Inspiration

AI is becoming increasingly capable of doing things on our behalf. It can find products, compare options and even make purchasing decisions. But as these capabilities grow, an important question arises: how much financial authority should we give an AI agent?

Imagine a small business owner who wants an AI assistant to purchase office supplies. They may be comfortable allowing it to find products and make purchasing decisions, but they would not want it spending beyond their budget, buying from unapproved merchants or making purchases without permission.

We wanted to address this gap by creating a way for people to delegate purchasing tasks to AI without giving up control over their money. That idea became MANDATE.

What it does

MANDATE is an AI-powered financial authorisation gateway for autonomous purchasing agents. It allows users to define purchasing instructions in natural language, including spending limits, approved merchants and when human approval is required.

An AI shopping assistant can discover products and propose purchases based on those instructions. Before a transaction proceeds, MANDATE independently evaluates it against the user's financial rules. Purchases that comply can proceed to PayPal checkout, while those that fall outside the mandate are blocked or escalated for human approval.

The idea is simple: people should be able to benefit from AI-powered purchasing without giving AI unrestricted access to their money.

How we built it

We built MANDATE by bringing together AI, product discovery, payment infrastructure and financial controls.

We integrated Google's Gemini to interpret natural-language purchasing instructions, assist with product discovery and review proposed transactions. Channel3 supports product discovery, allowing the shopping assistant to find relevant products for purchasing requests.

One of our most important design decisions was separating AI recommendations from financial authorisation. Rather than allowing an AI model to make the final payment decision, we implemented a deterministic policy engine that evaluates every proposed purchase against the user's actual spending rules. AI cannot override these restrictions.

We integrated PayPal's developer platform to support the payment journey, including order creation, buyer approval and payment capture through the Sandbox environment.

Supabase provides the data layer for user information, mandates and transaction records. We also implemented Row Level Security, audit events and mandate version history to improve data isolation, security and traceability.

Challenges we ran into

One of our biggest challenges was finding the right balance between AI autonomy and financial control. AI models can produce unexpected outputs or become unavailable, making it risky to rely entirely on them when authorising financial transactions. We addressed this by separating AI-assisted decisions from the deterministic rules that ultimately govern whether a purchase is permitted.

Integrating the different services was another challenge. Product discovery, AI processing, financial authorisation, human approval and PayPal payments each have different requirements. Bringing them together into one coherent purchasing experience required careful integration and testing.

We also had to consider what happens when multiple users interact with the platform. Each user's purchasing instructions, financial information and transaction records must remain separate. This made security and data access control an important part of our development process.

Accomplishments that we're proud of

We are proud of turning an idea about responsible AI purchasing into a working prototype that brings together AI-driven product discovery and financial authorisation.

One of our biggest accomplishments was implementing a deterministic policy engine that ensures purchasing decisions are checked against the user's instructions instead of relying entirely on AI-generated decisions.

We are also proud of integrating PayPal into the purchasing journey and building supporting features such as transaction records, audit logs and mandate version history. These capabilities help make purchasing decisions more transparent and traceable.

Most importantly, we built MANDATE around a problem that we believe will become increasingly relevant as AI agents gain the ability to transact on behalf of people and businesses.

What we learned

Building MANDATE taught us that introducing AI into financial workflows is not simply about making agents more autonomous. It is also about deciding where autonomy should end and human authority should begin.

We learned the importance of combining the flexibility of AI with predictable financial controls, making automated decisions understandable and treating payment security as a fundamental part of the product.

We also learned that integrating different technologies into one reliable experience requires more than getting individual components to work. The entire journey, from a user's instruction to the final transaction, must work together.

What's next for MANDATE - Financial Control for AI Agents

We envision MANDATE becoming a financial control layer for businesses that use autonomous AI agents to purchase goods and services.

As AI agents become more capable of taking actions in the real world, we believe people and businesses should not have to choose between the convenience of automation and control over their finances.

Our next step is to continue improving the purchasing experience, strengthen the payment and authorisation workflow, and explore how MANDATE can support more business purchasing scenarios.

Let AI handle the shopping. You stay in control of the money.

Built With

Share this project:

Updates

Submission history