Mercado: commerce infrastructure with AI reconciliation, and a person on every ledger change
Morphlix built Mercado's commerce platform on Amazon Web Services (AWS), with an Amazon Bedrock reconciliation agent that cut reconciliation of a standard batch from about two business days to under four hours. Ledger changes are confirmed by a finance reviewer.
- Commerce
- Agentic AI on AWS
- Studio build
- Industry
- Commerce Infrastructure
- Engagement
- Studio build
Results
< 4 hrs
per standard reconciliation batch
Down from about two business days, about 85% faster.
90% → 98%
partial-payment handling accuracy
After a targeted matching improvement.
88%
evaluation score
8%
finance override rate
With an 18-minute average approval time.
18 min
average approval time
3.0 → 1.2
notifications per batch
60% fewer, with small discrepancies logged without notifying anyone.
19
operational modules in one merchant console
121
console routes
44
granular permission actions
The challenge
Reconciliation by hand caps growth
Across African markets, commerce often starts on WhatsApp, with stock tracked in a notebook or spreadsheet and payments taken through several providers. As a business adds products, staff, channels and branches, stock counts drift, chat orders can miss the books, and payments can go unmatched against sales.
Three systems, cross-referenced by hand
For Mercado's merchants, reconciliation meant a finance analyst cross-referencing orders, payment settlements and stock movements across three systems by hand.
Two business days per batch
A standard batch took about two business days.
Hard cases need a senior analyst
Many discrepancies, such as partial payments, split settlements and duplicate webhooks, needed a senior analyst.
Without automation, merchants' growth would be capped by the size of their finance teams.
The solution
One platform for products, stock, orders, customers, payments and finance
Conversational ordering connected to orders, inventory and payments.
Inventory across locations, with adjustments, transfers and stock alerts.
Order tracking, returns, split payments, customer segments, wholesale and loyalty.
A double-entry accounting foundation.
APIs, integrations and role-based permissions across 44 granular actions.
Starter, Growth, Scale and Enterprise, including running in the customer's own AWS account.
The AI layer
AI reconciliation, with a person on every ledger change
Deterministic first
An AWS Step Functions pipeline fetches orders, settlements and stock movements, matches them and computes each discrepancy in code, so the arithmetic is testable and repeatable.
AI explains
Amazon Bedrock (Anthropic Claude) explains each discrepancy and proposes the correcting entry. Amazon Bedrock Guardrails help ground each explanation in the computed facts.
People approve
Agent roles have no permission to write to the ledger. Each proposal waits in an approvals queue for a finance reviewer, and two reviewers are required above the materiality threshold.
Less noise
A materiality band logs small discrepancies without notifying anyone, which reduced notifications per batch from 3.0 to 1.2.
Also live
A smaller Amazon Bedrock model triages payment events for review, and a read-only merchant assistant answers questions on each merchant's own sales, stock and orders.
Outcome
People stay in control without slowing the books
About 85% faster reconciliation: under four hours per standard batch, down from about two business days., Partial-payment accuracy up from 90% to 98% after a targeted matching improvement., 8% finance override rate and an 18-minute average approval time. and 60% fewer notifications per batch.
Architecture
Commerce and AI reconciliation on AWS
The infrastructure is defined in Terraform and released through GitHub Actions with an approved plan. The production network runs across two Availability Zones (10.40.0.0/16).
Application
Amazon ECS on AWS Fargate running 13 domain services, behind an Application Load Balancer, with images in Amazon ECR and AWS Lambda for supporting functions.
Reconciliation and AI
AWS Step Functions runs the reconciliation pipeline; Amazon Bedrock and Amazon Bedrock Guardrails explain discrepancies and propose entries.
Data
Amazon Aurora Serverless with Amazon RDS Proxy (Multi-AZ), Amazon DynamoDB, Amazon OpenSearch Service and Amazon S3.
Edge and security
Amazon Route 53, Amazon CloudFront, AWS WAF and AWS Certificate Manager at the edge, with AWS Secrets Manager, AWS KMS, AWS CloudTrail, Amazon GuardDuty and AWS Security Hub.
Technology stack
AWS services behind Mercado
Edge and delivery
- Amazon Route 53
- Amazon CloudFront
- AWS WAF
- AWS Certificate Manager
- Application Load Balancer
Compute and workflow
- Amazon ECS on AWS Fargate
- Amazon ECR
- AWS Lambda
- AWS Step Functions
AI
- Amazon Bedrock
- Amazon Bedrock Guardrails
Data
- Amazon Aurora Serverless
- Amazon RDS Proxy
- Amazon DynamoDB
- Amazon OpenSearch Service
- Amazon S3
Messaging
- Amazon SQS
- Amazon SNS
- Amazon EventBridge
Security and operations
- AWS Secrets Manager
- AWS KMS
- AWS X-Ray
- Amazon CloudWatch
- AWS CloudTrail
- Amazon GuardDuty
- AWS Security Hub
Tooling
- Terraform
- GitHub Actions
Governance
Audit, 14 October 2026
99.8% resource-tag coverage
432 of 433 resources carried the required tag.
0 infrastructure drift
The audited Terraform plan reported no changes.
8 of 8 target groups healthy
All eight load balancer target groups were healthy at the audit.
More from the library.
All case studiesAI · Productivity
Seshira: AI-native project management
An engineering assistant on Amazon Bedrock AgentCore that plans sprints, turns PRDs into tasks and flags blockers, with about 80% less planning effort.
Agriculture, Engineering & Automotive
Sallis: one digital front door for three divisions
One website and a Morphlix Connect enquiry workflow for an agriculture, engineering and automotive company, with 99.8% availability.
Work with the studio
Ready to future-proof
your infrastructure?
Tell us what you are building and where the current stack is holding you back. We will come back with an architecture and a plan.