-
-
PACT title card
-
Set the rules once ($150 item cap, auto-authorize up to $100)
-
Blocked by hard rules, with no override
-
Captured automatically and verified with PayPal (says the retailer step is a handoff only)
-
Above the auto limit: your approval is required
-
Every agent, every mandate, one Control Center
-
One gateway between any agent, any store and PayPal
-
Auto-authorized: $62.99 from a live Channel3 offer
-
Captured automatically and verified with PayPal (says the retailer step is a handoff only)
-
Spending analytics with AG Studio
-
Connect any agent in seconds pact-11-mcp.jpg: Agents can only ask
-
One approval, one exact purchase
PACT: The AI Payment Guardian
The financial authorization layer for autonomous AI agents
Let AI handle the shopping. Let PACT govern the spending. Let users stay in control.
1. Inspiration
AI agents are getting better at researching products, comparing prices and making decisions on behalf of users. But as they become more capable, there's one question we can't ignore:
How do we let AI act on our behalf without giving it unrestricted access to our money?
Traditional payment flows often require users to approve every transaction manually. This keeps spending under control, but defeats much of the purpose of autonomous agents. On the other hand, giving agents unrestricted access to payment methods creates obvious financial and security risks.
We wanted to find a middle ground.
That's why we built PACT, a financial authorization layer that lets AI agents make purchases within limits set by users.
2. What It Does
PACT sits between AI agents and payment execution. Every purchase request is checked against the user's spending mandate before it can proceed.
There are three possible outcomes.
AUTO-AUTHORIZE
The purchase meets all the mandate rules and falls within the user's automatic authorization limit. PACT allows it to proceed without a separate PACT approval.
REQUIRES APPROVAL
The purchase is outside the automatic authorization limit but is still permitted under the mandate. PACT asks the user to review and approve it.
BLOCKED
The purchase violates a hard restriction or security rule. PACT prevents it from proceeding.
Example
Suppose a user sets an automatic authorization limit of $100.
- A $62.99 purchase that meets all the rules is automatically authorized.
- A $109.99 purchase requires approval if the mandate allows exceptions.
- A purchase that violates a hard restriction is blocked.
PACT checks several things before making a decision:
- Spending limits and available budgets
- Merchant and product restrictions
- Purchase quantity and unit price
- Mandate validity and expiration
- Product offer integrity and price changes
- Agent identity and permissions
The important part is that AI does not make the final financial authorization decision. AI can find products and recommend purchases, but the PACT policy engine independently checks every request.
Once authorized, purchases follow PayPal's supported checkout and buyer authorization process.
3. How We Built It
PACT brings together AI, payment infrastructure, product discovery and agent interoperability.
Technologies and Integrations
Google Gemini
Used for AI-assisted purchasing, understanding product requests and generating recommendations.
Channel3
Used for discovering external products and retrieving retailer offers that PACT can evaluate against the user's spending rules.
PayPal Developer Platform
Integrated PayPal Sandbox Orders API v2 for order creation, buyer approval, payment capture and server-side transaction verification.
Model Context Protocol (MCP)
Allows compatible external AI agents to interact with PACT through MCP tools.
REST Agent Gateway
Provides APIs for external agents to connect to PACT and submit purchase requests using scoped credentials.
AG Grid and AG Studio
Used for transaction views, agent activity, analytics and the Control Center experience.
React, TypeScript, Python and FastAPI
Used to build the frontend, backend, policy engine and integration services.
Architecture
ANY COMPATIBLE AI AGENT
REST / MCP
|
v
PACT GATEWAY
|
v
DETERMINISTIC POLICY
ENGINE
|
+-------+-------+
| | |
v v v
AUTO APPROVAL BLOCK
AUTH REQUIRED
|
v
PAYPAL CHECKOUT
|
v
BUYER AUTHORIZATION
|
v
PAYMENT CAPTURE
|
v
AUDIT & ANALYTICS
Channel3 provides external product offers, while the Demo Store provides a controlled environment for demonstrating purchase scenarios.
