Every telecom operator has a digital transformation programme, a self-serve app and at least one AI pilot. Almost none of them have fixed the bill. Billing remains the number one source of customer confusion and support volume, accounting for up to 50 percent of contact centre calls, and most billing systems have not meaningfully evolved in over a decade.
The regulator’s mailbag says the same thing. Canada’s Commission for Complaints for Telecom-television Services accepted a record 23,647 complaints in the year to July 2025, a rise of 17 percent, and billing accounted for 46 percent of all issues raised. Billing is not a back office function. It is the most frequent and most emotionally charged conversation an operator has with its customers.
The challenge is not a lack of ambition. It is execution on a stack that was never meant to deliver modern experiences. Legacy billing platforms were designed when transparency meant mailing a paper invoice, not answering real-time questions, and certainly not for AI.
That does not mean the core has to be replaced. The decision facing transformation executives is narrower: keep funding a multi-year rip and replace, or bridge the gap with a governed intelligence layer that sits on top of the systems already running the business. This playbook sets out the second path in four steps, drawn from implementations at TELUS, Verizon and Vodafone.
Billing is the highest volume, lowest satisfaction interaction in telecom. It drives up to 50 percent of contact centre calls and 46 percent of all issues raised in Canadian telecom complaints, yet the systems behind it are a decade out of date.
Generative AI does not require replacing them. Layered over existing BSS and CRM through APIs and middleware, it explains charges in plain language, grounded in real account data, available around the clock.
The four-step playbook covers integration, feedback loops, technology stack and regulatory compliance. Operators applying it report 20 to 40 percent fewer billing calls. The constraint is not the model. It is integration discipline and governance.
By the numbers
50%
of telecom contact centre volume stems from billing-related issues, in the channel operators spend most to shrink
– Industry research, 2025
46%
of all issues raised in complaints accepted by Canada’s telecom watchdog were billing related, with incorrect charges and missing credits the biggest share
– CCTS, 2025
28%
more TELUS customers self-served successfully after a generative AI support assistant answered more than 50,000 queries within weeks
– TELUS, 2024
40%
increase in service channel sales at Verizon after generative AI was scaled enterprise-wide to assist 28,000 live customer service agents
– Verizon, 2025
0.52%
of total revenue is now lost to leakage across surveyed operators, down from 1.22 percent in 2021, so accuracy has improved even as billing complaints have not
– TM Forum Business Assurance Survey, 2025
$1.3T
in operator capital expenditure is forecast for 2024 to 2030, overwhelmingly network investment rather than the billing layer customers see
– GSMA, 2025
Why Billing Now Decides Whether Customers Stay
Customers expect clarity. Most billing systems were designed when transparency meant mailing a paper invoice. Today a customer wants to tap a screen and get a plain explanation of what they are paying for. Instead they wait on hold, navigate an IVR, and reach an agent reading from a script. That is attrition in progress.
Where Legacy Hurts the Most
For decades operators have relied on legacy software, including IVRs, outdated CRMs and antiquated ticketing systems. These tools once revolutionised service efficiency and are now the primary roadblock to modern expectations. Siloed databases and disconnected CRM and billing platforms leave agents with slow interfaces and incomplete data that forces repetitive information requests.
Integration, not intent, is the binding constraint. According to a 2024 ExecsInTheKnow study, “90 percent of telecoms have deployed AI, yet 74 percent still cite poor cross-channel integration as a CX blocker.”
The Evidence Points in Two Directions
The data on billing does not all run the same way. TM Forum’s Business Assurance Survey for 2024/25 reports revenue leakage down to 0.52 percent of total revenue from 1.22 percent in 2021, and more than 90 percent of communications service providers now running a dedicated business assurance function, up from around 40 percent in 2021/22. Operators have become better at making sure the bill is correct.
Complaint data points the other way. The CCTS recorded billing as the leading complaint issue at 46 percent of all issues raised, with billing complaints up 16 percent and at a five-year high. Both findings can be true at once. Accuracy has improved. Comprehensibility has not.
That reframes the investment. Revenue assurance tooling solved a calculation problem. What remains is an explanation problem, and further assurance spend will not close it. Billing is front-line customer experience now, and where generative AI delivers measurable return.
