I actually think this is one of the most exciting areas in payments over the next 5–10 years, and it aligns remarkably well with your experience.

The interesting part is that AI will not replace payment authorization—it will reshape the authorization decision itself.

Today, authorization is essentially a deterministic rules engine.

Merchant
    ↓
Acquirer
    ↓
Card Network
    ↓
Issuer
    ↓
Approve / Decline

The issuer evaluates:

  • Card status
  • Available funds
  • CVV/CVC
  • AVS
  • Velocity rules
  • Fraud score
  • 3DS result
  • Merchant category
  • Device fingerprint
  • Historical behavior

Everything ultimately reduces to a binary decision.

Where AI fits today

Today, AI is mostly being used as a risk scoring engine, not as the authorization engine itself.

For example:

Authorization Request
        │
        ▼
Feature Extraction
        │
        ▼
AI Fraud Model
        │
        ▼
Risk Score (0-100)
        │
        ▼
Issuer Decision Engine
        │
        ├── Approve
        ├── Decline
        └── Step-up Authentication

The important point is that AI advises; the payment network still makes a deterministic authorization decision because payment systems require consistency, auditability, and finality.

The next evolution: Human-in-the-Loop authorization

This is where things become much more interesting.

Imagine an authorization service that doesn’t just produce “Approve” or “Decline.”

Instead, it produces:

Approve (99.8%)

Approve (96%)

Unsure (61%)

Likely Fraud (23%)

High Risk (7%)

Now the architecture changes.

                AI
                 │
          Confidence Score
                 │
      ┌──────────┴──────────┐
      │                     │
High confidence      Low confidence
      │                     │
Automatic           Human review
Authorization

This is exactly how many safety-critical AI systems are designed.

Even better…

Instead of only approving or declining, AI can explain why.

Example:

Transaction

€480

Apple Store

London

↓

AI

Reasoning

• Customer usually spends < €150

• Device is new

• Merchant trusted

• Geolocation differs from home

• Similar purchases occurred before holidays

Risk = Medium

Now the fraud analyst isn’t starting from zero.

They review AI’s reasoning.

That’s Human-in-the-Loop.

Then comes Agentic Commerce

This is where Visa and Mastercard are investing heavily.

Instead of:

Human

↓

Clicks Buy

↓

Payment

we move toward:

Human

↓

"Find me the cheapest flight."

↓

AI searches

↓

AI compares

↓

AI books

↓

AI requests authorization

↓

Issuer

↓

Payment

The challenge is no longer just “Is this card valid?” but “Is this AI agent authorized to act on behalf of the cardholder?” That shifts authorization from authenticating a person to verifying delegated intent and permissions. Visa and Mastercard are building frameworks around explicit mandates, scoped permissions, spending limits, and cryptographic proof of authorization for AI agents.

I think the real innovation is elsewhere

This is the area I would personally research.

Imagine replacing static fraud rules with an adaptive policy engine.

Instead of:

IF amount > 1000

AND country != home

DECLINE

you express intent:

AI

This customer has

95%

probability of approving

because

• salary arrived yesterday

• annual vacation

• merchant trusted

• hotel reservation exists

Approve

This starts to look much more like a Staff Engineer problem than a machine learning problem.

Where I think you could make a real contribution

Knowing the kinds of topics you enjoy—state machines, idempotency, payment orchestration, EMV, authorization, and distributed systems—I wouldn’t focus on training better fraud models.

I’d focus on the architecture that safely integrates AI into payment authorization.

A possible architecture could look like this:

Authorization Request
        │
        ▼
Context Aggregator
        │
        ├─ Transaction history
        ├─ Merchant profile
        ├─ Device attestation
        ├─ Geolocation
        ├─ EMV data
        ├─ Customer profile
        │
        ▼
AI Risk Engine
        │
        ▼
Policy Engine
        │
 ┌──────┼────────┐
 │      │        │
Approve Review  Decline
 │       │
 │   Human Analyst
 │       │
 └───────┘
        │
        ▼
Issuer Authorization

Notice that AI never directly authorizes the payment. It provides a recommendation and supporting evidence, while the policy engine enforces deterministic business rules, regulatory constraints, and escalation thresholds. That separation between probabilistic intelligence and deterministic execution is increasingly reflected in industry thinking because payment systems demand explainability, auditability, and predictable outcomes.

I actually think this could become an outstanding technical series for CoreBaseIT. You could build it from first principles:

  1. Why rules-based authorization is reaching its limits.
  2. Designing an AI-assisted authorization engine.
  3. Human-in-the-Loop payment authorization.
  4. Explainable AI for issuer decisioning.
  5. Policy Engines: separating AI recommendations from payment decisions.
  6. Agentic Commerce: when AI becomes the cardholder’s delegate.

That sequence combines your payments expertise with emerging AI architecture and is the kind of content that’s still relatively rare in the payments engineering community.