WHITEPAPER : [S2-AIGAI] Telco AI Call Center Assistant : Simulating the resolution of [Roaming bill dispute] with AI Agent

The article is structured into following sections

1. Let’s talk about some of AI Agents, which can assist in potential telecom functions / processes

2. Let’s understand the reference architecture supporting telecom services & operations

2.1. Overall platform architecture with [AI Agent Layer]

2.2. AI Agent Framework Architecture

3. Explanation to AI Agent Framework Architecture

4. Conclusion

1. Let’s talk about some of AI Agents, which can assist in potential telecom functions / processes

Following are just very few examples to leverage [Telecom AI Agent] in various telco functions/processes.

1. AI Agent for NLP based queries of Telecom function or process (Prepaid, Postpaid, OTT, Roaming, Data, etc)
2. AI Agent for KYC Process
3. AI Agent for Waiver & Adjustments
4. AI Agent for KPI monitoring
5. AI Agent for Ownership Change
6. AI Agent for Change SIM/IMSI
7. AI Agent for Network Issues & Action
8. AI Agent for Credit Control & Payment Collection
9. AI Agent for Fraud Detection & Action
10. AI Agent for Error Analysis & Correction
11. AI Agent for Billing Dispute Handling (Local & Roam)
12. AI Agent for Service Termination
13. and many more ..

1.1. AI Agent for NLP based queries of Telecom function or process (Prepaid, Postpaid, OTT, Roaming, Data, etc)

Scenario : several prospects or existing customers want to understand about Telecom operator’s offering as per their usage purpose e.g. Head of the family asked the query via chat or voice AI Agent.

Example-1 : I am a student of [AI International University] situated in New Delhi, I take the metro from [AKV station], during metro journey I generally listen to subject related podcasts. I use a lot of internet for academic research purposes, I need to download / upload assignments and responses on the university portal. I need to make lots of unplanned calls to my friends & family for personal & academic discussions. Over the weekend (long hours) or occasionally on weekdays I access the social-media and digital channels for social & current news awareness.?

Please suggest the best plan from your offerings to fulfill my needs.

Example-2 : I need a plan which can cater my entire family for unlimited voice, SMS, data, OTT and social-media needs. We are 4 family members including my wife & 2 kids doing academics UG & PG. I also want to include 1 number for parents and 1 number for 1 in-laws.

Please suggest the best plan from your offerings to fulfill my needs.

1.2. AI Agent for KYC Process

Scenario : [Know your customer (KYC)] is the mandatory regulatory process required for customer acquisition. The KYC process includes customer biometric verification & address verification. The KYC process also includes document collection, scanning, digitization, and storing into a repository. These documents are customer acquisition form (CAF), proof of identity (PoI) & proof of address (PoA). Examples of PoI/PoA documents are UID (Unique Identity), Passport, Driving license etc.

Important point to note here is that customer physical verification of KYC process has improved from 8+ days to a range of instant verification to 3 days i.e. if the customer is not verified physically within a defined SLA period, telco services will be deactivated on the customer’s SIM. However, many of the governments have now cut down the paper based KYC to e-KYC process, which has brought the additional risk of SIM-frauds & cybercrimes.

Example-1 : [AI Agent for KYC] can bring the KYC process improvement for customer verification, can leverage telcos to prevent mobile-connections obtained by fraudster in convenience with point-of-sale (PoS), which are misused to commit cybercrimes/frauds.

1.3. AI Agent for Waiver & Adjustments

Telecom operators often provide waivers or adjustments to customers as part of their customer goodwill or satisfaction policies to maintain loyalty and build positive relationships. Below are some hypothetical examples categorized by typical scenarios:

Now let’s talk about how [AI Agent for Waiver & Adjustments] will assist telecom operators to speed up the [Waiver & Adjustments] process. Customer will be routed to [AI Agent for Waiver & Adjustments] and customer will explain the respective concern. [AI Agent for Waiver & Adjustments] will do the concern assessment by analyzing the relevant data fetched from the respective system through some API calls. Post assessment, the [AI Agent for Waiver & Adjustments] will share recommendations in the form of [Waiver & Adjustments], which shall be approved by concern authority.

1.4. AI Agent for resolving a customer issue

Scenario

Customer calls call center for roaming bill dispute

Scenario description

One subscriber has a billing dispute of 89 INR due to roaming outside India, while the fact is that he did not make any outgoing call nor received any incoming call while roaming. Apart from that he already had a foreign country local SIM for any local usage for the calls within? foreign country, so he firmly believed that there should not be any roaming bill component on my India number. So he decided to raise his concern to his home telecom operator, once he is back.?

Conversation history

Following is the conversation between customer and call-center user for his concern

In order to fetch the contextual data, following list of APIs would be invoked by [Retrieval Model] of [Telecom Agent AI]

2. Let’s understand the reference architecture supporting telecom services & operations

Below reference architecture diagram shows the standard telco functions at high-level, most of the components are well known to telco professionals. A new component [AI Agent Layer] in violet color is added to the platform architecture.

To my view [AI Agent Layer] is like [AI Agent In Middle] which will sit between Telco users & Telco Backend. Here Telco users can be of any type system users/events or human users including customers, partners & employees.?

[AI Agent Layer] with Direct database access (DDA) or via the Integration middleware layer.

