PACT — Verified Payments for AI Agents

Inspiration

Imagine hiring an AI agent to complete a task and paying it upfront. What happens if the agent disappears, delivers incomplete work, or simply fails to meet the agreement?

Traditional payment systems move money. They don't necessarily understand whether the work behind that payment has been completed.

As AI agents evolve from assistants into autonomous economic participants, we believe they need more than wallets and payment APIs. They need accountability.

That is why we built PACT: a payment coordination layer designed to connect agreed work, verifiable conditions, and payment settlement.

Our central idea is simple: payment should follow verified performance, not promises.

What it does

PACT — Verified Payments for AI Agents — introduces a structured approach to conditional payments between buyers and autonomous service providers.

It is designed around three stages:

1. Authorize: A buyer initiates and authorizes a payment through PayPal.
2. Verify: PACT evaluates whether the agreed delivery conditions have been satisfied.
3. Settle: The intended settlement logic captures payment when the required conditions pass or voids the authorization when they fail.

PACT also maintains persistent transaction records and an auditable history of payment-related events.

Rather than treating every payment as an isolated transaction, PACT creates a traceable relationship between the agreement, verification process, and payment outcome.

The goal is to make payments between AI agents and humans more accountable, transparent, and predictable.

How we built it

We built PACT as a web application using a combination of payment infrastructure, AI services, and persistent data storage.

Our implementation includes:

• PayPal Sandbox: Payment integration for testing the authorization and payment lifecycle.
• Groq AI: Integration using the "openai/gpt-oss-20b" model through Groq.
• Neon PostgreSQL: Persistent storage for application and transaction records.
• Vercel: Application hosting and deployment.
• Verification and audit infrastructure: Structured records and hash-chain verification designed to make transaction histories traceable.

We designed the architecture to separate payment operations, AI-assisted decision-making, and transaction recordkeeping.

This separation is important because an AI-generated recommendation should not automatically become an irreversible financial action.

Payment operations must remain governed by explicit conditions and the payment provider's authorization mechanisms.

Challenges we ran into

One of our biggest challenges was coordinating multiple systems that have different responsibilities.

PayPal handles payment authorization and payment operations. The AI component assists with decision-making. PACT must connect these components without allowing an AI response to bypass payment controls.

We also encountered challenges with asynchronous payment workflows.

A payment approval, a server-side callback, and a successful settlement are different events. Treating them as interchangeable can create incorrect payment states.

During Sandbox testing, we encountered a failure after the buyer approved the PayPal payment. This highlighted the importance of verifying the complete authorization-to-settlement lifecycle rather than assuming that an approval screen means a successful payment.

Maintaining durable transaction records and making the resulting history independently auditable introduced another layer of engineering complexity.

These challenges shaped our approach: financial state must be established through verified payment-provider responses, not assumptions.

Accomplishments that we're proud of

We are proud of building a working deployed prototype that brings AI-assisted payment coordination and PayPal integration into one application.

Some of our key accomplishments include:

• Integrating PayPal Sandbox into the PACT payment workflow.
• Connecting an AI model through Groq.
• Implementing persistent PostgreSQL-backed application storage.
• Building transaction recordkeeping and hash-chain verification.
• Designing a conditional settlement architecture around explicit delivery requirements.
• Deploying the application publicly so the project can be explored beyond a local development environment.

Most importantly, we have established the foundation for a payment system where the relationship between work, verification, and settlement is treated as a first-class part of the transaction.

What we learned

We learned that connecting an AI model to a payment API is relatively easy compared with building a system that can be trusted with financial decisions.

AI-generated decisions need boundaries. Payment authorization needs explicit handling. Transaction histories need to be durable and auditable.

We also learned that payment approval is not the same as payment capture, and that every transition in a financial workflow needs to be independently verified.

The biggest lesson was that the future of agentic commerce will not be defined simply by how quickly AI agents can move money.

It will be defined by how reliably they can justify those transactions.

What's next for PACT — Verified Payments for AI Agents

Our next priority is completing and verifying the entire PayPal payment lifecycle, including successful authorization, conditional capture, and voiding when delivery requirements are not met.

Beyond that, we want to expand PACT into a more comprehensive settlement infrastructure for agent-to-agent commerce.

Our roadmap includes:

• Stronger, independently verifiable delivery evidence.
• More sophisticated verification policies.
• Improved dispute and failed-delivery handling.
• Expanded auditability and transaction transparency.
• Support for additional payment and agent-service integrations.
• A developer-facing interface that makes conditional settlement easier to integrate into autonomous applications.

Our long-term vision is to make PACT a trust layer for the emerging agent economy.

Because when software starts doing business with software, trust cannot depend on promises.

It has to be built into the transaction.

Built With

• agentic
• agents
• ai
• api
• artificial
• chain
• financial
• fintech
• gpt-oss-20b
• groq
• hash
• intelligence
• neon
• payment
• paypal
• postgresql
• processing
• rest
• transaction
• vercel
• verification