How a Purchase Works
- An AI agent discovers a product and submits a purchase request to PACT.
- PACT identifies the agent and checks its permissions.
- The policy engine evaluates the request against the user's spending mandate.
- PACT automatically authorizes eligible purchases, asks for approval when needed, or blocks restricted purchases.
- Authorized purchases proceed through PayPal's supported checkout and buyer authorization flow.
- PayPal Sandbox processes the payment capture, and PACT verifies and records the transaction.
A note about the demo: Payments are processed using PayPal Sandbox. External retailer links take users to the retailer's website. They do not represent completed live purchases from those retailers.
4. Challenges We Faced
Balancing Autonomy and User Control
We didn't want users to approve every purchase, but we also didn't want agents spending money without proper limits. Configurable mandates and three authorization outcomes helped us find that balance.
Keeping AI Away from Final Financial Decisions
AI models can misunderstand instructions or be influenced by untrusted information. We made sure that financial authorization happens in our deterministic backend policy engine, independently of AI recommendations.
Handling Changing Product Offers
Product prices and availability can change between discovery and purchase. We implemented offer revalidation and purchase fingerprinting so that approvals remain tied to the exact purchase details. If important details change, an existing approval becomes invalid.
Preventing Unauthorized Spending
Giving an agent unrestricted payment credentials would defeat the purpose of PACT. We built the agent integration around scoped credentials, limited capabilities and controlled backend payment workflows.
Connecting Different Systems
Bringing together product discovery, external agents, policy evaluation, analytics and PayPal required careful handling of data and transaction states. We also had to make sure that a successful API response was not mistaken for a completed payment.
5. Accomplishments
Here's what we managed to build:
- An AI-assisted purchasing experience with external product discovery.
- Configurable spending mandates and automatic authorization limits.
- Three authorization outcomes: auto-authorize, requires approval and blocked.
- Live product discovery through Channel3.
- PayPal Sandbox Orders API v2 integration.
- Automatic payment capture after the required PayPal buyer authorization, without a separate manual Capture action.
- REST and MCP integrations for external AI agents.
- Scoped agent credentials and controlled purchasing capabilities.
- Purchase fingerprinting and offer revalidation.
- Transaction auditing and analytics through AG Studio.
We're particularly proud that we were able to demonstrate an external AI agent initiating purchases without having unrestricted access to the user's money.
6. What We Learned
Building PACT taught us that autonomous commerce involves much more than connecting an AI agent to a payment API.
We learned that:
- AI reasoning and financial authorization should be separate.
- Sensitive financial decisions need deterministic rules.
- Agents should only receive the permissions they actually need.
- Human approval should be reserved for exceptions rather than every routine purchase.
- Purchase details need to be checked again before payment.
- Payment authorization and payment capture are two different steps.
- Agent activity should be visible and auditable.
The biggest takeaway for us was that autonomy and user control don't have to be opposites.
7. What's Next?
We see PACT becoming a reusable financial authorization layer that works with different AI agents, regardless of which model or framework they use.
Some things we'd like to work on next:
- Support for more payment providers and payment methods.
- More e-commerce and product discovery integrations.
- Reusable spending mandates across multiple agents.
- Better risk detection and risk-based authorization.
- More detailed spending analytics and configurable policies.
- Stronger production security, compliance and deployment architecture.
Our goal is to make secure financial autonomy a standard part of the agentic economy.
Final Thought
AI discovers and recommends. PACT evaluates and authorizes. Humans handle exceptions. Payment providers execute payments through their supported flows.
AI agents should be able to act on our behalf without having unrestricted control over our finances.
That's what we're building with PACT.
Let AI handle the shopping. Let PACT govern the spending. Let users stay in control.
Built With
- ag-grid
- agentic-ai
- ai-agents
- artificial-intelligence
- channel3
- fastapi
- fintech
- generative-ai
- human-in-the-loop
- mcp
- payment-security
- paypal
- paypal-checkout
- python
- react
- rest-api
- typescript