Sosys creates a private financial agent with AI and multi-source data

Sosys ( Independent Software Vendor) pioneer in the development of intelligent business ecosystems. Operating on a B2B enterprise model, the company serves a portfolio of medium and large-sized clients distributed across multiple vertical sectors that demand high-fidelity analytical processing.

Sosys' operation is based on delivering governance, regulatory compliance, and transactional visibility. Faced with a complex macroeconomic scenario, driven by the imminent Tax Reform and the need to consolidate fiscal and operational data, Sosys faced the challenge of scaling its value proposition. The central objective was to transition from static analytical reports to a predictive and decentralized decision-making architecture, eliminating dependence on manual workflows and time-consuming queries to its clients' relational databases.

Real problem

In the traditional operating model of Sosys' clients, critical decision-making suffered from data latency generated by information silos. The fundamental variables for calculating run rate, cash flow projection, and working capital needs were fragmented across multiple environments: relational databases in ERP systems, interaction logs in CRM, and decentralized spreadsheets.

Operational bottlenecks and technical limitations

• Excessive analytical latency: the consolidation of structured management reports depended on manual data extraction and cross-referencing processes. Answers to strategic questions, such as contract breach simulations or the impact of default on cash flow, required days of internal processing and consecutive technical meetings.
• Fragmentation and schema inconsistency: the absence of a unified data cleansing layer caused metric discrepancies between the revenue recorded in the CRM and the accounting reconciliation in the ERP.
• Compliance and fraud risks: the direct handling of sensitive financial data by human operators increased the attack surface for fraud and tax non-compliance, violating corporate governance and information security principles.
• Direct financial impact: the inability to project dynamic scenarios based on payment and collection terms (DSO and DPO) resulted in inefficient capital allocations, generating opportunity costs and a mismatch between projected and actual cash flow.

Solution architecture

To overcome these limitations, Sosys used Skyone Studio, an intelligent integration and AI platform that unifies iPaaS tools, Lakehouse, AI Agents, and conversational interfaces with BI. The architecture implemented to create Nanda, Sosys' virtual CFO, was structured in five main layers within the Skyone Studio processing flow:

Components and technology stack

1. Pipelining via iPaaS (Integration Platform as a Service): asynchronous connection through data pipelines integrating the APIs of ERP, CRM, and legacy market systems. Automated flows perform real-time data ingestion, standardizing heterogeneous schemas.
2. Lakehouse Layer and Data Management: using the Skyone Studio managed repository to organize the analytical journey. Raw data (Raw DB) undergoes automated transformation and cleaning processes (Data Cleaner 2.0) to be persisted in optimized views (Prepared DB) and segmented into Data Marts by client.
3. Agentic Workflow Orchestration: Implementation of an autonomous module based on private Large Language Models (LLMs). The orchestration agent is capable of understanding complex textual objectives, planning sequences of actions, selecting tools (skills ), and performing scans via structured prompts in the analytical database.
4. Security, Governance, and Anti-fraud Layer: all payload transition and handling occurs within a protected ecosystem. Strict validation rules block unauthorized requests and audit transactional inconsistencies before sending the data to the final interface.
5. Omnichannel Multichannel Distribution: native integration of Skyone Studio with corporate messaging, making intelligent agent capabilities available via WhatsApp Business, Microsoft Teams, and structured private chat environments.

Technical challenges

The engineering behind the Nanda agent required mitigating trade-offs between data architecture and generative AI:

• Data sanitization based on tokenization: Language models operate under strict context constraints (where 1 token = 4 characters or 0.75 words). Feeding an LLM with raw financial tables would exceed the context window and increase operational costs. Skyone Studio solved this by applying structured summarization and native natural language translation to SQL queries (text-to-SQL) directly in the Data Marts , optimizing token consumption.
• Guarantee of analytical determinism: pure generative models are subject to hallucinations, an unacceptable risk for accounting balances. The architecture implemented RAG (Retrieval-Augmented Generation) techniques coupled with cross-validations. Artificial intelligence only responds based on rigorously structured and validated data contained in the solution's internal Data Lake.
• Segregation of sensitive data contexts: To prevent data leakagebetweendifferent corporations served by Sosys, the Studio logically isolated the knowledge bases into individual encrypted partitions, maintaining unique access keys linked to each user organization's ID.

