- Market
- Saudi Arabia
- Written for
- CIO, CTO, Head of Digital
- Decision
- Move an AI use case from pilot into an operational system
Where programmes stall
Pilots are built on curated data and judged on best-case output
We do not have an office in Saudi Arabia
We serve organisations there from our Riyadh office and our engineering hubs in Lahore and Karachi, travelling for the parts of an engagement that need to happen in person.
A demonstration runs against documents someone cleaned, questions someone chose, and a reviewer who knows what the system is good at. A production system faces the data the organisation actually has: three versions of the same policy, scanned pages, an export where the columns shifted in 2019, and a user asking something nobody anticipated.
What separates the two is unglamorous and mostly not about the model: where the data comes from and how fresh it is, what happens when the model is wrong or the provider is down, how much a request costs at real volume, who reviews outputs, and how you tell whether quality has drifted since launch. Those questions are the engagement. The model choice is usually the least contested decision in the room.
In Saudi Arabia there is a further constraint that shapes the architecture from the start: where the data is allowed to sit. Whether a hosted model provider can be used at all, and for which data classes, is a decision to make before design rather than during procurement. Where the answer is that nothing may leave, we design for self-hosted models and say plainly what that costs in capability — it is a real trade-off, not a footnote.
What we build
Six systems, not a technology list
Retrieval-grounded knowledge systems
Assistants that answer from your own documents and cite the source, so an answer can be checked rather than trusted. The engineering is mostly in chunking, freshness and permissions — a retrieval system that shows a user a document they should not see is worse than no system.
Arabic and bilingual conversation
Chat and voice that handle dialect variation, code-switching mid-sentence, and Arabic names and addresses without corrupting the record. We build the evaluation set in Arabic rather than translating an English one, because a translated test set measures the wrong thing.
Document and workflow processing
Extracting structured data from invoices, forms, contracts and reports and routing it into the system that needs it — with a human review step wherever the cost of an error justifies one. Usually the highest-return first project, because the task is bounded and the failure mode is visible.
Agents that take actions
Systems that do something in another system — create a ticket, update a record, book a slot. The work is the permission model, the audit trail and the stop conditions. An agent that can act without a reversal path is a liability, not a capability.
Forecasting on operating history
Demand, maintenance and risk models built on what actually happened in your operation, evaluated against outcomes rather than a held-out sample alone. We report accuracy honestly, including where a simpler method would have done as well.
Evaluation and monitoring
A test set reflecting real inputs including the awkward ones, quality measurement that continues after launch, and cost and latency tracked per request. Without this you cannot tell a good week from a degraded model.
Production readiness
What a pilot assumes, and what production has to answer
This table is the shortest honest summary of why the second half of an AI programme costs more than the first.
Data
A clean, curated extract
What production needs: A pipeline with a defined refresh, handling for the messy cases, and permissions that follow the source
Failure
A person was watching
What production needs: A defined fallback path, a confidence threshold that routes to a human, and a degraded mode when the provider is down
Cost
A few hundred requests
What production needs: Per-request cost modelled at expected volume, with caching and routing so economics stay visible before the first invoice
Quality
It looked right in the demo
What production needs: An evaluation set including the hard cases, measured continuously so drift is caught before users report it
Accountability
The project team owned it
What production needs: A named owner, an audit trail of what the system did, and a documented basis for decisions that affect a person
Residency
A hosted API was fine for testing
What production needs: A decision per data class on what may leave the Kingdom, and an architecture that enforces it rather than documents it
Integration surface
What these systems connect to
An AI system that cannot reach the systems of record is a demonstration. Most of the engineering effort is here.
Systems of record
- ERP
- CRM
- Core banking
- Ticketing
- HR systems
Channels
- Web
- Mobile
- Telephony
Data
- PostgreSQL
- Data warehouse
- Document stores
- Vector search
Platform
- AWS
- Microsoft Azure
- Kubernetes
- Self-hosted inference
Common questions
What buyers ask
Will our data be used to train someone else's model?+
Not without your explicit agreement. Which providers are used, under what terms, and whether data leaves your environment at all are decisions we put in front of you at design time rather than in a contract annex. Where the requirement is that nothing leaves, we design for self-hosted models and say plainly what that costs in capability.
How do you stop the system giving wrong answers?+
You cannot eliminate it, so the design has to assume it. Ground answers in retrieved source documents with citations, constrain what the system is allowed to answer at all, route low-confidence cases to a person, and measure quality continuously against a test set that includes the awkward inputs. Anyone promising a system that is never wrong is describing a demonstration.
What is a realistic first project?+
Something with a clearly measurable task, data that already exists, and a tolerable failure mode — document extraction with a review step, or an internal knowledge assistant over documentation your staff already search badly. A customer-facing system that takes actions is a better second project than a first one, because by then you have an evaluation harness and a sense of where the model is weak.
Do we need a data platform before we can do anything with AI?+
Not a complete one, but the specific data the use case depends on has to be reliable and reachable. Building an entire platform before any use case is a common way to spend a large budget with nothing to show at the end of year one. Build the domain the use case needs, prove it, then extend.
In this cluster
Next step
Test a use case against your data
A short, fixed-duration test of one use case against your real data, ending with a written answer on whether it is viable and what production would take. One hour to scope, no cost, no commitment.
Test a use case against your data
What happens next
- 1We reply within one business day.
- 2A 30-minute scoping call with a senior consultant.
- 3A written proposal with scope, team, and timeline.
Or reach us directly

