Everyone has a demo. Fewer ship.
AI that answers from your live data, respects who may see what, and never acts without approval. Arabic and English, wired into the systems you already run.
The hard part is not the model.
Getting a language model to say something impressive takes an afternoon. Getting it to answer correctly from your database, in front of a regulator, without emailing a customer by accident, is the actual work.
Demo versus production
A demo answers a question. Production answers to an auditor.
Most AI pilots die in the same place: the gap between a convincing conversation and a system anyone will let near real customers, real money and real records.
A model with no grounding will produce a confident number that does not exist in your data.
If the assistant ignores permissions, one question exposes salaries, margins or a competitor's pricing.
An agent that can send is an agent that will send, at some point, to the wrong person.
"The AI said so" survives exactly one conversation with a regulator or a board.
- —Start where being wrong is cheap, not where it is catastrophic
- —A narrow assistant that works beats a broad one that impresses
- —If the data is a mess, AI makes the mess faster, not smaller
- —Some problems are a report, not a model — we will say so
Grounding, so answers come from your records. Permissions, so the assistant sees only what its user may see. Approval gates, so nothing leaves without a human. And an audit trail, so every answer can be traced back to the query behind it.
The distinction that matters
A chatbot answers. An agent does the work.
This is the whole difference, and most vendors blur it. Ask both the same thing and watch what happens next.
- —Chasing receivables with correct invoice references
- —Qualifying inbound enquiries before a human picks them up
- —Extracting supplier documents into records
- —Compiling the answer to a recurring management question
Tools it can call against your real systems, a definition of what it may and may not do, memory of the conversation, and a gate before anything leaves the building. Remove any one of those and you are back to a chatbot with better manners.
Connected, not copy-pasted
It reads your database. Not a copy of it.
- —Natural-language questions over live business data
- —Answers cite the records they came from
- —The assistant runs as a user, with that user's permissions
- —No parallel copy of your data to secure and keep in sync
- —Open standard, so you are not locked to one vendor
Why open standards matter here: the connector is not a proprietary bridge only we can maintain. If you replace us, the wiring stays.
The part imported products get wrong
Your team does not speak in one language.
Real Saudi business conversation switches mid-sentence: an Arabic message with an English product code, a WhatsApp reply that starts in one language and finishes in the other. Systems built elsewhere treat that as an error.
- —Understanding that does not break at the language boundary
- —Replies in the language the customer used, not a default
- —Arabic that reads as written by a person, not translated
- —Product codes, numbers and units preserved exactly
- —Right-to-left handled as engineering, throughout
Where this is decisive: customer-facing automation. An assistant that answers Arabic in stilted, obviously-machine phrasing damages the brand more than having no assistant at all.
Where Saudi business already happens
The channel your customers already answer on.
- —Quotations, confirmations and payment reminders with real references
- —Approve and ask buttons that write back into the record
- —Every message logged against the customer, searchable later
- —Arabic and English templates, pre-approved by Meta
- —An agent that drafts; a person who releases
The quiet cost centre
Somebody is retyping a PDF right now.
Supplier invoices, delivery notes, purchase orders, contracts. They arrive as documents and leave as data — and in most companies a person is the conversion layer.
- —Bad scans stay hard; we set a confidence threshold, not a promise
- —Low-confidence extractions go to a human, not into the ledger
- —Every extraction keeps a link to its source document
- —Start with your highest-volume supplier, not all of them
Not the typing — the checking. A person retyping an invoice is not comparing it to the order. Extraction that also reconciles is what catches the wrong quantity and the price that quietly changed.
What was said, at scale
Nobody listens to a thousand calls.
Sales and support conversations are the richest data a company generates and the least used, because reviewing them costs a manager's week. Transcription and analysis turn them into something readable.
- —Which objections come up repeatedly, and where deals stall
- —Whether the team is actually saying what training said
- —Summaries and next actions written straight into the CRM
- —Sentiment and escalation risk, flagged while it still matters
- —Arabic and English handled in the same pipeline
Recording and analysing conversations is a people question before it is a technical one. We build it with disclosure, retention limits and role-based access from the start, because deployments that surprise staff get quietly sabotaged, and rightly so.
The part that gets it approved
AI proposes. People decide.
- —Read-only, draft-only, or act-with-approval — chosen per action
- —Outbound messages queue for release, never send silently
- —The assistant inherits the permissions of the user asking
- —Every action logged: who asked, what ran, what was returned
- —A kill switch that a non-technical manager can reach
This is also the compliance answer. When a regulator or auditor asks how a document was produced, the trail is the answer — not a description of how the model works.
Where the data goes, and which model sees it
Two questions your board will ask first.
Before capability, most Saudi buyers want to know where their data physically sits and who the vendor behind the model is. Both answers should be a decision, not a default.
