Login
Register
Sponsored
Himanshu Polavarapu is AVP – Tech Strategy and Enterprise Architecture at Verizon. In this interview with TM Forum, he provides fascinating insights into the operator’s progress with Agentic AI for customers, employees and the network. He explains priorities, approaches, lessons learned and possible next steps.
Himanshu Polavarapu
Sponsored by:
How Verizon’s work on Agentic AI boosts experience for customers, employees and the network
At Innovate Americas, Himanshu Polavarapu (HP), Associate Vice-President, Tech Strategy and Enterprise Architecture at Verizon, will be in conversation with Sharad Sachdev, Global Lead for Gen AI-Powered Customer Experience at Accenture. Their topic is From vision to value: How Verizon is operationalizing GenAI at enterprise scale, a key part of the opening Ready, Set, Scale session which starts at 11am on September 10. This interview with TM Forum sets the stage.
HP: For us it’s always been a productivity and efficiency play with big agent ecosystems focused on customer, network and employee experiences.
In customer experience our platform is the digital touch points or IVR and interaction with an agent for sales or service. We developed and deployed chatbots specifically for those chat experiences, to address and anticipate that customer’s need based on our information about them and their behavior to reach that point of engagement.
We looked at the tools frontline agents need to serve customers and now they have a dedicated personal research assistant (PRA) so that they don't have to trawl knowledge articles to figure out the best offer or solution for that customer or the best explanation for billing fees. They ask their PRA.
We deploy the same agent ecosystem in our developer tools to standardize tools across our developer community – with orchestration behind the scenes. Developers need more than one agent because they could work on anything from a serving requirement to developing code for testing and deploying code.
HP: The first thing we needed to understand was how to achieve efficacy – agents’ performance depends on how much and the type of information they receive. We went through three big iterations to get the PRA agent to the highest efficiency.
We also quickly identified a learning gap and realized that to produce the right prompt, we need to translate what is intended by the questioner so the agent understands. This meant we needed to educate the entire user base, which we’ll come back to.
Verizon measure PRAs’ success with tracking metrics for their answerability, accuracy and efficacy, which relies heavily on a T&E ( testing and evaluation) platform that Verizon built to keep agents up to scratch. We found their efficacy slips unless they receive updated information daily. Currently the PRA is serving a lot of front-end employees and deployed at scale for the whole user base.
HP: We had initial productivity gains from code-assist tools but quickly pivoted to transforming the whole software development lifecycle (SDLC), starting with standardizing the process of how we perform SDLC across different organizations.
Next we provided tools for all the user personas: for example, the business persona who gives requirements to the architect designing a solution; the developer working on the solution; and the solution’s tester. A big gain is that the business product owner no longer needs a four-hour call to explain their requirements with 20 or 30 user stories. Now the business owner has an AI tool – they come up with one simple requirement and AI generates the user stories.
These stories can be complex, such as looking at what applications are impacted, the required design elements and sequence flows, and appropriate test scenarios we need to develop. This has brought efficiency and productivity to every user persona.
Third, developers receive a predefined artefact of changes; they are moving from primarily producing code to reviewing it because the code snippets are pre-generated. If developers accept what they receive, it is automatically generated.
We home in on three big KPIs – speed, quality and cost optimization.
HP: Yes. We’ve talked about customer and employee experience, but if network experience is poor, customer experience automatically takes a hit first. Verizon’s network needs many people to run it and this workforce engages with the network through assets in teams like planning, engineering, system performance and service assurance. To leverage AI here, we analyze what they do, day to day, to see where we can help and make them more productive.
This workforce needed something similar to the PRA, a predictive model, for instance to warn if a cell will become overloaded and suggest the best reroute for traffic without impacting customers. We created the ‘Network Genie’ so network staff can ask what to do in a certain situation. The Genie engine correlates what’s happened and offers a recommendation.
Verizon also has a product that readies an end-to-end orchestration for an employee to solve an issue and if the user accepts it, automatically deploys. This is our strategy: a tiered competency approach within the network journey that invokes different capabilities in different scenarios.
HP: Verizon had four data lakes holding customers’ data, corporate function data, business data and petabytes of network data. The lakes were on-prem and various public clouds. We brought them together under the Verizon Data Program, conceived by our CIO to bring many benefits as well as enabling AI. The program has another 18 months to run.
We realized we only needed specific data sets for AI, delivered at the right time. We addressed this by curating appropriate data into ‘data products’ just for AI.
To deal with the huge volumes of unstructured data – images, PDFs, Excel spreadsheets, Word documents and more – dispersed across the company, we created One Functionality which is an internal knowledge service. It ingests unstructured data and generates agent- or LLM-ready data. This is when we started seeing the best outcomes from AI agents and the exercise is ongoing but I’m confident we have the right platform – data ingestion pipelines, if you will.
HP: At the start with GenerativeAI, everybody had use cases they wanted to deploy so we
set up an AI Council to prioritize and put a responsible AI framework in place – a tenet of our approach. The Council is drawn from business partners and product owners across the organization.
We honed use cases down from about 140 to our top 20, using return on investment as the first element in the process. If approved, a guide created by the Council supplies the appropriate design packet for that particular scenario. Verizon uses a closed loop governance model in this second part of the process, first running a proof of concept, then deployment at scale. We measure who it benefited, how and the ROI.
Regarding people, we have upskilled user personas across all the hierarchical levels starting with master classes and C-suite. We will continue with more exercises at the working folk level as part of our dedicated, intentional upskill initiative. Education will be ongoing because the entire space constantly evolves. A case in point is that we already have a team in place looking beyond Agentic AI, working on what the model content protocol (MCP) means to us and how we could we deploy it responsibly.
Register for Innovate Americas here.
