, and human accountability—not only a capable language model.

MCP provided an effective way to make application functions available to Gemini while preserving those controls.

What’s next

We plan to expand governed agent workflows, improve explainability and review controls, add more accounting integrations, and continue independent security and professional testing.

Submission scope

Only AttestWork and WPOSBooks are included in this competition submission. Other WorkpaperOS applications and unfinished modules—including payroll, tax, and other works in progress—are outside the submission scope.es

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for WorkpaperOS: AttestWork + WPOSBooks

Built With

Share this project:

Updates