01 — Why CX programmes stall
Buying tools before fixing identity
The usual sequence: a personalisation tool for the website, a separate messaging platform, a loyalty system from a different vendor, and a contact centre that has been in place for a decade. Each works. Together they produce a customer who gets a welcome offer for a product they bought last week.
The reason is almost always identity. The website knows a cookie. The app knows a device. The loyalty system knows a phone number. The contact centre knows a national ID. Nothing reconciles them, so there is no such thing as "the customer" — only four partial versions who never meet.
A CX platform is the layer that makes those four one. Unified data, a decisioning engine that can act on it, and orchestration that decides which channel does what. Tools bolted on above that layer work. Tools bolted on without it produce the welcome offer.
| Moment |
Siloed channels |
Unified platform |
| Customer browses on mobile, buys in store | Two unrelated sessions | One journey, attribution intact |
| Complaint raised in the app | Agent starts from nothing | Agent opens with full history |
| Customer churns | Noticed at renewal | Flagged weeks earlier by behaviour |
| Marketing sends a campaign | Segment from stale export | Audience from live profile state |
| Customer asks for deletion | Handled system by system | Propagates from one request |
02 — The data layer
A customer data platform, and what it is not
A CRM records what your organisation did to a customer — calls logged, tickets raised, opportunities opened. A CDP records what the customer did, across every channel, resolves it to one identity, and exposes the result in real time to systems that need to act on it. One is a system of record. The other is a system of behaviour.
Buying a CDP product does not produce this. The work is in identity resolution rules, event schema design, and deciding what a profile is allowed to remember — which under PDPL is a governance decision as much as a technical one.
- Identity resolution — deterministic matching on national ID, phone and email, with probabilistic matching only where the risk of a wrong merge is acceptable
- Event schema — a defined vocabulary of customer events, so "purchase" means the same thing in the app as in the store
- Real-time profile access — sub-second reads for channels that must decide during a session, not overnight batch exports
- Consent state on the profile — marketing permission travels with the customer rather than living in one channel
- Retention and deletion — a deletion request propagates to every downstream store, provably
- Audience governance — data collected for one purpose cannot silently become a marketing segment
03 — The journey
Five stages, and where each one usually breaks
Journey maps are common; instrumented journeys are rare. Each stage below lists the channels involved and the failure we most often find when we measure rather than assume.
-
Discover
Search · social · app stores · Arabic SEO
Where the customer first meets you, usually on a phone and usually in Arabic.
Common failure: Arabic search terms unmapped to English catalogue data, so the product exists but cannot be found.
-
Engage
Web · app · WhatsApp · chat · contact centre
The stage with the most tools and the least consistency between them.
Common failure: Channel switching resets context — the customer repeats themselves, and satisfaction drops on the second telling.
-
Convert
Checkout · mada · Apple Pay · BNPL
The narrowest part of the funnel and the most instrumented, though rarely instrumented well.
Common failure: Payment options that miss local preference, and address entry designed for a Western format.
-
Deliver
Order tracking · SMS · push · email
Post-purchase, where expectation is highest and attention is lowest.
Common failure: Proactive updates absent, so the contact centre absorbs volume that automation should have prevented.
-
Retain
Loyalty · offers · re-engagement · service recovery
Where margin actually lives, and the stage most often owned by nobody.
Common failure: Retention treated as a discount problem rather than a behavioural one — no propensity signal, so spend goes to customers who were never leaving.
04 — Where AI actually helps
Six applications that survive contact with production
AI in CX is oversold in general and undersold in specifics. It does not repair a broken journey or unreliable data — it amplifies whatever the underlying experience already is. Applied to these six, on a working data layer, it earns its cost.
Ranking and recommendation
Where a catalogue is too large to merchandise by hand, learned ranking beats manual rules quickly. The measurable win is usually in search results and category ordering rather than the recommendation carousel everyone builds first.
Search ranking · Next-best-offer
Propensity and churn modelling
Scoring who is likely to lapse, upgrade or respond lets retention spend go where it changes an outcome instead of discounting customers who were staying anyway. See also our propensity modelling work.
Churn · Propensity · LTV
Arabic conversational AI
Modern Arabic models handle Modern Standard Arabic well and Gulf dialect reasonably — enough to resolve a meaningful share of routine contacts. The design decision that matters is the handover, not the deflection rate.
