Case Studies

AI Agents for Site Reliability in a Payments Platform, On-Premises

Written by Dr. Jagreet Kaur | 11 Nov 2026

Client

Large payments company, India

Industry

Financial services, payments

Product

AgentSRE on ElixirClaw

Engagement

9-week proof of concept

In brief

  • The reliability team wanted to know whether AI agents could detect anomalies, correlate incidents and draft root-cause analyses on its own infrastructure, without sending logs to a public AI service.
  • XenonStack deployed AgentSRE on-premises, connected to the client's existing Prometheus and Loki, using language models through private network endpoints only.
  • The agents covered the full path from signal to ticket: observability, anomaly detection, correlation, root-cause analysis, remediation suggestions and Jira updates.

The challenge

A payments platform runs many services that customers depend on every minute, including consumer-credit applications. When something degrades, engineers piece together logs and metrics from several tools, decide which alerts belong to the same incident, and write the root-cause analysis by hand. That takes time, and the quality depends on who is on call.

AI agents could take on much of that work, but a regulated payments company cannot send production logs to an outside AI service or let an agent act without controls. The test had to run inside the client's environment, on its existing monitoring stack, with every step reviewable.

What we built

1. AgentSRE, deployed on-premises

The agent stack ran on the client's own Kubernetes environment, next to its user-acceptance-testing systems, not in a XenonStack cloud.

2. Connected to the monitoring the team already uses

AgentSRE read logs from Loki and metrics from Prometheus, so engineers needed no new tool and the agents worked from signals the team already trusts.

3. Private access to language models

The agents used large language models on Amazon Bedrock through private network endpoints only, so log data never crossed the public internet.

4. A chain of specialised agents

  • Observability agent: answers questions about logs and metrics in plain language.
  • Anomaly agent: spots unusual behaviour, with thresholds tuned during the engagement to cut noise.
  • Correlation agent: groups related alerts and signals into one incident.
  • Root-cause agent: drafts the analysis and suggests remediation for an engineer to approve.

5. Built into the incident workflow

Findings flowed into Jira as ticket and incident updates, so the agents worked inside the team's existing process rather than beside it.

How we delivered

Weeks 1–2

Onboarding, least-privilege access with role-based permissions, validation of monitoring data

Weeks 3–4

Core agent deployment, secure ingestion pipelines, model integration

Weeks 5–6

Observability and anomaly detection, with threshold tuning

Weeks 7–8

Correlation and root-cause analysis, validated end to end from anomaly to root cause

Week 9

Demonstration, feedback, sample root-cause and remediation reports, final proof-of-concept report

Security was agreed up front: encryption in transit and at rest, multi-factor sign-in, data classification for logs and metrics, and tracked approvals for every access request.

Why it matters

Most organisations can see the value of AI agents in operations, but few can test them without moving sensitive data or adding risk. Running agents on-premises, on the existing monitoring stack, with private model access and a human approving each fix, lets a regulated business find out what agents can do for reliability before it commits.

Read more about ElixirClaw, how the stack fits together on the XenonStack ecosystem page, and AgentSRE. To see what happens before you book, read how the Context Workshop works.
View full post