Solution Pattern · AQL «عقل»
How an operational system — an ERP, a CRM, a case-management or billing platform — grows into an intelligent decision platform: a conversational interface in plain language, live dashboards on real-time data, and predictive models, all layered on top of the existing system rather than replacing it.
A three-layer design that upgrades any operational system into a decision platform — without rebuilding it. Each layer is independent, sellable, and adoptable on its own: conversational access, live analytics, and (when the data is ready) prediction.
Operational systems are excellent at recording transactions and poor at answering questions. Executives want real-time insight on their screens, analysts spend their time preparing data instead of analysing it, and global platforms now ship AI by default — raising what users expect from every system.
The system stops being a place where data goes in and starts being a place where decisions come out — and its owner gains a tiered offering that raises contract value per customer and creates recurring revenue, using the investment already made instead of replacing it.
01 — The problem
Most organisations run at least one system that faithfully records everything they do — orders, invoices, cases, enrolments, movements. The recording works. The deciding doesn’t: the people who need answers can’t query the system themselves, the reports that exist describe last week, and every non-standard question becomes a ticket in someone’s queue.
Meanwhile expectations have moved. Users who talk to AI assistants every evening no longer accept clicking through five screens for a number at work. And the large global platforms have started bundling AI layers into their products by default — which resets what buyers expect from every system in the market, including yours. The pressure lands hardest on locally-built and legacy platforms, whose owners face an ugly choice: rebuild from scratch (slow, expensive, risky) or watch the product age.
This pattern is the third option: keep the system, add the intelligence.
02 — The design
The layers are independent and complementary. You can start with any of them and add the others as the organisation matures — no big-bang project required.
A sales manager, finance director, or CEO asks in natural language — Arabic or English — and gets a structured answer immediately: the figures, the comparison, the context. “How did the northern branch sell last week against the same period last month, and what were the top three product categories?” No new interfaces to learn, no waiting on IT.
Role-based dashboards on real-time data: sales, margins, inventory, liquidity, customer satisfaction. Decisions ride on the state of the moment, not on last week’s email attachment.
Machine-learning models move the conversation from “what happened?” to “what will happen?”: which customers are likely to leave, which stock lines will run out, which opportunities are most likely to close. Deliberately sequenced last — it needs the historical data quality the first two layers create, and an organisation ready to act on predictions.
03 — Architecture
The existing system stays exactly where it is. The layers connect through its APIs — or a secure read path to its database — and inherit its permission model.
04 — Under the hood
A purpose-built MCP (Model Context Protocol) server sits between the language model and your data — connecting through your APIs where they exist, or a secure database path where they don’t. Users see only what their system role permits. Model choice stays flexible: hosted frontier models, or open-source models running entirely inside the client’s environment for high-confidentiality settings.
A six-part reference architecture: scheduled or change-data-capture extraction; a lakehouse or warehouse in three zones (raw → clean → ready); a central semantic model where business logic is defined once; governance and security with lineage and sensitive-data discovery; and role-based dashboards with mobile support. Designed to run on whichever cloud matches the client’s existing investment.
A full machine-learning lifecycle — training, deployment, monitoring, retraining — built only after the first two layers have matured the data quality and the organisation’s habit of acting on evidence. Typical first models: customer churn, inventory demand, fraud and anomaly detection.
05 — Sequencing
The practical recommendation: start with the conversational layer as a fast, visible win, prepare the analytics layer in parallel, and put prediction on the roadmap for when the historical data — and the organisation — are ready for it.
06 — Delivery
Workshops with the technical and functional teams; assessment of the current architecture, data sources, and data quality; use-cases prioritised by value and feasibility. Deliverables: assessment report and a proposed roadmap.
Architecture design for the full solution and a working proof of concept for the first use-case, reviewed jointly with the client’s team. Deliverables: the PoC and technical design documents.
Component development to the approved design; unit, integration, performance, and security testing; full technical documentation including architecture decision records. Deliverables: the solution ready in a test environment, fully documented.
Production rollout to plan, training for technical and functional teams, and a gradual handover of ownership with accompanied support. Deliverables: a fully operational solution and a team qualified to maintain and extend it.
07 — The payoff
Clear differentiation against local competitors and a credible answer to global platforms with built-in AI; higher contract value per customer through the tiered path; recurring revenue from value-added services; faster time to market than building the expertise in-house; and lower risk through staged, tested delivery.
Immediate answers through natural language instead of report queues; lower analysis cost because routine questions no longer need a specialist; visibility from every branch up to headquarters; and a modern experience that matches what a new generation of users already expects.
Keep the system that already runs your operations — and layer onto it the three things it was never built to do: converse, visualise, and predict. The system’s owner gets a competitive product with a tiered growth path; its users get answers at the speed of the question.