Why Billing Is the Right Place to Start with Generative AI
Telecom companies underestimate how much damage unclear billing does, until call volumes spike, agent turnover rises and customers quietly leave. The damage is not one problem. It is seven, and each lands on a different line of the profit and loss statement.
| Billing pain point | What it costs the business |
|---|---|
| Unclear invoices | Bills are built for internal accuracy, not customer clarity. Fragmented, technical charges erode trust and generate avoidable service volume. |
| No interactive support | Legacy platforms were never built for real-time questions or self-service, so customers hit slow portals and call queues instead. |
| Inbound call overload | As much as 50 percent of contact centre volume is billing related, clogging channels and delaying every other issue. |
| Increased cost to serve | Repetitive inquiries tie up agents, extend handling time and inflate staffing, while adding nothing for the customer. |
| Agent burnout | Teams spend the day walking customers through line items, which drives disengagement, turnover and training cost. |
| Poor satisfaction | Long waits and limited digital options leave customers feeling undervalued, especially when billing is their only regular contact. |
| Brand erosion | Billing friction does not only depress net promoter scores. Frustrated customers churn, and say so publicly in reviews and social channels. |
The last row is where the argument sharpens. Retention economics in telecom are unforgiving, and billing is one of the few touchpoints an operator controls every month. For the design side of the same problem, see the five design-led UX steps to reduce telecom churn.
The Opportunity Hidden in the Complaint Volume
Generative AI can explain a bill in natural language, personalised to the individual, grounded in real account data and available around the clock. It does not replace the agent. It scales the agent. It sits on top of the existing stack and connects to CRM and billing through APIs.
In pilots designed for telecom operators in exactly this space the pattern repeats: fewer calls, higher app usage, and agents free to solve real problems rather than explain line items. Operators deploying targeted generative AI against billing report 20 to 40 percent fewer billing-related calls within months.
Billing is the best first use case for generative AI in telecom precisely because it is unglamorous.
It is high volume, highly repetitive, grounded in structured data the operator already owns, and measured by metrics the business already reports to the board. Value is provable inside a quarter, and failure surfaces early and cheaply.
ML arteka works with operators to place this intelligence layer over systems that are staying exactly where they are, so the first win funds the next one.
What Leading Operators Have Already Proven
This is not a theoretical playbook. Several carriers have deployed generative AI against service and billing problems, and the results are public. They matter as evidence of which architectural choices survive a legacy estate.
TELUS: A Self-Service Lift Through Chatbot Integration
TELUS introduced a generative AI support chatbot in 2024 using Azure OpenAI and the Fuel iX engine. It answered over 50,000 queries within weeks and enabled 28 percent more customers to self-serve successfully. Built on responsible AI principles and Privacy by Design, it integrates more than 1,000 vetted support articles. The grounding in vetted content, not the model, made the answers defensible.
Verizon: Live Agent Augmentation
Verizon adopted generative AI to augment, not replace, its 28,000 customer service agents. The assistant, powered by Google’s large language models, listens to live calls and recommends real-time responses or next-best actions. Since scaling enterprise-wide in early 2025, Verizon reports a 40 percent increase in service channel sales, achieved without the customer ever meeting a bot.
Amdocs: A Vendor Platform Aimed at Billing Queries
Amdocs, an industry vendor, built the amAIz platform for telecom billing queries using large language models fine-tuned on billing data. Applying retrieval-augmented generation on NVIDIA infrastructure, it improved answer accuracy and cut inference costs by 40 percent. That figure is the vendor’s own, but the lesson holds: a model grounded in billing data beats one reasoning about bills.
Vodafone and AT&T: Two Different Bets
Vodafone used AI chatbots for common billing and plan inquiries, reducing customer wait times by over 50 percent. AT&T took the opposite route, focusing on backend AI for network optimisation and predictive maintenance. Not customer facing, but the reliability gains benefited experience indirectly.
Whether customer facing or agent augmenting, the pattern holds. AI improves experience, shortens resolution times and reduces cost, without replacing legacy systems.
The Four-Step AI Playbook for Legacy Billing
Implementing generative AI in a telecom environment is not about chasing a trend. These four steps form a pragmatic, low-risk framework for layering AI onto existing systems, moving from pilot to production without disrupting operations or customer trust.
| Playbook step | What it delivers |
|---|---|
| Step 1: AI integration strategies | Custom APIs, middleware and knowledge base connectors that let an assistant read and act on legacy BSS, OSS and CRM data, piloted narrowly first. |
| Step 2: Feedback loops and continuous learning | Refreshed training data, user feedback, human review of failed queries, multi-turn context handling and performance monitoring. |
| Step 3: Technology stack | Large language models, machine learning for personalisation and analytics, responsible AI guardrails, and a deliberate cloud or on-premises decision. |
| Step 4: Regulatory and ethical considerations | CCPA and PIPEDA compliance, consent and transparency, CPNI-grade security and auditability, and FCC and CRTC consumer protection obligations. |
Step 1: AI Integration Strategies for Legacy Telecom Systems
Legacy CRM, billing and support systems often lack modern APIs and behave like walled gardens of data, so integration needs planning before modelling does. Five practices have emerged.
