Meet Our Customer
Our customer is one of Saudi Arabia's largest and longest-established grocery retail chains, operating a nationwide network of stores and serving millions of customers across the Kingdom.
As the business grew, so did its data. Every day its stores generate an enormous volume of operational detail — sales, inventory, stock movement, promotions and procurement — flowing from SAP S/4HANA into an Amazon Redshift data warehouse that now holds more than 600 million transaction line items and 125 million baskets, growing by over half a million line items each day. The company had built a genuine enterprise data asset. What it wanted next was to put that asset directly into the hands of the people running the business, in the language they think in, at the moment they need it.
The Challenge
The data existed. Access to it did not.
Store managers, category buyers and executives depended on BI dashboards and analyst queues for even routine questions — how a store performed yesterday, which products were running low, whether a promotion moved volume. Answers took hours or days, arrived in English only, and required desk access. A store manager and a member of the executive team waited in the same report queue.
At the same time, the highest-volume operational processes remained manual. Daily purchase-order creation for fresh and non-fresh goods, demand forecasting, competitor price tracking and customer win-back campaigns were spreadsheet-driven and dependent on a small group of specialists — work that consumed skilled hours without scaling.
The customer set NorthBay a deliberately demanding brief: a single, trustworthy assistant that employees could simply talk to. It had to understand retail language in both Arabic and English, answer from governed enterprise data rather than guesswork, respect each employee's role and permissions, and safely take action — not just report on it. It had to serve every persona in the business, from the store floor to the boardroom, run inside the tools employees already use, and scale from a first release to the entire organisation.
The Vision
NorthBay designed an agentic AI platform running on Amazon Bedrock inside the customer's own AWS environment, so enterprise data stays under the customer's governance and every persona receives a different, role-appropriate answer from the same trusted foundation.
A supervisor-and-specialist agent architecture
Rather than a single monolithic bot, NorthBay built a multi-agent system on Amazon ECS Fargate in which a supervisor agent interprets each request and routes it to one of nineteen specialists — sales analytics, inventory, stock movement, sales budgets, product catalogue, competitor pricing, predictive forecasting, purchase-order creation and inquiry, HR knowledge, IT incidents, business events, weather and scheduled reporting among them. Each specialist pairs model reasoning with governed tools: text-to-SQL over the Redshift warehouse, retrieval-augmented generation over HR and policy documents on Amazon Aurora Serverless v2 with pgvector, ServiceNow ticketing, SAP purchase-order APIs, and Airflow job scheduling on Amazon MWAA. Agents either complete their task or escalate back to the supervisor, so a conversation can move across domains without the employee needing to know which system holds the answer.
Reasoning that holds up against a real enterprise schema
The hard part of retail analytics is not generating SQL — it is generating correct SQL against a warehouse that encodes years of business rules. Anthropic's Claude models on Amazon Bedrock translate ambiguous, bilingual retail questions into precise queries across 130+ tables, respecting posting versus transaction dates, fresh versus non-fresh classification and store hierarchies. Vision capability reads product photos, shelf images and competitor brochures; voice queries flow through Amazon Transcribe into the same agents. One assistant detects and responds in each user's language, so Arabic and English speakers share a single platform rather than two.
Governed by design
Amazon Bedrock Guardrails, role-based access control enforced at the agent layer and prompt-injection hardening keep responses grounded and workplace-appropriate. The platform is deployed with infrastructure-as-code, CI/CD with environment promotion and explicit production-deploy authorisation, Amazon SageMaker AI pipelines for demand forecasting with SHAP explainability, and event-driven integration to SAP and messaging channels. Security follows an enterprise programme of IAM least-privilege, KMS encryption, managed secrets and OIDC-only pipelines, with PDPL-aligned data handling and full auditability of agent actions.
