AWS Builder Center
What do you show a bank when the customer who disputed a charge never clicked anything?

What do you show a bank when the customer who disputed a charge never clicked anything?

How I built Recourse, a deterministic evidence system for disputed AI-agent transactions — for the Bharat Builds Tour, and the three real AWS/API failures it took to get there.

In February 2026, Razorpay and NPCI launched an agentic payments pilot on Claude — letting people order from Zomato, Swiggy, and Zepto by conversation, using UPI Reserve Pay. A user pre-authorises a spending cap for a merchant, and from then on, an agent can transact within it without a PIN prompt every time.
It's a genuinely good idea. It also breaks something quietly: when one of those transactions gets disputed, the evidence a merchant's ops team normally reaches for — a device fingerprint, a click trail, a browser session — doesn't exist. No human clicked anything. An agent did.
I built Recourse over the Bharat Builds Tour's First Commit hackathon to answer one question: what replaces that evidence when the buyer wasn't a person?



The chain instead of the click

Recourse replaces the missing click trail with three things: the mandate (what the user actually authorised — a cap, a merchant, a validity window), the agent's own action log (search, select, confirm, pay, each timestamped), and the fulfilment record. Six deterministic checks run against those three: is the amount within cap, does the merchant match, is the mandate still active, was the order actually fulfilled, is the timeline internally consistent, is there no suspicious near-duplicate order under the same mandate.
Three of those checks are Cedar policies, evaluated by Amazon Verified Permissions — not a hand-rolled if-statement pretending to be a policy engine, an actual AWS authorization service. I chose it specifically because Cedar's native language bindings don't have a clean Python story, and Python is where the rest of the system lives.



The part I'm actually proud of

If all six checks pass, a language model writes a short narrative explaining the recommendation. Here's the thing: that narrative gets checked before a human ever sees it. It's rejected if it mentions anything structurally impossible for an agent-initiated transaction — a device fingerprint, an IP address, a browser session, a human clicking anything — or if it cites a rupee figure that doesn't appear anywhere in the real, structured facts. One generation attempt, no retries. A failure means the dispute escalates to a human, full stop.
My first version of this grounding check didn't work this way. It asked a second model to judge whether the first model had hallucinated — which meant a probabilistic system was quietly deciding whether to trust another probabilistic system, exactly the failure mode the whole architecture exists to prevent. I caught it and rewrote it as plain deterministic code: pattern matching and numeric comparison against structured facts, nothing that touches a network or a model.



What actually broke, in order

I'd love to say the AWS side was smooth. It wasn't, and I think that's worth saying plainly rather than editing out of the story:
Bedrock access got blocked by an account-level eligibility hold — a restriction new AWS accounts without billing history run into, unrelated to which model you're requesting. I filed a support case, tried the (unofficial but reported) fix of triggering a small EC2 instance to verify the account, and confirmed the restriction was account-wide, not model-specific, by testing Amazon's own Nova models and hitting the identical error.
WeMakeDevs confirmed Bedrock isn't actually mandatory for eligibility, so I pivoted to calling an LLM API directly instead. First attempt hit a billing wall. Second attempt got blocked by a Cloudflare firewall rule — Python's default HTTP client sends a User-Agent string that reads as non-browser traffic and gets flagged. Fixed that, then hit a deprecated model name that had been retired for free-tier accounts three months earlier. Each of these is logged, with the real error text, in Recourse's own audit trail — the same append-only log the architecture uses for actual disputes now has a record of its own build process failing and recovering, which felt like a fitting place for it to live.



The numbers

Precision and recall on 20 held-out synthetic disputes both come out to 1.0. Worth being honest about what that does and doesn't mean: this is a deterministic rule engine, not a trained model generalising to unseen data, evaluated against synthetic cases with unambiguous ground truth. A correctly implemented deterministic system scoring close to perfect on clean-cut cases is expected behaviour, not a claim that it's infallible on messier real-world data.



What I'd build next

Real mandate revocation mid-transaction. Partial refunds instead of binary contest/accept. Confidence calibration against genuinely ambiguous cases, not just clean synthetic ones. And eventually, the actual Bedrock integration this was originally meant to use — the account hold will clear eventually, and the whole system was built provider-agnostic specifically so that swap is a config change, not a rewrite.




Built for Bharat Builds Tour — First Commit, WeMakeDevs × AWS Builder Center.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article