Customer service agent
Service in Arabic and English, with a person one step away.
The customer service agent answers customers on web chat, WhatsApp and, where it makes sense, the phone — in Saudi dialect, Modern Standard Arabic or English. It answers only from content you have approved, looks up order and case status in your systems, and hands over to a person with the full conversation when it should. Your service team keeps the conversations that need judgement.
The problem
What customers and teams run into
- 01Routine questions — where is my order, when are you open, what is the status of my request — take agents’ time away from the cases that need it.
- 02Customers write in Saudi dialect, MSA and English, often in the same message, and scripted bots miss the meaning.
- 03Generic chat tools answer from what the model thinks it knows, not from your policies, and can invent terms you never offered.
- 04When a bot gives up, the customer has to start again from the beginning with a person.
How it is built
How the agent is built
The model does not work alone. It answers from an approved knowledge base, calls a small set of tools to look up status, and follows hand-over rules you control. Everything it says and does is logged for quality review.
Layer 1
Customers and channels
- Web and in-app chat
- WhatsApp Business Platform
- Voice, optional
- Service team inbox
Layer 2
Conversation and control
- Dialect-aware understanding
- Answers from approved content
- Guardrails and topic limits
- Hand-over rules
Layer 3
Tools and knowledge
- Order and case lookups
- Customer verification step
- Versioned knowledge base
- Conversation records
Layer 4
Infrastructure
- Private model serving
- Speech recognition and synthesis
- Integration layer to CRM and OMS
- Quality dashboards and monitoring
Capabilities
Customer service agent
Answers only from approved content
The agent draws on FAQs, policies and product information your team has approved and versioned. If the answer is not there, it says so and offers a person, rather than improvising terms or prices.
Saudi dialects, MSA and English
The agent is tested on real transcripts in Najdi, Hijazi and Gulf phrasing, Modern Standard Arabic and English, including switches between them mid-message. It replies in the customer’s language, in a register your brand approves.
Lookups, within limits
Tools let the agent read order status from your OMS or case status from your CRM once the customer is verified. Tools are read-only by default; refunds, cancellations and other actions need a person or an explicit approval rule.
Hand-over with full context
When the customer asks for a person, the topic is sensitive or the agent is unsure, the conversation moves to your team’s inbox with the transcript, detected intent and lookups already done. The customer does not have to repeat themselves.
Quality monitoring
Conversations are sampled and scored for accuracy, tone and correct hand-over, with dashboards for supervisors. Failed answers become test cases, so the same mistake is checked for on every future change.
Red-teaming before and after launch
We test the agent against attempts to extract other customers’ data, override its instructions or make it promise what you do not offer — in Arabic and English. The tests repeat whenever the model or the content changes.
How it works
A conversation, step by step
- Customer writes or calls
- Intent and dialect understood
- Answer from approved content
- Status looked up securely
- Hand-over when needed
Connects to
Works with
- WhatsApp Business Platform
- Website and in-app chat
- Salesforce Service Cloud
- Microsoft Dynamics 365 Customer Service
- Zendesk and Freshdesk
- Genesys and other contact-centre platforms
- SAP and Oracle order data
- Salla, Shopify and custom order systems
- SIP telephony for voice
Where it runs
On-premises
Models and data run on hardware in your own facilities. Nothing leaves your network unless you allow it.
Private cloud in the Kingdom
Dedicated, isolated capacity in a Saudi data-centre region of a provider you select — the elasticity of cloud without a shared public AI service.
Hybrid
Sensitive workloads stay on-premises; less sensitive ones use private cloud capacity. You set the policy for what goes where.
What we measure
Measures we agree with you
We agree these measures with your service leads before launch and report against them on a reviewed sample of real conversations. They are targets we track together, not results we promise in advance.
01
Correct resolution rate
Share of conversations the agent resolved correctly, checked on a reviewed sample — not simply conversations that ended.
02
Answer accuracy by language
Share of sampled answers that are correct and supported by approved content, reported separately for Arabic and English.
03
Hand-over quality
Share of hand-overs made at the right moment, and how often staff had to ask the customer to repeat something.
04
Customer experience
Post-conversation rating and time to first response, compared with the same channels before launch.
Indicative delivery path
Indicative delivery path
- 01
2–3 weeks
Content and scope
We pick the first intents from real transcripts, gather and approve the content that answers them, and agree hand-over rules and measures with your service leads.
- 02
4–6 weeks
Build and integrate
The agent, the channels and read-only lookups to your CRM or OMS are built, then tested on Arabic dialect and English transcripts, including red-team rounds.
- 03
2–4 weeks
Controlled launch
The agent goes live on one channel or for a share of traffic, with close review of conversations and quick fixes to content and rules.
- 04
Ongoing
Expand and monitor
More intents and channels — and voice, if you choose it — are added, with continuous quality sampling and periodic red-team rounds.
Durations are indicative and depend on scope, data and integrations.
Questions
Customer service agent
Will it make things up?
It is built not to. The agent answers only from approved content and tool results, is tested with questions designed to tempt it to improvise, and hands over when it cannot find a supported answer. No language model is perfect, which is why sampled quality review continues after launch.
Where do WhatsApp conversations go?
WhatsApp messages pass through Meta’s WhatsApp Business Platform before they reach your systems, however the agent itself is hosted. The model, knowledge base, conversation logs and customer lookups can all stay in your environment. We document the data flow so your compliance team can assess it against the Personal Data Protection Law and your own policies.
Can it take actions, such as cancelling an order?
It can, but we start read-only. Actions with financial or customer impact — refunds, cancellations, booking changes — are added one at a time, with limits and, where you choose, a staff approval step. Every action is logged with the conversation that led to it.
Does it work on the phone?
Voice is optional: it adds speech recognition and speech synthesis to the same agent. Arabic speech recognition on real call audio, with dialects and background noise, needs testing on your own recordings before we recommend it. We usually suggest starting with chat and WhatsApp, and adding voice once the content and hand-over rules have proven themselves.
Other solutions
All solutionsPrivate AI platform
Open-weight and commercial models served inside your boundary, behind one governed gateway.
View solutionKnowledge assistant
Cited answers from your documents and systems, in Arabic or English — limited to what each user may open.
View solutionDocument intelligence
Invoices, contracts and applications read, checked against your records and posted — people review the rest.
View solutionAgentic operations
AI agents that handle procurement, invoice, HR and IT requests end to end — within limits you set.
View solution
Have a process that should run itself?
Tell us about it. We will help you work out whether AI is the right answer, what it would take, and what it should return.

