Case study

AI Approvel. A decision with a receipt.

AI Approvel puts a person between an AI agent and a consequential action. We built the AWS platform behind that exchange: a request, a binding human decision, and a signed record that the customer can verify independently.

Practice AWS Cloud · DevOps Engagement Platform build · AI governance · mobile apps
Designed capacity
1M
Approvals per month in the platform sizing model; not a reported traffic figure.
Approver apps
2
iOS and Android, alongside WhatsApp and SMS delivery.
Proof of concept
6 weeks
Concurrency, signing, summaries and push delivery tested before the full build.
Availability target
99.9%
Monthly API availability objective, not a measured uptime result.
The brief

Sending a message does not prove approval.

The prototype had established demand, but enterprise buyers needed more control over isolation, recovery and the evidence behind each decision. Duplicate replies, retries and late messages could not be allowed to change an outcome.

Approvers also needed context. A raw payload on a phone is difficult to judge, especially when the action involves a payment or a customer record. The platform had to make requests understandable while keeping the decision with the person.

What we did

Make the decision binding and the evidence portable.

Keep the first valid reply binding

Transactional constraints and duplicate protection resolve concurrent replies into one decision. Requests expire on a schedule, and retries cannot reopen an answer that has already been recorded.

Sign the sequence of events

Ordered events form a hash chain with Ed25519 signatures. Exportable evidence bundles let customers verify the trail offline, with immutable storage for retained records.

Use AI to explain the request

Amazon Bedrock produces a short summary and risk flags for the approver. It supplies context; the approve or reject action remains human, including in the mobile apps.

Build for a small operating team

Managed containers on ECS Fargate, Aurora PostgreSQL and queued delivery separate the API from background work. Terraform, deployment pipelines and operational runbooks give the team a repeatable way to run the platform.

The outcome

Human oversight with an auditable record.

The delivered platform brings the approval workflow, signed evidence, mobile apps and multi-channel notifications into one AWS environment. A six-week proof of concept exercised the concurrency and signing paths before the production build.

The design targets one million approvals a month, 99.9% monthly API availability and a notification latency below five seconds at the 95th percentile, excluding carrier delay. These are capacity and service objectives, not measured production results.

The brief

The person you brief is the person who solves it.

One conversation · 30 minutes · no monologue