Arabic NLP · Dialect · Handover
Sentiment and feedback analysis
Reading Arabic reviews, survey verbatims and contact transcripts at volume surfaces the recurring complaint that dashboards average away. Usually more actionable than the NPS number it sits behind.
Arabic sentiment · Theme extraction
Send-time and channel selection
Choosing when and where to reach someone is a smaller, more reliable win than choosing what to say. It also degrades gracefully — a wrong channel is recoverable, wrong content is not.
Timing · Channel · Frequency capping
Generative content variation
Producing message and creative variants in Arabic and English at volume, with human review before anything customer-facing ships. Useful for throughput; risky as an unreviewed pipeline.
Variants · Bilingual copy
05 — Saudi-specific requirements
What platforms built elsewhere get wrong here
CX platforms configured for other markets tend to fail in the same predictable places. These are specified during design rather than discovered after launch.
- Arabic as a first language, not a locale file — right-to-left designed rather than mirrored, and Arabic content written rather than machine-translated from English source
- Dialect in conversational channels — customers write in Gulf dialect and expect to be understood, even where your published content is Modern Standard Arabic
- Hijri calendar and seasonal demand — Ramadan, Eid and national day shift behaviour far more than the Gregorian retail calendar most platforms assume
- mada before international schemes — local card coverage is the dominant rail; launching without it visibly costs conversion
- BNPL as mainstream, not niche — Tamara and Tabby are ordinary checkout expectations for a large share of shoppers
- WhatsApp as a primary service channel — often higher engagement than email, with different consent and template rules to respect
- Nafath where identity assurance matters — national single sign-on for verified identity rather than a bespoke KYC flow
- PDPL consent as platform architecture — lawful basis, purpose limitation and retention affect schema and hosting region, not just the privacy page
06 — How we deliver
Data layer first, then one journey end to end
The failure mode of CX programmes is breadth: every channel improved slightly, nothing measurably better. We take one journey all the way through instead, prove the lift, then widen.
-
01
Experience audit
Instrument the current journey and measure it. Where customers drop, where they repeat themselves, what the contact centre absorbs that automation should prevent.
2–3 weeks
-
02
Identity and data design
Resolution rules, event schema, consent model and retention policy. The least visible phase and the one everything later depends on.
3–4 weeks
-
03
Platform build
Profile store, real-time access, decisioning and orchestration into the channels in scope. Integrations built against what each system actually supports.
6–8 weeks
-
04
First journey live
One journey end to end, measured against the audit baseline rather than against expectations.
3–4 weeks
-
05
Widen and tune
Additional journeys, AI models introduced where the data now supports them, and continuous measurement against the same baseline.
Ongoing
A first useful release typically lands in 10 to 16 weeks. Full multi-channel programmes run 6 to 12 months, though value should be measurable long before the end of that.
07 — Questions
Answered plainly
What is a customer experience platform?
The layer that lets separate channels behave as one system: a unified customer data layer, a decisioning or personalisation engine, and journey orchestration across web, app, messaging, contact centre and store. Without the data layer underneath, most CX programmes are a set of disconnected tools.
How is a CDP different from a CRM?
A CRM records what your organisation did with a customer — calls, tickets, opportunities. A CDP records what the customer did across every channel, resolves those events to one identity, and makes the profile available in real time to other systems. CRM is a system of record; a CDP is a system of behaviour. Most enterprises need both.
Where does AI genuinely improve CX?
Four places consistently: ranking and recommendation where catalogues are too large to merchandise manually, propensity and churn modelling to direct retention spend, Arabic language understanding for service automation, and send-time or channel selection. AI amplifies the underlying experience — it does not repair a broken journey or unreliable data.
How does PDPL affect personalisation?
Personalisation processes personal data, so lawful basis, purpose limitation, retention and data subject rights become platform requirements. Consent state travels with the profile, deletion propagates to every downstream system, and audiences cannot quietly be built from data collected for another purpose.
Do Arabic chatbots work well enough for service automation?
Well enough to resolve a meaningful share of routine contacts. Modern Standard Arabic is handled well and Gulf dialect reasonably. The decision that determines whether customers accept it is the escalation path: a bot that hands over cleanly with full context outperforms one that tries to resolve everything and pushes people to call anyway.
How long does a CX platform take to build?
A first useful release typically lands in 10 to 16 weeks, covering identity resolution, one unified profile and one or two orchestrated journeys. Full programmes run 6 to 12 months, with value measurable well before completion.