An agent that can complete a purchase can also be manipulated into paying for something the user never asked for. PayWall.ai authorizes every agent-initiated payment against the user's intent, policy, and risk — before it ever reaches Razorpay.
Guard Decision
Payment request
Nike Running Shoes
₹4,799
Matches user intent, within budget, low-risk merchant.
The problem
Traditional fraud detection
“Is this transaction fraudulent?”
PayWall.ai
“Is this transaction authorized, consistent with the user's intent, within policy, and safe to execute?”
How it decides
Layer 1
Deterministic checks — agent status, merchant blocklist, transaction and daily limits, category authorization. Any violation blocks outright, no ML or LLM required.
Layer 2
An XGBoost classifier scores every transaction 0–100 on amount deviation, merchant risk, velocity, and budget headroom.
Layer 3
Compares what's being paid for against what the user actually asked for. Falls back to a deterministic scorer and fails closed to REVIEW if the LLM is unavailable — never guesses.
Revenue growth, not just safety
The Upsell Agent proposes one relevant, budget-aware add-on after a purchase — and it has no special authority. Its proposal is evaluated by the exact same Guard as any other payment.
Legitimate upsell
Running socks, ₹149 — offered after a shoe purchase
Same category, within remaining budget, relevant to intent
Manipulated “upsell”
Protection plan, ₹14,999 — disguised as a checkout add-on
Unrelated category, exceeds authorization, low intent match
Architecture
User states intent
“Buy running shoes under ₹5,000”
Agent proposes a payment
POST /transactions/evaluate
Guard decides
Policy → Risk → Intent, combined
Razorpay executes
Only after ALLOW, or REVIEW + human approval