Case study 01 · Telecom
Etisalat Self-Care Portal
Six weeks of fieldwork, 84 users, and more dropped calls than we’d like to count. Here’s how we rebuilt Etisalat’s self-care portal around the problems subscribers actually have — and the moments they currently find out about them.
The problem, in plain terms
We spent six weeks talking to subscribers across Dubai, Abu Dhabi, Sharjah and Ajman. The same frustrations kept coming up, almost word for word.
Reactive support
Found out something was wrong only when a call dropped, a payment failed or a video froze mid-meeting.
Bill confusion
Of call-centre contacts are bill-related. Only 3 of 24 participants could explain a “VAS bundle auto-renew” charge.
Repetition on escalation
Had to repeat their issue to two or more agents — who had no idea what they’d already tried in the app.
Six methods, because no single one tells the whole story
What people say in a survey and what they do in the moment can be very different things. So we combined contextual inquiry (90-minute sessions in homes and offices), moderated usability testing, a 7-day diary study, an 8-minute intercept survey, a heuristic evaluation by four independent evaluators, and a competitive analysis of du, Zain, STC, Ooredoo, T-Mobile and Singtel.
Six screen-reader users were recruited through Al Noor Training Centre from day one. Accessibility was built in from the start, not patched in at the end.
- Wk 1–2Recruit & screen340 contacted, 84 matched
- Wk 2–4Contextual inquiry18 sessions in homes & offices
- Wk 3–4Diary study12 participants, 7 days
- Wk 4–6Usability testing24 sessions vs. current app
- Wk 5–6Survey fieldwork312 responses
- Wk 6–7Synthesis310 data points, journey maps
“Watched Fatima try to dispute a roaming charge from her Abu Dhabi apartment. After 17 minutes of dead ends in the app, she gave up and called support.”
Contextual inquiry, week 3Affinity mapping gave us 310 data points. “Unexpected charge” came up 72 times — more than twice as often as the next theme. Journey mapping showed a 47-minute average gap between the network having a measurable problem and the user finding out.
| Segment | Age | Data usage |
|---|---|---|
| Prepaid | 22–35 | ≥ 3 GB / month |
| Postpaid | 30–50 | ≥ 10 GB / month |
| SMB | 38–60 | ≥ 50 GB total |
Three people we kept coming back to
These aren’t invented archetypes. All three emerged directly from fieldwork, and in a few cases what we found surprised us.
Ahmed Al Mansoori
Digital marketing manager · 32 · Dubai
The “two taps or I’m calling” type“If I need to tap more than twice, I’ll just call. I really don’t have time for menus.”
Fatima Al Mansoori
HR manager · 38 · Abu Dhabi
The one who needs to know“My son burned through 22 GB in one week playing games online. I had absolutely no idea until the bill showed up.”
Omar Al Mansoori
General trading business owner · 54 · Sharjah
The “just tell me what it means” type“What does ‘ISD bundle’ even mean? I just need to know — did that charge come from my trip to London?”
What the research actually showed
By the time people knew something was wrong, it had already cost them
In 15 of 18 sessions, users only learned about a network problem because something had already broken. Telemetry showed the issue had typically been detectable for 47 minutes. The 8% already receiving proactive alerts from a competitor were noticeably happier: NPS +34 against a baseline of +12.
Once the bill was written in plain language, people could understand it
Same charge, different wording: with the annotated prototype, 21 of 24 identified the “VAS bundle auto-renew” charge correctly within 8 seconds.
| Metric | Baseline | Annotated bill | Change |
|---|---|---|---|
| Charge identified | 12.5% | 87.5% | +75pp |
| Time to identify | 94 sec | 8 sec | −91% |
| Escalation intent | 54% | 18% | −36pp |
| Trust score (5-pt) | 2.3 | 4.1 | +1.8 |
| “I understand my bill” | 29% | 84% | +55pp |
Nobody wants to explain their problem three times
When users escalated from the app to live chat, 87% said the agent had no idea what they’d already tried. Showing a summary of what had been passed to the agent before joining the queue made the wait feel 23% shorter — with identical actual wait time — and lifted CSAT by 18 points.
Free text beats menus — mostly
The NLP wizard beat the category menu on almost every measure. But vague inputs like “phone is slow” were misclassified 27% of the time, and Arabic and Urdu accuracy came in at 61% against an 85% launch threshold.
| Metric | Category menu | NLP wizard |
|---|---|---|
| Time to first fix suggestion | 42 sec | 18 sec |
| Correct routing accuracy | 71% | 89% |
| User satisfaction (1–5) | 3.1 | 4.3 |
| Task abandonment | 31% | 14% |
| Ambiguous input failure | — | 27% misclassified |
Ask when unsure, never make people start over
The findings turned into two flows that carry most of the portal’s value. Smart Diagnose lets people describe a problem in their own words; when the model isn’t confident, it asks instead of guessing. And wherever a user gives up on self-service, everything they’ve tried travels with them to the agent.
