AI agents for financial reporting without code

Automating Financial Reporting with a Chain of AI Agents, No Code Required

A system of 17 specialized AI agents has fully automated financial reporting for a B2B platform with over 200 partners. Each agent handles a specific task: from daily reports to monthly reconciliations and contract generation. Integration via MCP servers provides direct access to MS SQL, Google Sheets API, Drive, and GitLab. In three months, routine operations that took 2–3 hours daily were reduced to minutes without writing any code.

Specialized Agents in a Unified Ecosystem

Agents are organized as skills in Claude Code—markdown files with instructions. Each skill focuses on a single operation, interacting through shared infrastructure:

• MCP server for read-only access to MS SQL.
• Google Sheets API for creating and formatting spreadsheets.
• Google Drive for uploading supplier data exports.
• GitLab API and task trackers for productivity analysis.

Agents don't communicate directly but use shared resources. This allows the system to scale without restructuring.

Financial Report Generation Algorithm

The command /financial-report Partner Period initiates the process in 40 seconds:

• Resolve Counterparty: mapping aliases to UUIDs from the internal database.
• SQL Query: building a funnel (leads → transitions → payouts → commission) considering deal statuses.
• Aggregation: grouping by day, calculating conversion rates and commissions based on individual splits.
• Write to Sheets: creating a spreadsheet with data, totals, and formatting.
• Verification: re-reading and cross-checking with source data to prevent row loss.

Formatting is applied after writing to avoid styling empty cells.

Monthly Reconciliations: Processing Multiple Sources

Reconciliation across 3–4 suppliers involves a chain of steps:

• Uploading files from Drive.
• Mapping deals with the database.
• Calculating commissions based on splits.
• Generating sheets with formulas for each partner.

Time reduced from 8 hours to 3–5 minutes. The agent accounts for different export formats and matching rules.

Full Range of Automated Tasks

| Skill | Task | Time |

|-------|------|------|

| Financial Report | Partner funnel | 40 sec |

| Daily Digest | Report for all partners | 2–3 min |

| Monthly Reconciliation | Reconciliation with suppliers | 3–5 min |

| Contract Generator | Standard contract | 1–2 min |

| Team Tracker | Team management | interactive |

| GitLab Analytics | Repository activity | 2–3 min |

| Meeting Report | Meeting transcription | 3–5 min |

The system uses 9 MCP servers (SQL, Sheets, Drive, Calendar, Gmail, GitLab, Tracker, Telegram, GitHub). Logic is described in markdown without code.

Advantages of Claude Code Over Alternatives

• ChatGPT + Zapier: no database access.
• LangChain/LangGraph: requires developers and long iteration cycles.
• Custom API + GPT-4: changes need deployment.

Claude Code enables a 30-second feedback loop: edit markdown → test → run. MCP servers provide infrastructure access.

Architecture:

• Claude Code as the runtime.
• MCP servers as bridges to systems.
• Skills as business logic in markdown.
• Sheets/Drive as the interface.

Key Problems and Solutions

• Sheets API loses rows: mandatory verification after batch writes.
• Database documentation mismatches: joining reference tables instead of using glossaries.
• Formatting: a separate step after data entry.
• Context loss: logging steps as external memory.
• Duplicates in GitLab: filtering merge commits and auto-generated content.
• Aliases: mapping names between systems.

Automation Results

| Metric | Before | After |

|--------|--------|-------|

| Daily Reports | 2–3 hr/day | 5–10 min |

| Monthly Reconciliation | 8 hr | 5 min |

| Contract Generation | 2–3 hr | 2 min |

| Errors | 3–5/month | 0 |

| People on Routine Tasks | 2 | 0 |

Savings of ~60 person-hours/month at $200/month for Claude Code.

Key Takeaways

• Specialized agents with MCP access replace manual labor in finance and analytics.
• Verification at each step prevents silent failures in API integrations.
• Markdown instructions speed up iterations without deployment.
• The system scales from one report to 17 skills.
• A product owner can build automation faster than a dev team.

— Editorial Team