AI in Telecom: One Customer, One Decision
A customer is about to run out of data. The operator knows the balance, the usage pattern and the available offers. Yet the next message may still be the same campaign sent to thousands of other people.
That is the opportunity I see for AI in telecom: bring the quality of the decision closer to the richness of the information we already have.
Customer value management—CVM—is the discipline of growing and protecting the value of existing customer relationships. AI can change its operating model. Instead of starting with a campaign and finding an audience, start with the customer and choose the most useful next action.
One customer. One decision. Including the decision to do nothing.
The unit of work should be the customer
Segmentation remains useful. It helps us understand needs, design propositions and allocate resources. But two customers in the same segment can need very different things today.
One may benefit from a larger data allowance. Another may have an unresolved service problem. A third may be perfectly happy and need no contact at all.
My proposed Self-Driving Value Engine treats these as individual commercial decisions. People define the objectives and limits; software evaluates the available actions, coordinates execution and learns from the response.
The goal is a more valuable customer relationship, with fewer irrelevant interventions. More messages are not a success measure.
Six jobs behind a useful offer
I separate the work into six functions. They can be implemented as specialist agents, conventional models, rules or a combination. Calling something an agent does not make the decision better.
- Value: Identify where the relationship could develop, without confusing higher peer spending with proven customer demand.
- Profile: Understand the customer’s services, usage, experience and relevant permissions. Poor data should reduce confidence, not trigger more aggressive selling.
- Decision: Compare eligible actions—including no action—against expected incremental value and customer benefit.
- Content: Explain an approved offer clearly. AI may adapt the language; it should not invent prices, benefits or eligibility.
- Timing: Choose an appropriate moment and channel within the customer’s preferences and contact limits.
- Orchestration: Resolve conflicts, enforce the budget and decide whether anything should be sent.
That last function matters. Separate teams can each make a reasonable offer and still create an unreasonable experience when all their messages reach the same person.
A value gap is a question, not a sales target
Consider an illustrative prepaid customer, Orxan. He spends 12 AZN a month; comparable customers spend 24 AZN. The 12 AZN difference is worth investigating.
It does not follow that Orxan can—or wants to—spend twice as much. The comparison may miss differences in income, needs, another SIM or shared household services. The peer benchmark is a signal, not a measure of what the customer owes the operator.
Suppose his data repeatedly runs out late in the month. A relevant short-term pack could be useful. But if he recently complained about billing, resolving that problem may be the better next action.
Even an accepted offer can destroy value if it discounts a purchase he would have made anyway. The decision therefore needs a counterfactual: what would probably happen without this intervention?
The household changes the commercial question
A SIM is a product relationship. A household may contain several people, several devices and services bought from different providers.
Looking at one line at a time can miss that context. An offer that adds a family member or solves a shared home-internet need may create more value than another upgrade on the original line.
The following example shows how the question changes. It does not predict a real family’s spending or an operator’s returns.
The commercial objective is to earn a larger share of the family’s digital needs through useful services. That requires coordination. Members should not receive competing offers, and linking a household must respect consent, access rights and individual privacy.
Loyalty belongs in the same conversation. A useful offer can develop the relationship; recognition and dependable service can help sustain it. Neither effect should be assumed simply because both programmes exist.
Measure what changes because of the decision
Campaign sales are easy to report. Incremental value is harder—and more important.
A properly designed randomized holdout helps estimate what would have happened without the intervention. Compare equivalent populations over the same period, report uncertainty and account for offer costs. This follows the logic of controlled experimentation.
For household offers, consider whether the household should be the assignment unit: treating one member can influence another. Individual assignment may obscure that effect.
I would review outcomes at three levels:
- Commercial value: incremental revenue and contribution after incentives and relevant delivery costs—not attributed sales alone.
- Customer relationship: retention, useful engagement, complaints and contact fatigue.
- Decision quality: compliance with eligibility and consent rules, conflicting offers, exceptions and the ability to explain an action.
One further distinction matters: the actions most likely to produce a response are not necessarily the actions most likely to change behaviour. Strong CVM needs to learn the difference.
Autonomy should be earned
The destination is not a system nobody watches. It is a system that can handle more decisions safely while remaining accountable to people.
Begin with recommendations that people review. Let the system execute proven playbooks within defined limits. Expand only when the evidence supports it, with monitoring, exception handling and a way to stop or reverse an action. That emphasis on oversight and ongoing measurement is consistent with the NIST AI Risk Management Framework.
The commercial team still owns the hard choices: what value means, which tradeoffs are acceptable and when the customer’s interests require restraint. Models can inform those choices; they cannot remove responsibility for them.
The question I would take to the leadership team
“Can we use AI to run more campaigns?” is too small a question.
I would ask: Can we make a better decision for each customer, coordinate it across the household and demonstrate the value it adds?
That changes the investment conversation. The priority becomes the data, decision rules, measurement and accountability needed to improve customer relationships—not the volume of content the system can generate.
A self-driving value engine earns its name through the quality of those decisions. Sometimes it sends a better offer. Sometimes it fixes a problem. Sometimes it leaves the customer alone.
The framework is a proposed operating model. Customer and household figures are illustrative examples, not measured business results or forecasts.