First commercial AI use case for a telecom operator: how to choose

Seven evaluation parameters

The first AI use case is often picked on “what is hot in the industry” or “where we have more data”. Both criteria are sub-optimal. A hot use case requires a mature operating machine that does not yet exist. “Where data is greater” does not mean “where value is greatest” — data may exist but action does not.

The systematic approach is to evaluate use cases on seven parameters, each with its own weight. These seven parameters form a decision matrix.

Data readiness. Are there data of the required quality and volume for training? No data, no ML. If data still has to be collected — 6-12 months of preparation. If it is there — start in 60 days.

Business sponsor. Is there an owner who carries P&L and takes the decisions? Without a sponsor the model is built into thin air. With a sponsor it is integrated into operations from day one.

Integration complexity. How many systems must be integrated for the action to occur. If the pilot needs integration with billing, CRM, app, contact centre — that is heavy build. If it lives in one system — light.

Regulatory exposure. What regulatory risk does the use case carry. AI taking financial decisions — high. AI classifying complaints — low.

Talent availability. Are there people on the team who can build and maintain it. If not — external resource or 6-12 months of hiring.

Time to value. When will the pilot show a measurable result. A use case with a 3-month feedback loop is easier than one with 12.

Downside risk. What happens if AI is wrong. A misclassified support ticket is a small price. A wrong fraud decision means lost trust and compensation.

Five candidates and their scores

Take the five most common first use cases and score each on the seven parameters.

Candidate 1: support deflection chatbot

Parameter	Score
Data readiness	High — many contact histories
Business sponsor	High — head of customer care
Integration complexity	Medium — chatbot plus handoff to live agent
Regulatory exposure	Low — general tariff questions
Talent availability	Medium — NLU specialist or vendor
Time to value	High — 3-6 months to first results
Downside risk	Low — customer escalates to agent

Total: 6/7 high, 1/7 medium. One of the best entry points.

Candidate 2: churn prediction model

Parameter	Score
Data readiness	Medium — needs good signals and labelled data
Business sponsor	High — head of retention
Integration complexity	Medium — needs a retention process for action
Regulatory exposure	Medium — using data for personalisation
Talent availability	Medium — data scientist plus retention manager
Time to value	Medium — 6-9 months to validation
Downside risk	Low — false positive is just an extra call

Mixed. Good if the retention process already works; risky if retention is reactive.

Candidate 3: general fraud detection

Parameter	Score
Data readiness	Medium-high — fraud cases labelled
Business sponsor	Medium — fraud team exists, usually not at C-level
Integration complexity	High — real-time decisions across systems
Regulatory exposure	High — false positive blocks the customer
Talent availability	Low — fraud ML specialists are scarce
Time to value	Medium — 6-12 months
Downside risk	Medium — wrongly blocked customers get angry

Strong value, hard entry point. Better as the second or third use case once operational AI processes have stabilised.

Candidate 4: network anomaly detection

Parameter	Score
Data readiness	High — network data is there
Business sponsor	Medium — usually CTO, commercial impact unclear
Integration complexity	Medium — fits NOC processes
Regulatory exposure	Low — internal tool
Talent availability	Medium — network ML specialist or vendor
Time to value	Medium-long — 9-12 months
Downside risk	Low — false alert handled by NOC

A reasonable choice if CTO is willing to cooperate with CCO on using the results in proactive customer care.

Candidate 5: Next Best Offer (NBO)

Parameter	Score
Data readiness	Medium — historical offer data needed
Business sponsor	High — head of digital monetisation
Integration complexity	High — decision engine plus multi-channel orchestration
Regulatory exposure	Medium — consent for personalisation
Talent availability	Medium — ML plus decision engineer plus experimentation
Time to value	Long — 12-18 months to scaled value
Downside risk	Low — wrong offer the customer ignores

The most ambitious and the hardest first use case. Not recommended as first — better as 2nd or 3rd after simpler wins.

Which first use case is usually chosen, and which should be

Most often it is NBO or AI marketing. The board sees margin and assumes AI will deliver fast results here. In practice it is the hardest entry point with the longest time to value.

Correctly chosen — the support deflection chatbot. Easy entry, fast ROI, low risk, and the team builds initial ML experience.

A good second use case is often fraud detection — once the chatbot is running and the team has built up experience.

NBO is better as the third or fourth, once a proven decision engine, experimentation culture and data foundation exist.

What separates an AI-ready organisation

Several capabilities must be in place before AI makes sense to launch.

Master customer data. One customer, one ID across all systems.

Event collection. A pipeline for real-time or near-real-time data.

Operating routine. A weekly review where results are discussed and decisions taken.

Experimentation culture. Willingness to test hypotheses and accept results that contradict the initial opinion.

Audit and governance. A process for approving new use cases, monitoring running models, retiring underperforming ones.

A talent core. At minimum 1 ML engineer, 1 data scientist, 1 ML ops. Less than that — a vendor partnership with a clear mandate.

Without these, AI becomes a one-off project that does not scale beyond 12-18 months.

What ROI to expect from the first use case

No promises and no fantasies. Realistic numbers for typical first use cases.

Support chatbot — 15-30% call volume reduction in routine inquiry categories within 6-9 months. The annualised impact depends on contact centre size.

Churn prediction — 10-20% retention improvement in the pilot segment (typically premium customers) within 9-12 months. Translated into P&L through preserved ARPU.

Fraud detection — 20-50% reduction in fraud losses in detected categories. Depends on the current level — if fraud is 0.5% of revenue today, doubling detection delivers 0.25% of revenue. At a large operator that is a meaningful sum.

Network anomaly — 15-30% incident reduction or 20-40% MTTR reduction. Translates into support cost savings and customer experience improvement.

NBO — 15-30% conversion lift over baseline campaigns. But it requires 18-24 months of build-up.

When AI should wait

If the organisation structurally lacks master data — a year of work before AI. Not a “lost year”, a foundation that accelerates everything afterwards.

If there is direct risk that ROI gets invalidated by regulatory changes (for example new restrictions on personal data use) — wait for clarity.

If experimentation culture is missing and every decision is taken “the way it always was”, AI recommendations will be ignored. Culture first, AI second.

If the CFO treats every pilot as “trying to find ROI”, AI will be cancelled within a quarter. AI requires continuing investment of 12-18 months.

Discussion points for the committee

Which three data foundations are best prepared (customer, network, transactional)? This defines the first AI candidate.

Which business owner is ready to put P&L under the pilot result? Without the owner — no project.

What downside risk are we ready to accept? This determines how ambitious the first pilot should be.

What 6-9 month timeline will deliver a meaningful result? Less — pilot is a toy. More — hard to keep sponsorship.

How SamaraliSoft can help

Telecom AI Use Case Selection Framework — a decision matrix specific to the operator. Scoring of five-to-seven candidate use cases on seven parameters with numbers and context. Recommendation of the first 90-day pilot with a measurement plan. Pilot architecture: data, model, integration, operating routine. And an escalation roadmap from the first use case to an AI portfolio over 18 months.

Related reading

• /en/insights/telecom-ai-native/ — where AI pays back
• /en/insights/telecom-subscriber-intelligence-operating-model/ — operating model
• /en/use-cases/telecom-churn-war-room-mnp/ — churn AI inside retention
• /en/architecture/telecom-around-core-architecture/ — where AI lives

Sources

• Digital Government Portal — digital infrastructure UZ
• Ministry of Digital Technologies — AI strategy 2030
• UzDaily — telecom market