- —Cloud regions inside the Kingdom where the platform offers them
- —Regional hosting where in-Kingdom is not available
- —Self-hosted open models where data must never leave your estate
- —Retention and logging policies set per deployment, in writing
- —Frontier models for reasoning-heavy work and Arabic quality
- —Smaller, cheaper models for high-volume classification
- —Open models when residency or cost dictates it
- —The architecture stays the same; models are swappable
We will not tell you a model is "secure" as though that settles it. The questions that matter are where the data rests, who may read it, how long it is kept, and whether it trains anything. Those get written into the engagement, not assumed.
The unglamorous half that pays for itself
Most of what people want is not an AI problem.
A large share of the work we are asked to solve with AI is a plumbing job: a form that should create a record, a file that should reach a folder, a reminder nobody sends. Deterministic automation does it cheaper, faster and without a probability attached.
- —The rule can be written down completely
- —The same input must always give the same output
- —The step is high-volume and low-judgement
- —An error must be impossible, not merely unlikely
- —The input is unstructured — language, documents, speech
- —The rule exists but nobody can fully articulate it
- —The variety is too wide to enumerate
- —A confident draft for a human beats a blank page
- —Onboarding and offboarding checklists that run themselves
- —Approval routing with escalation when nobody responds
- —Document filing, naming and retention
- —Reminders that reference the actual outstanding item
Make.com for service-to-service flows, Power Automate inside Microsoft estates, self-hosted tooling where volume or data residency requires it, and custom code only when no-code genuinely does not fit. The tool is chosen after the problem, not before.
Our own platform, stated plainly
We built the pattern before we sold it.
Rather than describe an architecture in the abstract, we built one: a multi-tenant conversational agent that answers business questions from a live ERP over WhatsApp. It is our platform, not a client deployment, and we would rather say that than imply otherwise.
- —Proves: the architecture runs, end to end, against real ERPs
- —Proves: multi-tenant separation and tool-calling work as described
- —Does not prove: scale under a large customer's load
- —Does not claim: a named client running it in production today
When a backend is not connected, the platform returns clearly flagged sample data and the assistant says so in its reply. It is a small thing that reflects a rule we apply everywhere: a system should never let you mistake a demonstration for the truth.
Tobrecards
A product where the AI is the product.
- ·Generic cards; no personalisation at the point of purchase
- ·Design effort required per card, limiting range
- ·Output needed for both digital delivery and print
- ·An image-generation pipeline producing card artwork
- ·Personalised message generation from the buyer's input
- ·A layout engine composing print-ready and digital output
- ·The surrounding web platform and API
- ·Generative AI inside a real product flow, not a demo
- ·Output quality controlled well enough to print
- ·A pipeline, not a single model call
Generative work is judged by its worst output, not its best. The engineering here is in the constraints — templates, validation and fallbacks — that keep the range acceptable every time, not just in the screenshot.
The unglamorous kind that pays for itself.
- ·Work spread across tools that did not talk to each other
- ·Information re-entered by hand between systems
- ·Knowledge held in individuals rather than a shared base
- ·Automation workflows connecting the existing toolset
- ·A structured workspace as the shared source of record
- ·Cross-tool data synchronisation replacing manual re-entry
- ·Documentation so the client's own team can extend it
- ·Steps that were somebody's routine now run unattended
- ·One place to look rather than three to reconcile
- ·The client can maintain and extend the flows themselves
No model was involved and none was needed. This is the work that quietly returns hours every week, and it is usually where we suggest starting before anything is pointed at a language model.
Two platforms, under agreement.
- ·An intelligent assistant handling customer conversation
- ·Question answering across the company's own knowledge
- ·Automated generation of proposal documents
- ·Session capture to see where users actually struggled
- ·Analysis of recorded calls at a volume no manager could review
- ·Structured scoring of candidate interactions in hiring
- ·Question-and-answer analysis across conversations
- ·These engagements are covered by confidentiality
- ·The website presents them the same way, unnamed
- ·Scope is verifiable in a reference conversation
We would rather show you an unnamed engagement described accurately than a named one described loosely. If a supplier's case studies never carry a constraint, ask what else has been smoothed over.
Small, reversible, measured
Start where being wrong is cheap.
The failed AI projects we are asked to rescue share a shape: too broad, too central, and impossible to evaluate. We deliberately start at the edge and move inward.
A task that is repetitive, language-shaped, and where a wrong answer is caught by a human before it costs anything.
Before building, we write down what "working" means and how it will be judged. If we cannot define that, we do not start.
Real users, real data, drafting only. They keep the judgement; we watch where it fails and fix what actually breaks.
Either the measure moved and we extend, or it did not and we say so. Both are acceptable outcomes; only pretending is not.
AI has a per-use cost that traditional software does not, and it is the part most proposals leave vague. We estimate it per conversation or per document before you commit, and design to keep it predictable — smaller models for volume, larger ones only where judgement is needed.
- —Where does my data rest, and does it train anything?
- —What happens when the model is confidently wrong?
- —Can it act on its own, and who can stop it?
- —What does this cost per use at our volume?
- —What do we own if we replace you?
The next step