2025-09-16
- Customer
- Planning
- Project manager on the customer side
-
- IT Provider
-
Squad Builder
- Year of project completion
-
2025
- Project timeline
-
August, 2025 - September, 2025
- Project scope
-
200 automated workstations
- Goals
-
Our project, Conciliador AI, was created to eliminate one of the most invisible yet critical pain points for finance teams: the manual reconciliation of accounting data. On average, accountants spend 3 hours per client every month on reconciliation. For firms with 2,000 clients, that means 6,000+ hours lost monthly to repetitive, error-prone work. The goal of Conciliador is to drastically reduce this time through specialized AI agents that clean, align, and reconcile data automatically. By leveraging natural language instructions, users can instantly generate reusable data-cleaning templates and execute reconciliations with speed, accuracy, and full auditability. The outcome is not only efficiency and cost reduction, but also governance and transparency: every action is versioned, traceable, and aligned with compliance requirements. Our vision is to free professionals from operational bottlenecks so they can focus on strategic insights rather than clerical tasks.
- Project Results
-
The project delivered a measurable breakthrough in efficiency. A reconciliation process that previously consumed 3 hours per client is now completed in an average of 5 minutes, representing a 36x productivity gain. This shift has freed thousands of hours per month, allowing professionals to redirect their time toward analysis and strategic decision-making rather than clerical tasks. Results are now fully auditable, with every action versioned and logged, eliminating the need to hunt for wrong formulas hidden in piles of Excel sheets. Quality is reinforced by our evals pipeline, ensuring accuracy and compliance. The combination of drastic time reduction, reliability, and auditability demonstrates the tangible value of agentic AI in solving critical but often overlooked operational bottlenecks.
The uniqueness of the project
Conciliador AI is unique because it goes far beyond traditional RPA or reconciliation software. Instead of requiring rigid rules or complex setup, our solution uses natural language to generate reusable data-cleaning templates on the fly, turning accountants’ instructions into automated workflows. Every reconciliation is fully auditable and versioned, embedding governance by design. We leverage evals as a continuous quality control layer, ensuring accuracy and reliability. Our agents also learn from mistakes and improve over time, creating a compounding effect that static tools cannot replicate. Unlike niche point-solutions, Conciliador is packaged as a “one-click hire” AI team: users do not configure, they activate. This mix of agentic architecture, embedded compliance, continuous learning, and auditability makes Conciliador uniquely positioned to transform reconciliation at scale.
- Used software
-
Tech stack
- NextJS: frontend
- Mongodb: database layer
- OpenAI GPT-4.1: LLM
All hosted on Render.com with a fully automated DevOps pipeline.
- Difficulty of implementation
-
The greatest challenge was that our initial approach attempted to solve the entire reconciliation process directly with GenAI, which proved unreliable and inconsistent. Through trial and error, we realized that large models are powerful for interpreting intent but not for executing complex, deterministic operations. We shifted the design: GenAI is now used exclusively for translating accountants’ natural language instructions into code scripts. These scripts are then stored as reusable templates that execute deterministically across datasets. This hybrid approach preserved the flexibility of natural language while ensuring accuracy, repeatability, and auditability. Achieving this balance between AI creativity and deterministic execution required multiple iterations, extensive testing, and the design of an evals pipeline to guarantee quality control. The final architecture is both practical and scalable, but it emerged only after overcoming the limitations of a GenAI-only solution.
- Project Description
-
Accounting reconciliation is one of the most time-consuming and error-prone processes inside finance teams. On average, firms spend three hours per client per month cleaning, aligning, and checking spreadsheets. For organizations handling thousands of clients, this means more than 6,000 hours of repetitive work every month. The cost is not only measured in time and salaries but also in delayed closings, missed discrepancies, and exposure to compliance risks.
The project was designed to eliminate this bottleneck by transforming reconciliation into an automated, intelligent, and auditable process. At its core, the system leverages specialized AI agents capable of interpreting natural language instructions and converting them into reusable data-cleaning templates. Instead of building rigid scripts or macros, accountants can simply describe what needs to be done, and the agents generate the appropriate transformation logic. Once created, these templates are versioned, stored, and applied consistently across new datasets.
Execution is asynchronous and fully tracked. Each reconciliation request creates a task with its own ID, enabling users to monitor progress and retrieve results with complete traceability. Every action taken by the agents is logged and auditable, supporting both internal governance and external compliance requirements. This ensures that automation does not compromise accountability, a key weakness in many AI deployments.
Reliability is reinforced by continuous evaluation pipelines. A custom evals framework measures the quality of each agent’s output against defined criteria, acting as a quality control layer. When discrepancies or errors are detected, agents learn from their mistakes and improve over time. This feedback loop transforms the system from a static automation tool into a self-improving process engine, driving consistent gains in accuracy and efficiency.
Tech stack: the frontend is built with Next.js to deliver a clean and intuitive user experience. Large language model reasoning is powered by OpenAI GPT-4.1. Data, templates, and reconciliation histories are persisted in MongoDB with full versioning and audit trails. The entire solution is hosted on Render.com with a fully automated DevOps pipeline for build, deploy, environment management, and rollback. Auxiliary systems include an evals pipeline for continuous quality checks and observability for monitoring agent performance.
The impact is immediate and measurable. Hours of manual work are reduced to minutes. Accountants gain time to focus on higher-value analysis instead of clerical adjustments. Leaders benefit from faster and more reliable closings, while organizations improve both productivity and compliance.
The broader vision is to extend this approach to other invisible but indispensable processes across finance and operations. By packaging specialized AI agents as “one-click hire” teams, the project demonstrates how critical back-office activities can be transformed with scalability, auditability, and continuous improvement at the core.
- Project geography
-
The project was launched in Brazil, but its flexible architecture makes it global-ready from day one.
