When machines are in charge of making real-time analytics-based decisions for better customer experiences, accuracy and consistency of data with up-to-date contextual and behavioral information are very important. But is that sufficient?
The Telecom Scale
With the enormous growth in internet data usage over the past few years, modern telecom marketing analytics systems need to munch real-time datasets such as deep packet inspection for browsing alongside ever-changing locations and other streaming information at rates northwards of 1 million events per second. Telcos need to reach out to customers on the right channel at the right time using the right offers before the context is lost, as relevance holds the key to achieving the desired outcomes. Typically, promotional messages are piggybacked to events like post-call notifications, balance checks and other informational API accesses initiated by the end customer, and in large telcos, this could easily translate to an ask for over 30,000 to 40,000 decisions per second.
The Engineering Challenges
Business intelligence systems that operate at such scale are distributed and replicated by design with decisioning logic evaluated at one of the many peers based on load balancing algorithms with continuous state synchronization between the peers. However, given practical considerations around network latencies, system faults, inherent concurrency among different events, data streaming challenges and asynchronous transfers, decisions taken for a subscriber at two different peers can, at times, end up being different owing to different states, and there may not be a canonically correct way for the conflicting peers to arrive at a common plane using well-understood CRDT (conflict-free/convergent/commutative replicated data types) techniques.
A Closer Look At Reality
Consider that Bob, who has a $2 balance on his sim card while surfing Facebook in the background, drops a message to Dick to recharge for $25 and also makes a call to Alice requesting her to top up $15, and they both oblige nearly instantly. Let’s assume there are two event interceptors, call termination and Facebook browsing at low balance, and the former sees $15 recharge first while the latter sees $25, and they both send two different offers to Bob reflecting the best recommendation from their standpoints — only to realize in a while on receiving the next recharge that they both are wrong and when they try to sync up with each other, they can neither defend their recommendation nor accept the one received from the peer. An engineering solution can help them reconcile on the balance, and a technique like “last writer wins” may crudely choose one of the two offers, but a business solution needs more than that as it puts the end customer at the center of focus.
The Goal
A business-centric system needs to work a way out in the interest of the end customer, and in line with “N=1, R=G,” the calls taken may be different from case to case. It may involve soliciting both offers with no recall so as to not confuse the customer, or it may be to retain better of the two offers than going for an altogether new one, or it may even be to send a polite on-demand explanation revoking both of the offers in lieu of another more personalized offer. This decision should be taken at the level of the individual customer.
The business end of real-time personalization is where context-aware, user-centric resolution takes precedence over preset mechanized norms that may be acceptable in certain areas of data engineering but fall short in realizing the objective of retaining the best interests of every single customer even at scale. Ensuring accuracy and eventual consistency of data by itself is not sufficient in meeting the goal. An enterprise has to go beyond that to stay relevant and offer the highest value to its end customers.
How Telcos Can Get There
- Reassess the offered recommendations under real-time interaction management (RTIM) to check if they continue to be in the best interests of the customer as they transit through the micro-moments. Steer away from the “arrows fired” mindset, and don’t hesitate to recall grossly unfavorable ones only for the fear of confusing or getting misjudged by the customer. When conveyed the right way, they should be happy to see how far you have walked out of the way only to offer them the best outcome.
- Invest in technology that can help resolve real-time decisioning conflicts by plugging in user-centric business logic and let enterprises work with practical challenges around asynchronous, partial and out-of-order information delivery. Shy away from vendors who say their algorithms work only when you get them all of the data points in the right order. A proper solution should be able to learn from mistakes in real time and offer better alternatives on the go.
- Close the feedback loop and validate the intended value-add by running in-place surveys around the micro-moments of your intervention. Only ask pointed questions and seek their satisfaction levels around your personalization in that particular context. Instead of asking customers how they would rate your service, it is better to ask something like: “Did you find our ‘Evening 5GB social media on Tube Train every week’ offer useful?”
Telcos need to rise above the personalization limits of classical data engineering and work toward sustained customer-centricity even beyond the moments of intervention by staying open to real-time course corrections at scale in order to preserve the best interests of end users.
This article originally appeared here
