An AI agent that can move money, and prove every decision.

• Six real actions: refunds, cancellations and address changes for the retailer; EMI date changes, card blocks and card disputes for the lender.
• Four gates in a fixed order (identity, eligibility, penalty, logic) return exactly one of three routes: execute now, hold for a person, or deny citing the rule.
• Rules are rows in a policy table, not sentences in a prompt: a 7-day return window, a ₹5,000 auto-refund limit, a 15-day EMI shift, 30 days past due.
• Card disputes follow the RBI customer-liability framework (RBI/2017-18/15): reported within 3 working days means zero liability, later reports are capped and go to a person, and the shadow-credit and 90-day decision dates are computed per case.

• Execute or approve anything. It can only propose; an executed action exists only if the gateway returned “execute” or a person approved it.
• Pick which record an action touches. Order, account and transaction ids are resolved on the server from that customer’s own data, so another customer’s order comes back “not found”.
• See what it doesn’t need. Phone numbers are masked, the one-time code never enters the prompt, and for the lender, balances and transactions are withheld until identity is verified.
• Write to memory or email anyone it chooses. Customer memory is curated; email goes only to the address on file, and in this sandbox to a simulated inbox.

• The gate trace on every action: each rule, the policy value and the value it found.
• The evidence ledger: every message, proposal, decision and bill line is chained with HMAC-SHA256 inside the database. “Verify chain” re-walks it and names the first event that fails.
• The bill: each line points at the ledger entry that earned it, a failed or denied action costs nothing, and a disputed line is credited.
• The regression suite: 18 governance scenarios run against the live policy table nightly and again as this page loads, including 23 hours on, so no promise on this page decays.
• Act as the operator: approve or reject the held actions in your own sandbox, in the chat or the console.

• Connectors. Tools write to sandbox tables that stand in for an order system, a loan book and a card processor; in production each becomes an API or MCP connector behind the same gateway.
• Voice and WhatsApp. One web channel; the gateway and ledger are channel-agnostic.
• Policy authoring and versioning. Policies are rows today; a what-if simulator that replays past actions against a draft policy is the next step.
• Key custody. The ledger key sits inside a database function, so the chain is tamper-evident to anyone without database-owner access; production keeps it in a KMS and anchors each day’s chain head outside the database.
• Bank holidays in working-day maths, and certifications (SOC 2, ISO 42001).

$1.00 per completed action and $0.50 per resolution the customer confirms, capped per tenant per month; anything past the cap is recorded as capped, not billed. The rate card mirrors Azeon’s published pricing on azeon.ai. The difference studied here is what makes the meter trustworthy: “resolved” is the customer’s call, and every charge is backed by a ledger entry.

Azeon is Azilen Technologies’ agentic customer-service product (azeon.ai). This is an independent study, built by Kaushal Khodifad, of how a platform like it can make autonomous actions provable. It is not affiliated with, endorsed by or commissioned by Azilen, and it does not describe their implementation. Customers, tenants and data are synthetic.