- Data integration via APIs or middleware. Most legacy BSS and OSS platforms were never designed to talk to external AI services. Identify the critical data first, typically customer profile and bill detail, then build the connectors that expose it. A chatbot needs an API to pull the bill before it can explain a $20 overage fee.
- Connection to knowledge bases. Many queries can be answered from FAQs, support articles and device manuals that already exist. Plugging the assistant into those repositories gives current answers aligned with company policy.
- Legacy CRM and billing hooks. An assistant handling a dispute may need to trigger a credit or open a ticket, which means exposing selected write operations through secure APIs or robotic process automation.
- Compatibility and performance. Old systems may not tolerate high-frequency calls. Cache common data, queue heavy processing, and index billing records so answers return within seconds.
- Phased rollout and testing. Pilot against a subset or copy of the data, test with employees, then release to one region before going wide.
Modernisation happens in parallel rather than first. Custom APIs bridge legacy systems to AI platforms while the IT infrastructure is gradually modernised.
Step 2: Feedback Loops and Continuous Learning
Deploying AI is not a set and forget endeavour, particularly in telecom, where products, policies and customer behaviour change continually. The system needs a feedback loop for continuous learning.
- Ongoing training data updates. Offerings and support issues evolve, so the knowledge base and training corpus must stay current, which may mean fine-tuning on transcripts of unresolved queries as new topics emerge.
- User feedback integration. Collect explicit ratings after each interaction. Negative feedback, and cases where the customer still calls support afterwards, should be analysed for failure modes.
- Human oversight and agent review. Queries the assistant could not handle, or answered without confidence, go to a subject matter expert whose resolution feeds back into the knowledge base. Supervised feedback is what improves edge cases.
- Multi-turn context handling. Billing conversations involve follow-ups. Maintain context across turns, and treat a repeated follow-up as a signal that the first answer was not clear enough.
- Performance monitoring and iteration. Track resolution rate, satisfaction, interaction time and escalation rate continuously. A drop usually means the environment changed before the assistant did.
A working loop looks like this. The customer marks an answer unhelpful. A trainer reviews the logged exchange, sees the assistant missed a one-time equipment charge, and adds an example teaching it to check for equipment charges. The next similar question is answered correctly.
Step 3: The Technology Stack for AI on Legacy Systems
A legacy environment demands a stack that interfaces with old systems while delivering modern capability. Four components carry the weight.
- Large language models for natural language interaction. Telecom queries range from “my internet is down” to “explain this charge on page 3 of my bill”. Options include calling a hosted model such as OpenAI GPT-4 or Azure OpenAI Service, or fine-tuning in house. Some carriers choose open-source models fine-tuned on telecom data for privacy reasons.
- Machine learning for personalisation and analytics. Operators sit on usage records, network performance and customer profiles. Include predictive analytics for churn risk and upgrade propensity, a personalisation engine to select the best response, and a dashboard business users can tune.
- Responsible AI and security components. Content filtering and guardrails, encryption in transit between the AI and legacy systems, redaction or tokenisation before data reaches a third-party model, and clear disclosure that an AI is assisting with a visible path to a human agent.
- Platform and infrastructure. Cloud offers agility, on-premises and hybrid offer governance, and roughly 40 percent of telecom firms prefer on-premises AI deployments to control latency and data handling. A common compromise keeps the model cloud-hosted while customer data stays in the operator’s private environment.
Step 4: Regulatory and Ethical Considerations
Telecom operators navigate a complex regulatory landscape, particularly on data privacy, security and transparency. Both the United States and Canada enforce privacy laws relevant to customer communications, and telecoms face stricter rules given the sensitivity of usage data.
- Data privacy laws. The California Consumer Privacy Act gives United States consumers rights over how their personal data is used, and PIPEDA governs personal data handling in Canada. An AI reading billing information and usage history must use that data only for the consented purpose.
- User consent and transparency. Customers should know when they are interacting with an AI and what data it uses. The Government of Canada’s Voluntary Code of Conduct for Generative AI specifically calls for transparency in AI interactions and decisions.
- Security and confidentiality. Call records and locations are highly sensitive, and some are classified in the United States as Customer Proprietary Network Information with strict access rules. Enforce the same authentication and access control as existing systems, and log every AI data access and response for audit.
