Proofed — AI Can Buy Anything. But Can It Prove You Got It?

AI agents are learning how to shop, negotiate, and pay. But there's something they still can't reliably do: know whether you actually received what they purchased.

Imagine asking your personal AI agent to buy something.

It finds the product, checks your budget, and processes the payment. Everything happens in seconds.

But what happens next?

The courier delivers to the wrong address. A package arrives damaged. A vending machine fails to dispense the item.

The payment worked. The physical transaction didn't.

And suddenly, a seamless autonomous purchase becomes a customer support problem, a refund request, or a payment dispute.

This is one of the missing pieces in agentic commerce.

As AI moves beyond digital services into physical transactions, we need infrastructure that connects financial execution to what actually happens in the real world.

That's why we built Proofed.

Authorize now. Capture only after verified handover.

The Missing Half of Agentic Commerce

Most AI commerce solutions focus on helping agents discover products and complete purchases.

But checkout is just the beginning.

A successful payment authorization doesn't guarantee that a product was delivered, that a machine dispensed an item, or that a physical service was completed.

We wanted to explore a different approach:

What if AI agents could initiate purchases, but financial settlement depended on evidence that the physical transaction succeeded?

Proofed connects three capabilities:

1. Agentic purchasing. An AI assistant understands customer intent, searches available inventory, and prepares transactions within defined budgets.

2. Physical-world verification. Computer vision and additional verification signals evaluate evidence associated with a delivery or handover.

3. Conditional payment execution. PayPal authorizations allow the system to separate buyer approval from final payment capture, enabling different financial outcomes depending on fulfillment.

The result is an agentic commerce workflow that accounts for more than a successful checkout.

We Built Two Outcomes. One Payment Architecture.

We wanted the demonstration to go beyond a chatbot connected to a payment API.

So we built a controlled physical handover prototype using real photographs of packages and doorways, our delivery application, multimodal AI, and PayPal Sandbox transactions.

The scenarios are staged to demonstrate the system's behavior; they are not completed commercial deliveries. No real purchases or transfers of customer funds are involved.

What we wanted to test was the complete interaction between AI reasoning, physical evidence, deterministic verification policies, and payment execution.

Scenario 1 — The Purchase Succeeds

A shopper opens Proofed and asks:

"I have a $20 budget. What coffee do you have?"

Our MiniMax-powered agent searches the catalog and recommends Ethiopian Yirgacheffe coffee for $18.50.

But there's an important distinction.

The AI doesn't decide what something costs.

Product prices, budgets, and transaction amounts are validated by a deterministic TypeScript policy engine using integer-cent arithmetic.

The agent prepares the purchase, and the shopper proceeds through PayPal's approval flow.

Using PayPal Orders v2 with intent=AUTHORIZE, Proofed creates an authorization rather than immediately capturing payment.

The purchase is authorized. The payment has not yet been captured.

Scenario 2 — Bringing the Physical World Into the Transaction

Next, we simulate the handover.

Our delivery application collects evidence from the physical environment. We use MiniMax multimodal vision to interpret a doorstep photograph, identify the apartment number, and assess the visible package condition.

Proofed also incorporates location and recipient-confirmation signals into its verification workflow.

The objective is to establish whether the available evidence satisfies the conditions for completing the transaction.

In our successful demonstration, the verification flow produces a positive handover result and the application proceeds with PayPal's authorization-capture operation.

The sandbox transaction illustrates the intended payment lifecycle:

Buyer approval → PayPal authorization → Handover verification → PayPal capture

Scenario 3 — What Happens When Delivery Fails?

Here's where Proofed becomes more interesting.

We deliberately introduce a failure into a second staged delivery scenario.

The courier approaches the wrong apartment.

The submitted photograph doesn't match the expected destination.

Proofed's verification workflow rejects the handover, and the transaction is not treated as successfully fulfilled.

A shopping agent might consider the purchase complete. Proofed recognizes that fulfillment remains unresolved.

But detecting the problem is only half the solution.

The agent also needs to help the customer resolve it.

Our conversational interface explains the failed verification and presents recovery options, including another delivery attempt or cancelling the order.

When the buyer selects cancellation, Proofed demonstrates voiding the uncaptured authorization through PayPal.

No successful handover. No captured payment.

Two Transactions. Two Different Outcomes.

In our PayPal Sandbox demonstration:

Successful handover	Failed handover
$18.50 authorized	$38.00 authorized
Physical verification succeeds	Address verification fails
Authorization captured	Authorization voided
Successful payment outcome	No payment captured

This is the central idea behind Proofed.