Outcome
The platform has been in production since 2025, delivered through four major releases and two funded extension phases, with a third in planning — sustained commitment rather than a one-off pilot. It runs across development, QA and production environments with formal release management, a 12-week software-quality-assurance cycle and hotfix-governed production freeze discipline. It is delivered inside Microsoft Teams, with an embedded web chat option, so adoption required no new application: employees message the assistant like a colleague.
19 agents
Nineteen specialists and a supervisor, live in production across analytics, procurement, HR, IT and more
Seconds, not days
Around 30 seconds for document look-ups and 35–45 seconds for live analytics, against hours or days through reporting queues
600M+ records
Transaction line items across 130+ warehouse tables, reachable in plain language in Arabic or English
200+ monthly users
Monthly active employees across stores and head office, growing as onboarding continues
08:30 daily
SAP-integrated purchase-order creation covering every store in each day's forecast, each order carrying an explainable rationale
6 of 7 functions
Business functions served in production today, with customer marketing next on the roadmap
The change is as much economic as technical:
- Self-service analytics. Questions that previously entered a reporting queue are answered conversationally, removing a dependency that constrained decisions at the point of operation.
- Procurement automation. Purchase orders are created automatically every morning at 08:30 KSA, integrated end to end with SAP, replacing a manual spreadsheet process and returning buyer hours to negotiation and assortment strategy.
- Forecasting with accountability. SageMaker demand forecasts with SHAP explainability inform ordering, with drivers — events such as Ramadan and Eid, weather, promotions — visible to the business rather than buried in a model.
- Enterprise data at conversational reach. Records that were the preserve of data specialists are now open to any authorised employee, under the same governance.
- Measurable internal communication. Organisation-wide announcements carry per-recipient delivery tracking, on a pipeline validated at 1,000+ recipients.
Value across the business
Because the supervisor-agent architecture separates routing from expertise, the same platform serves each function differently — without separate tools or integrations to manage. Six of seven functions are live in production today.
| Function | Status | Before | With the assistant |
|---|---|---|---|
| Store Operations | Live | Waited hours or days on BI and analyst queues for routine performance and stock questions | Inventory, stock movement, vendor delivery, weather and forecasting answered in natural language, in Arabic or English, from within Teams |
| Merchandising & Procurement | Live | Manual, spreadsheet-driven purchase orders; slow competitor price tracking | Automated daily purchase-order creation and inquiry with explainable rationale, plus competitor price tracking across major KSA retailers |
| Finance & Leadership | Live | Limited, delayed visibility into demand drivers and enterprise performance | Self-serve sales and budget-versus-actual reporting, and SHAP-explainable forecasts tied to real business drivers |
| HR | Live | LivePolicy questions handled manually, one employee at a time | Bilingual HR knowledge available on demand, grounded in approved documents, freeing the team for higher-value people work |
| IT Service Desk | Live | Incidents logged and triaged manually | Employees raise ServiceNow tickets directly from chat, speeding capture and routing |
| Corporate Communications | Live | Broadcasting to a distributed workforce was slow and untracked | Announcements and scheduled email reports across Teams and web chat, with per-recipient delivery tracking |
| Customer Marketing | Roadmap | Dormant customers and lapsed products went unaddressed | Consent-managed, bilingual win-back campaigns over WhatsApp and email, matched to live promotions |
Work now underway extends the same governed foundation into wider competitor price intelligence, deeper purchase-order history analytics, and personalised customer re-engagement. Each increment reuses the existing agent architecture, shortening time-to-value with every release.
AWS Services
The project heavily utilized AWS services to ensure scalability and efficiency:
Why NorthBay?
From the outset this was built for daily operational use across an entire workforce — which meant release management, security review, bilingual quality assurance, and the discipline to ship one use case, prove it in production, and expand from there. NorthBay is an AWS Premier Consulting Partner with an AI Consulting Services Competency and deep delivery experience in the region, and applied the same engineering rigour to an agentic system that it would to any enterprise platform: governed data access, explainable automation, and a production operating model from the first release.