- Telecom-specific regulation. The FCC and the CRTC apply consumer protection obligations to AI-driven communication as they do to human channels, and anything touching emergency calls or outage information enters strictly protocolled territory.
Treat the assistant as a regulated system from the first pilot, not from the first audit.
Every decision in steps 1 to 3 has a step 4 consequence. Which data the API exposes determines the privacy assessment. Where the model runs determines the residency argument. Whether interactions are logged determines whether a dispute can be resolved at all. Retrofitting those controls after a pilot is how billing programmes stall.
How Should Leaders Measure a Billing AI Pilot?
A billing assistant is easy to demonstrate and hard to defend without numbers. Agree the measurement set before the pilot begins. Four categories are enough.
| Metric category | What to track |
|---|---|
| Deflection and containment | Call deflection rate, digital containment, and the share of billing queries resolved without a human. Operators report 20 to 40 percent fewer billing calls. |
| Efficiency | Average resolution time, handling time for escalated interactions, and cost to serve per billing contact. |
| Experience | Satisfaction on billing interactions specifically, repeat contact rate within seven days, and portal usage for billing tasks. |
| Assurance | Escalation rate to agents, accuracy sampling on answers, and the share of interactions with complete logs for dispute or audit. |
Report the categories together. Deflection alone will make almost any assistant look successful, because in a deflection metric a customer who gives up is indistinguishable from a customer who was helped. Pairing it with repeat contact rate and satisfaction separates containment from silent abandonment.
The assurance row matters most in a regulated estate. If an operator cannot reconstruct what the assistant told a customer about a charge, it cannot defend the interaction or resolve the dispute. Logging is the condition on which the capability scales.
Why the Real Benchmark Is Not Another Carrier
Generative AI in legacy environments is no longer a long shot. It is achievable, with three near-term benefits that compound.
AI Augmentation Improves the Digital Experience
Customers get instant support around the clock, clearer answers and proactive recommendations, which increase satisfaction and loyalty. In a saturated market where competitors sell near-identical products, customer experience is the battleground. Reducing billing confusion and hold times turns frustrated customers into promoters.
Early Movers Build a Durable Advantage
Early adopters develop internal expertise and data insight that laggards struggle to replicate, and they can market the improvement directly, which is difficult with a network upgrade. Industry analysis suggests AI-driven experience gains help operators grow revenue faster than the market.
Responsible Implementation Reduces Regulatory Risk
Implementing AI with responsible principles bolsters trust and mitigates regulatory exposure. Operators that lead on ethical AI position themselves as trusted providers, which carries more weight in telecom because the data is so sensitive.
One framing should govern the investment case, and it comes from outside the industry.
If the digital experience does not meet that bar, it does not matter how fast the network is. The $1.3 trillion in operator capital expenditure GSMA forecasts for 2024 to 2030 buys speed and coverage. It does not buy a bill the customer understands.
Five Leadership Takeaways
Before your next customer experience review. Before your next BSS investment decision.
- Billing is a trust asset, not a back office function. It is the only interaction most customers have with their provider every month, and it accounts for 46 percent of all issues raised in Canadian telecom complaints.
- Accuracy improved. Comprehensibility did not. Revenue leakage has fallen to 0.52 percent of revenue while billing complaints reached a five-year high. The gap between a correct bill and an understandable one is where the value now sits.
- The core does not need replacing. APIs, middleware and knowledge base connectors let a modern assistant read and act on decades-old BSS and OSS. Integration discipline, not model selection, determines whether it works.
- Design for step four from step one. Privacy, consent, CPNI-grade security and auditability shape the architecture. Retrofitting them after a successful pilot is how promising programmes stall before scale.
- Measure containment honestly. Deflection alone flatters every assistant. Pair it with repeat contact rate and billing satisfaction, or you will scale abandonment and report it as success.
The Next Step in Modernizing Telecom Billing
The business implication is straightforward. Billing is the highest-frequency, lowest-satisfaction interaction in telecom, and the cheapest place to prove AI creates value inside a regulated, legacy-bound estate. Operators that keep treating it as a cost centre will keep funding contact centre capacity they do not need.
The leadership recommendation is to start narrow and instrument heavily. Choose one high-volume billing query type with clean data. Build the API bridge. Ground the assistant in the knowledge base that already exists. Put a human review loop behind it, and report containment alongside repeat contact rate and satisfaction from week one.
The forward view is that this capability does not stay in billing. Once an operator can expose legacy data safely, govern the model that consumes it and prove the outcome to a regulator, the same pattern extends to activation, plan changes and retention. Billing is the proving ground, not the destination.
