Finance Copilot: From Spend Visibility to Action

Inspiration

Finance Copilot was inspired by my earlier project Chotu — a Hindi-first, voice-powered AI assistant designed to help neighborhood shopkeepers run their entire dukaan (sales, khata, stock, and payments) using simple spoken commands.

While building Chotu, a key insight became clear:
people don’t struggle because systems lack data — they struggle because systems don’t clearly guide them on what to do next.

Shopkeepers using Chotu didn’t want dashboards or reports. They wanted clarity and action. That same philosophy shaped Finance Copilot — this time applied to finance teams dealing with expense data instead of shop-level operations.

What It Does

Finance Copilot transforms raw expense data into actionable insights.

It automatically:

Categorizes expenses using explainable rules
Evaluates tax impact
Flags high-risk or compliance-sensitive transactions

Instead of requiring users to manually scan dashboards, Finance Copilot highlights what needs attention and notifies teams through Slack with direct links back to Tableau for investigation.

The focus is not just spend visibility — it is decision support.

How We Built It

The project is built around Tableau Cloud as the core analytics and logic layer.

Expense data is ingested as CSV and modeled using a semantic layer inside Tableau. Rule-based logic is applied to categorize expenses, calculate tax impact, and assign risk scores.

Expense Categorization

Expenses are categorized using keyword-based rules on transaction descriptions (e.g., Travel, Software, Utilities), with an explicit Unknown category to surface ambiguity instead of hiding it.

Risk Scoring Logic

Risk is calculated using a transparent, additive rule-based model:

$$ \text{Risk Score} = \mathbb{1}(\text{High Amount}) + \mathbb{1}(\text{Missing Tax}) + \mathbb{1}(\text{Unknown Category}) $$

Where:

High Amount = transaction amount exceeds a predefined threshold
Missing Tax = tax amount equals zero
Unknown Category = expense could not be confidently categorized

Risk levels are derived as:

$$ \text{Risk Level} = \begin{cases} \text{High}, & \text{Risk Score} \geq 3 \ \text{Medium}, & \text{Risk Score} = 2 \ \text{Low}, & \text{Risk Score} = 1 \ \text{None}, & \text{Risk Score} = 0 \end{cases} $$

Dashboards

Insights are presented through three dashboards:

Finance Overview — spend, tax impact, and trends
Risk & Compliance — flagged transactions and risk distribution
Expense Drill-Down — vendor- and category-level investigation

To demonstrate actionable analytics, Slack notifications are integrated using incoming webhooks, linking alerts directly to Tableau dashboards.

The system is intentionally explainable and auditable, avoiding black-box models in favor of business trust.

Challenges We Ran Into

One major challenge was working within platform constraints while still demonstrating real-world workflows. Native automation capabilities were limited in the development environment, requiring careful design of the alerting flow to remain honest, demo-ready, and enterprise-aligned.

Another challenge was scope discipline — resisting the temptation to add machine learning or overengineering, and instead focusing on clarity and impact within a limited hackathon timeline.