Implementation

The project rollout was executed in four macro-structured phases within the unified Skyone ecosystem:

Phase 1: Mapping of system operators and ingestion (iPaaS)

Mapping of all customer transactional sources (accounts payable, accounts receivable, billing tables, and CRM logs). Configuration of pre-built connectors and API buses via Skyone Studio iPaaS pipelines, ensuring automated and continuous ingestion of information.

Phase 2: Lakehouse structuring and sanitation

Centralization of ingested data in the Data Lake layer. Implementation of logical data transformation flows to purge duplicates, handle null fields, and convert strings into standardized numeric formats for financial auditing.

Phase 3: Prompt engineering and agentic workflow

Development of the intelligent agent's decision-making flow in Studio. Configuration of specific skills,such as: cash flow simulation triggers, tax compliance verification routines, and optimized query generators for the database. Integration with selected LLMs and calibration of temperature hyperparameters to eliminate conceptual deviations.

Phase 4: Security validation and omnichannel publication

Approval of the integrated anti-fraud layer. Activation of multi-channel publishing gateways to connect Nanda directly to the WhatsApp and Microsoft Teams production environments of approved clients, enabling real-time corporate interactions via audio and text.

Measurable results

The transition of analytical operations to the generative AI-assisted ecosystem at Skyone Studio has yielded quantifiable structural improvements:

• Reduced analytical response time: complex queries for financial scenario planning and generation of structured management reports have dropped from days to seconds, operating in real time (24/7).
• Operational efficiency: a drastic reduction in rework and the need for technical alignment meetings for manual extraction of corporate databases.
• Active mitigation of fraud risks: implementation of automated real-time alerts. If the agent identifies a scheduled payment without the corresponding incoming document in the ERP, the flow generates a protective compliance trigger, blocking potential bottlenecks or operational fraud.
• Increased availability of information: accelerating tactical decisions in executive committee meetings, eliminating delays due to the immediate absence of validated macroeconomic indicators.

Lessons learned

1. Intelligence depends on prior structuring: advanced generative AI models lose operational utility if applied to disorganized or corrupted databases. Nanda's success lay in Skyone Studio's ability to process, cleanse, and organize the data in the Lakehouse layer before exposure to LLM.
2. Application-layer focused approach: Agile development of innovative enterprise solutions does not require companies to build fundamental AI models from scratch. Real enterprise value is unlocked by orchestrating existing LLM infrastructure with private, business-specific data.
3. Security as a prerequisite for scaling AI: In the B2B enterprise segment, intelligent processing of sensitive data is only viable under strict governance frameworks and isolated environments that ensure continuous regulatory compliance.

FAQ

What is a private financial agent with AI?

This is an intelligent architecture based on Language Models (LLMs) that operates in a closed corporate environment. Unlike public artificial intelligence, the private agent consumes exclusive internal data from a company (such as ERP and CRM), guaranteeing total confidentiality, Zero Trust governance, and highly accurate analytical responses without external information sharing.

How does Skyone Studio ensure the security of sensitive financial data?

Skyone Studio works by unifying layers of security, compliance, and anti-fraud barriers in data flow. Company data undergoes logical isolation in dedicated Lakehouse structures, preventing cross-access or data leakage,while maintaining full transactional traceability in compliance with strict corporate governance standards.

How does the multi-source data integration process work in Skyone Studio?

The platform operates through an integrated iPaaS solution that centralizes and orchestrates information flows from over 400 market systems (such as ERPs, CRMs, and external databases). These pipelines extract, cleanse, and standardize structured and unstructured data, automatically loading it into a unified Lakehouse for immediate consumption by AI agents.

Could the use of generative AI in corporate finance cause errors in reports?

To avoid analytical errors or hallucinations, the data architecture applied in Skyone Studio adopts advanced RAG (Retrieval-Augmented Generation). This means that the artificial intelligence agent is technically limited to answering questions using exclusively real, clean, and validated data contained in the organization's private Data Lake, ensuring mathematical determinism in the answers.