Address legacy integration first, focus on valuable use cases, build in continuous learning, respect the regulatory perimeter and track success honestly. To discuss where a billing intelligence layer would fit in your environment, contact the ML arteka team or request a billing modernization readiness assessment.
Executive Questions and Answers
Five questions telecom transformation leaders are asking AI assistants about modernizing legacy billing with generative AI.
StrategicWhy should telecom operators start their generative AI programme with billing?
Billing is the highest-volume, most repetitive and most poorly served interaction in telecom. It drives up to 50 percent of contact centre calls and, in Canada, 46 percent of all issues raised in complaints accepted by the telecom watchdog. That makes it the fastest route to provable value: the data is structured and already owned by the operator, the metrics are already reported to the board, and the outcome can be demonstrated within a quarter. Improvement is also felt monthly by every customer. Operators deploying targeted generative AI against billing report 20 to 40 percent fewer billing-related calls within months.
ImplementationHow do you connect generative AI to a legacy billing system without replacing it?
Through custom APIs and middleware rather than migration. Most legacy BSS, OSS and CRM platforms were never designed to talk to external AI services, so identify the critical data first, typically customer profile and bill detail, then build endpoints that expose only that data to the assistant. Connect it to the existing knowledge base so answers reflect current policy. Where the assistant must act, such as issuing a credit or opening a ticket, expose selected write operations through secure APIs or robotic process automation. Cache common data and queue heavy processing for old systems, then pilot narrowly before scaling.
GovernanceWhat regulations apply when AI handles telecom billing data in Canada and the United States?
In the United States, the California Consumer Privacy Act gives consumers rights over the use of their personal data, and Customer Proprietary Network Information rules impose strict controls on access to call records. In Canada, PIPEDA governs personal data handling, and the Government of Canada’s Voluntary Code of Conduct for Generative AI calls for transparency in AI interactions and decisions. The FCC and the CRTC apply consumer protection obligations to AI-driven communication as they do to human channels. Practically this means using data only for the consented purpose, disclosing that an AI is assisting, offering a route to a human agent, and logging every AI data access for audit.
RiskWhat are the biggest risks of deploying an AI billing assistant?
Three risks matter most. The first is ungrounded answers: a model reasoning about a bill rather than reading it will invent explanations, which in billing is a regulatory and trust problem, not an inconvenience. Grounding in real account data and vetted support content is the mitigation. The second is silent abandonment, because deflection metrics cannot distinguish a customer who was helped from one who gave up, so containment must be reported with repeat contact rate and satisfaction. The third is retrofitted governance: privacy assessments, data residency and logging shape the architecture, and adding them after a pilot is why programmes stall.
OperationalWhat KPIs should telecoms track in an AI billing pilot?
Track four categories rather than one headline number. Deflection and containment covers call deflection rate, digital containment and the share of billing queries resolved without a human. Efficiency covers average resolution time, handling time for escalated interactions, and cost to serve per billing contact. Experience covers satisfaction on billing interactions specifically, repeat contact rate within seven days, and portal usage. Assurance covers escalation rate, accuracy sampling on the assistant’s answers, and the share of interactions with complete logs for dispute or audit. Reporting all four prevents an operator scaling a metric rather than an outcome.
Legacy telecom billing is the largest untreated source of customer friction in the industry: it drives up to 50 percent of contact centre calls, and Canada’s Commission for Complaints for Telecom-television Services reported billing as 46 percent of all issues raised across a record 23,647 complaints in the year to July 2025, with incorrect charges and missing credits the largest share. The evidence runs in two directions. TM Forum’s Business Assurance Survey for 2024/25 shows revenue leakage falling to 0.52 percent of total revenue from 1.22 percent in 2021, so accuracy has improved while comprehensibility has not. Generative AI addresses the explanation problem without replacing the core. The four-step playbook is: AI integration strategies using custom APIs, middleware and knowledge base connectors against legacy BSS, OSS and CRM; feedback loops with human review and performance monitoring; a technology stack combining large language models, machine learning personalisation and responsible AI guardrails, noting roughly 40 percent of telecom firms prefer on-premises deployment; and regulatory compliance covering CCPA, PIPEDA, CPNI, the Government of Canada’s Voluntary Code of Conduct for Generative AI, and FCC and CRTC obligations. Published results include TELUS enabling 28 percent more successful self-service, Verizon reporting a 40 percent increase in service channel sales across 28,000 agents, and Vodafone cutting wait times by over 50 percent.