It is observed that the industry is talking about implementing [AI Agent Layer] with direct database access (DDA) to fetch the contextual data and prepare the NLP response for users. Technically direct database access (DDA) is the doable approach, but not the best practices because for following reasons

• In industry some domains have really very large and complex IT ecosystems which have 100s or 1000s of applications. Allowing [AI Agent Layer] to access databases of all these applications will bring single point of security risk for entire ecosystem
• These IT platforms have evolved after huge efforts of several years, now they become highly mature and stable systems. Business will not be keen to disturb all running system just to introduce new technology
• These platforms have also evolved architecturally, they have microservice layer, observability layer, decoupled databases, and other technical layers. These architecture layers will remain the same best practices until they become completely irrelevant.

Sametime, people may find direct database access (DDA) fast & easy option for small size enterprises, which have a small IT ecosystem, which do not have complex infrastructure and a large ecosystem, who don’t want to implement the best architecture practices because of higher maintenance cost. So this should be well thought architectural decision to go with [Direct database access (DDA) or via the Integration middleware layer]

2.1. Overall platform architecture (Direct Database Access (DDA))

2.2. Overall platform architecture (Integration Middleware Layer)

3. AI Agent Framework Architecture

Description

• Analyzing the whole situation, it is found that there will be two types of queries

1. Queries which can be responded by already developed some capabilities e.g. API get-prepaid-balance() to get prepaid balance by MSISDN, select statements [SELECT <CDR fields> FROM <CDR DB tables> WHERE <subscriber-number=MSISDN>]
2. Other NLP based queries, where [Telecom AI Agent] will be leveraged to prepare the most relevant response for the NLP query. In case of hallucination, the user will be routed to designated human-agent. Responses will be combinations of structured & unstructured data.

• Query from user (customer) will be routed to request receiver [Load-balance, Integration Middleware] or [Telecom AI Agent] which is situated next to the conventional [Integration Middleware].
• [Telecom AI Agent] will analyze the NLP query, using it’s knowledge the [Telecom AI Agent] will decide that further contextual data is required or not to prepare the appropriate response.?
• Depending upon the need of contextual data, [Telecom AI Agent] communicates with [Retrieval Model], which will decide probable [task-list] using its groomed knowledge. The [task-list] will vary depending upon the customer’s NLP query. Here, point to note that the task-list is not the same decomposition-rule which is configured in the telecom order management system.?
• Based on tasks identified & groomed knowledge of telecom IT ecosystem, [Retrieval Model] will get the contextual data using [Integration API or direct Provider API] or using respective database query in case of data fetch from [Data store] or using API/DB call in case of unstructured data.
• [Telecom AI Agent] will consolidate (orchestrate) the original query & multiple contextual data and will pass the [Consolidated user information] to [Response Generator].
• Using the [Consolidated user information], the [Response Generator] will generate the NLP response which will be reviewed by [Response Reviewer].
• The [Response Reviewer] further can be supported by the [Telecom AI Agent] framework, where [Response Reviewer] will act as a [Software QA Engineer]. Here the idea is that [Consolidated user information] & [Response Generated] can be verified by executing relevant test-cases based on query & generated response.
• In case of hallucination, [Consolidated user information] & [Response Generated] will be forwarded to [Human Agent] and the user will be connected for the human voice conversation, else it would be [AI Agent] for voice conversation.

4. Conclusion

Thank you for reading the article. I will be happy to collaborate & discuss further on more scenarios and guide on the technical aspect. Based on my diversified experience, I believe that TELECOM is one of the business domains, which follows the best industry practices including decoupling & functionally segregated architecture. Being one the most complex domains, Telcos can be the leader of [AI Agentic Framework] for supporting their own customer services in the area of L2C, T2R & C2M journeys. On the architecture aspect, my view is that [AI Agentic Framework] the similar paradigm shift, what we saw in transforming from [monolith to microservice] i.e. mindset has to be tuned per say [AI Agentic Framework].

In this article, I wanted to cover user-journey and implementation-approach of each use-case in detail, and to keep the document size small, complete detail of [User-journeys & Implementation approach] is kept in separate document, please visit (Telco AI Call Center Assistant - User journey & Implementation Approach) for reference where it is explained that how [Telecom AI Agent] will integrate with existing API & database to fetch contextual data for [query - request - concern (QRC)] by user.

Sharing few reference link of source as follows, which inspired my thought process.?

1. TelecomGPT: A Framework to Build Telecom-Specific Large Language Models
2. AI Chat Agent, the game changer

About author

Profile : Rajesh Verma - Brief profile

Source : link for this article here

Series : S2 (Artificial Intelligence)

Episode : S2E2 (WHITEPAPER : [S2-AIGAI] Telco AI Call Center Assistant)

Author’s approach : Rajesh wants to share his learning & experience gained throughout his career from various sources. Author started the series on architectural topics including AI/ML & GAI topics and this article is one of the episodes in that attempt. Author feels that lots of information is available on various forums, but scattered here & there. Episodes in this series will be designed for most of the relevant topics in architecture-&-design, published gradually and organized in logical sequence. Principally episodes will have linkage with other episodes, so that readers can have proper connection among the topics and would be able to correlate with ongoing activities in their software life. Topics for example will be related to functional architecture, integration architecture, deployment architecture, microscopic view of mostly architecture-building-blocks (ABBs), security guidelines & approach to comply, performance KPIs & engineering, git branch & DevOps enabled automation strategy, NFR aspects (e.g. scalability, high-availability, stability, resiliency, etc.), commonly used architecture styles & design patterns, cloudification approaches, multi-tenancy approach, data migration, channel-cutover & rollout strategy, process standardization & simplification, greenfield rollout & brownfield transformation journeys, etc.

Thank you for reading the post, please stay connected.