AI can initiate commerce. But the outcome of a physical transaction should matter to how the payment is completed.

The Engineering Behind Proofed

We designed Proofed around one fundamental constraint:

Probabilistic AI should never have unrestricted authority over deterministic financial operations.

That shaped how we separated the system into five components.

AI Agents — Understanding Intent and Exceptions

We use MiniMax for natural-language interaction, product discovery, tool calling, and multimodal image interpretation.

The agent helps customers find products and understand unexpected delivery outcomes.

Financial calculations and permission checks remain outside the model.

PayPal — Authorize, Capture, Void

PayPal provides the financial infrastructure for our prototype.

We integrated:

PayPal Orders v2: Creating checkout orders with intent=AUTHORIZE.
PayPal JavaScript SDK: Buyer-facing checkout and approval.
PayPal Payments v2: Authorization capture, void, and refund functionality.
Transaction lifecycle management: Tracking the relationship between order states, verification outcomes, and payment operations.

Rather than implementing another payment network, we're exploring how PayPal's existing authorization capabilities can support more accountable agentic workflows.

Multimodal AI — Interpreting Physical Evidence

MiniMax's multimodal capabilities allow Proofed to analyze photographs of physical delivery environments.

Our prototype examines visible address information and package condition, incorporating these findings into the verification process.

This also exposes an important engineering challenge: computer vision is evidence, not certainty.

A photograph can be ambiguous, GPS can be inaccurate, and an image hash alone cannot prove that a customer received an item.

That's why we're exploring verification workflows that combine multiple signals rather than relying on a single AI classification.

Deterministic Financial Controls

We developed backend policies for catalog pricing, budget limits, and transaction states.

Money is handled using integer cents, avoiding floating-point arithmetic and model-generated price calculations.

An order's payment status is tracked separately from its fulfillment status.

This separation is critical when an AI agent interacts with real payment infrastructure.

Merchant Operations and Auditing

We also developed a merchant dashboard with AG Grid to inspect orders, payment activity, authorization states, and fulfillment exceptions.

The system includes order-event logging and database components built with SQLite and Turso/LibSQL.

Our stack combines Next.js, React, TypeScript, Node.js, MiniMax, PayPal APIs, AG Grid, and Vercel.

The Hardest Part Wasn't Making the Agent Buy Something

The real challenge was designing a system where autonomous reasoning interacts with financial operations and imperfect physical evidence.

We encountered several important problems while building Proofed.

Buyer consent cannot be confused with agent intent.

The AI can prepare a purchase, but the customer must approve the relevant PayPal payment flow.

Financial decisions need deterministic boundaries.

Language models are useful for understanding ambiguous requests. They shouldn't be responsible for calculating prices or bypassing payment policies.

Physical verification is inherently uncertain.

A photograph might contain the correct address without proving the package was handed over. We had to think carefully about what each verification signal actually establishes.

Failures require a product experience, not just an error code.

When verification fails, the customer needs an explanation and a practical resolution. This is where agentic reasoning becomes especially valuable.

And we learned that testing these systems requires more than happy-path demonstrations. We built automated tests and scenarios for successful purchases, fulfillment failures, conflicting evidence, and repeated recovery attempts.

The goal isn't to make AI infallible.

It's to build systems that can recognize uncertainty and keep financial operations within clearly defined rules.

Beyond the Prototype

Proofed is currently a proof of concept demonstrated with controlled physical scenarios and PayPal Sandbox.

But the problem we're exploring is much larger.

Imagine AI agents purchasing from automated retail kiosks, coordinating deliveries, reserving merchandise, or interacting with connected dispensers.

In all of these environments, a successful digital transaction doesn't necessarily mean that physical fulfillment succeeded.

We envision Proofed evolving into an infrastructure layer that connects autonomous purchasing to verifiable fulfillment events and payment workflows.

That could eventually mean integrating with delivery providers, warehouse systems, connected hardware, automated retailers, and other physical-commerce platforms.

The long-term opportunity isn't simply another AI shopping assistant.

It's infrastructure for making AI-initiated commerce accountable beyond checkout.

PayPal provides the authorization and payment execution capabilities. AI helps interpret intent and real-world exceptions. Proofed explores how those capabilities can work together.

Because as AI agents become capable of spending money on our behalf, we believe one question will become increasingly important:

Did we actually get what we paid for?