- StartDescribe the issue“My calls keep dropping at home”
- AIClassify intentUnder 2 seconds
- CheckRun diagnosticsLine, device & tower status
- DecisionConfidence ≥ 70%?Gate before routing
No — ask, don’t guess
- Clarifying question2–3 options: “indoor coverage or device speed?”
- User picks oneTwo seconds, not a wrong route
Back to classification with the new signal
Yes — show fixes
- Ranked fixesOrdered by real success rate, with time to complete
- Rate each step👍 / 👎 retrains the model
- DoneResolved
- AppSelf-service attemptFixes tried, diagnostics run
- UserEscalate to chatFrom any screen
- Summary“Here’s what we’ve shared”3 bullets: tried, flagged, last contact
- CRMAgent briefed firstBefore accepting the session
- QueueWait over 10 min?Live estimate every 60 s
- ResolveChat or callback71% prefer a callback slot
Nine screens, each backed by evidence
Every screen in the portal traces back to something we saw in the field. A few of the decisions that mattered most:
Dashboard & account overview
Status ribbon at the very top, a burn-rate sparkline (“projected to last until Jun 28”) and quick actions two taps from home.
Plain-language bill explainer
“Why this charge?” on every line, about 20 words each. Localised copy in Arabic, English and Urdu.
Recommended plans
A bottom sheet, not a modal, showing the exact saving — “Save AED 45/month” — and the usage behind it.
Smart Diagnose
Free-text symptom wizard with a confidence gate, ranked fixes and an escalation path that never starts over.
Network dead-zone map
Show the map first, ask later. Every pin is fuzzed to a 200-metre radius; tower upgrade ETAs reduce churn intent.
Family plan management
Per-line allocation with steppers on small screens, no auto-save, PIN-protected parental controls and an 80% usage alert.
Usage analytics
A 30-day rolling model, tap-to-drill app breakdown and a plain-text fallback for screen readers.
Live support
“Agent has been briefed” confirmation, live wait estimate and a callback offer once the queue passes 10 minutes.
Profile & settings
Separate toggles and channels per alert type, coverage contribution off by default, four interface languages.
The accessibility issues we found
Four evaluators went through the prototype independently — two sighted, one low-vision, and one who relies entirely on a screen reader. Two findings are launch blockers.
| WCAG criterion | Issue | Severity | Fix |
|---|---|---|---|
| 1.4.3 Contrast (AA) | Amber status ribbon fails on white at small sizes | Critical | Darken to #B8860B; pair colour with icon and text |
| 1.1.1 Non-text content | Sparkline has no screen-reader alternative | Critical | aria-label: “Data projected to last until June 28” |
| 4.1.3 Status messages | NLP input lacks an error announcement | Serious | aria-live for errors and the classification result |
| 4.1.2 Name, role, value | Parental toggles lack descriptive names | Serious | Include the line name in each toggle’s label |
| 2.4.7 Focus visible | Bill line expansion has a weak focus indicator | Moderate | 3px teal focus ring; move focus into the panel |
What changed
These are projections from usability and concept testing, not production data — but the differences are large enough to take seriously.
Correct charge identification
Up from 12.5%
To understand a charge
Down from 94 seconds
Escalation intent
Down from 54% after viewing the bill
Task abandonment
Down from 31% with the NLP wizard
Bill trust score
Up from 2.3
NPS with proactive alerts
Up from a +12 baseline
How we sequenced the work
Order was driven by three things: how strongly research backed the feature, how complex the build is, and what unresolved risks we’d be carrying.
Foundation
Hits all three core pain points. The bill explainer alone could take a meaningful chunk out of call-centre volume.
Intelligence
Gated by NLP accuracy reaching 85% in Arabic and Urdu, and legal sign-off on the dead-zone map under UAE data law.
Completion
Context handoff needs a CRM integration; family controls need a secondary PIN system and audit log storage.
Five things we’re confident about
Give the NLP wizard a way out when it’s not sure
Below 70% confidence, show two or three clarifying options instead of guessing.
Evidence: 27% misclassification on ambiguous inputs; flagged independently by three of four evaluators.
Let people see the map before asking them to share anything
Read-only first, then offer the 500 MB reward. Asking cold significantly lowers opt-in.
Evidence: opt-in rose from 44% to 71% with the incentive.
Tell users what you’ve already shared with the agent
Three items is enough: what they tried, diagnostic flags, last contact.
Evidence: perceived wait −23%, CSAT +18; 18 of 24 called it “a relief”.
Fix Arabic and Urdu accuracy before launch, not after
61% accuracy isn’t a follow-up item — it’s a blocker for some of the fastest-growing subscriber segments.
Evidence: tested with 6 participants in both languages; target 85% before Sprint 2.
Alert family admins before the line runs out, not after
An 80% usage alert per line, once a month, removes most bill surprises.
Evidence: 13 of 16 admins rated it “very useful”; 78% had a surprise bill in 3 months.