Real-time Personalized Recommendations

Customer experience personalization is all about data first. Get the data right and you can shape the overall customer experience by applying insights learning. Recommendation engines are very powerful personalization tools because it’s a great way to discover what products they like, what their preferences are and so on. They help in enhancing visitor’s experience by offering relevant items at the right time and on the right page.

The importance of suggesting the right item to the right user can be gauged by the fact that 35% of all sales in Amazon are estimated to be generated by the recommendation engine. The core strategy accounts for more revenue per customer through predicting recommendations that a user is likely to buy, which they might not be aware before. Continue reading →


Data Foundation for AI Implementation

AI is a simple science of being able to decide and carry out functions in real-time just like how human workers do

Just like the way human brain makes decisions, systems that can create a memory of patterns can quickly detect/recognize patterns from past memories, allowing them to make decisions which can be translated to predictive decisions, recommendations or the shortest path to the goal. Based on the status of these states, the system fires the next appropriate function, which could be an engagement with a customer or an insight to a marketer.

This simple mechanism allows the user to observe every step of the learning process, keeping the system transparent and agile. Without transparency, your AI investments could be a very risky proposition Continue reading →

Remembering past interactions is critical for engagement automation

How does the customer engagement system or the marketing automation platform know the past behavior or preferences of your customer? Does it have a memory? Is it making decisions based on past experiences?

To initiate true AI driven customer engagement, it is important that the system learn and create a memory file around each customer. This has to be constantly updated after every interaction so that memory consolidation and references are maintained, and accessible to any touch-point application in runtime.

The memory files will allow the machine to truly behave like a human: exhibiting relevance and context in its communication and driving customers towards their goal. These memories can be easily integrated with customer engagement applications to make them aware of customer states and preferences; helping them deliver highly effective communication and reducing the time to revenue.

This is exactly how Plumb5’s Memory Stack is designed to perform real-time contextual customer engagement. By ingesting individual customer data (created by the unification feature), Plumb5 extracts spatial and temporal patterns to arrive at states based on weight attribution, creating individualized memories. It also checks for similarities in other individual’s memories and creates collective memory states, before creating/updating the final memory file(output file).

The Plumb5 Unified Customer Stack, which forms the foundation to these memories, collects data from all data sources or data lake or data warehouse and creates a single customer file based on tags and unique identifiers. The Unified Stack, which is nothing but a centralized data repository of all individual customers (i.e., a transformed and harmonized dataset), has all the data parameters required to analyze/learn or make decisions.

The data repository contains each customer’s explicit or implicit data collected through touch-point applications. This individual stack, when extracted, presents the relationship between the customer and all other input parameters spanning products, channels, campaigns, feedback and other. On further extraction, we can find a relationship between the customer and associated factors from product interactions, or channel responses or preferences associated with the customer id, forming spatial relationships.

These relationships are converted to a set of tags with their own weights, which is used to understand the importance of the relationships in order to to prioritize the next best action. The weights can also be used to arrive at the nearest neighbor for firing a new event For the purpose of analysis, one can cluster customers based on these tags and their weights

This data, when laid out sequentially by time, presents the temporal pattern of the individual: allowing one to predict the next probable direction of the customer.


Both spatial and temporal relationship weights are combined to arrive at the net weight, which updates the final individual memory states. These memory states or patterns of a single individual are compared with other individual stacks and patterns to arrive at collective memory states; important for predicting recommendations or probable next actions based on similar profiles.

All of these states are consolidated into a single string (JSON) as memory states. On the collection of data from data sources, these memory states are updated at runtime using continuous-time Markov decision processes, allowing the string to keep updated information at all times. These strings can be consumed by customer engagement platforms for real-time contextual and relevant engagement.

These memory files will make it simpler for the proliferating customer touch-point space. With more businesses wanting to track, collect or engage customers using IoT devices, data needs to be organized for quick extraction of information and for runtime analysis and decision generation. The memory file will make it simpler for any such device that requires it to present quick relevant communications.

Currently, the robots deployed for customer service have no consolidated data about a customer and merely serve to answer generic questions. To take robotic customer service to the next level, these robots will need to have access to memories to make customer service far more contextual and relevant.

These memory files can also be integrated with NLP tags, where contextual conversation can be carried out at runtime, exhibiting to a complete AI based contextual experience for every customer or visitor.


Optimizing Digital Ad Spends

In order to measure ROI on Ad Spends(ROAS), it is important to track the entire journey of the ad visitor up to the point of purchase. This could happen either on their first visit or during subsequent visits. Measuring conversions, or measuring interest generated by a particular ad, is dependent on the interactions that happen post clicking the ad. By filtering out bot traffic, one can identify the right channel that drives visitors with intent. Continue reading →

How Intent States work in Plumb5?

Intent Scoring helps in understanding customer’s possible intent towards a product, which can be used in personalizing content or offers based on his intent state. The state can be identified by comparing different behavioral parameters.

These states help the automation engine to pick up relevant messaging and engage the customer automatically. Intent states are used between each stage of the customer lifecycle in order to understand user’s intent to move towards the next stage.

These intent states can be applied to any group created by you to understand customer’s or prospect’s intent within the group. You could quickly visualize your segments by their intent states and roll out a campaign based on their intent


The intent scale range is easily configurable. For experimenting marketers, the scoring module allows users to assign their own scores and come up with states using score-ranges.

However, for marketers with no exposure to scoring, they can use the inbuilt algorithm which clusters in real-time arriving at states by using ranges from different parameters.

Plumb5 Unified Customer Model allows the system to easily arrive at intent states for each customer. The algorithm would select each customer data and extract session-level data to get unique pages to total page ratio which is further compared to the recency weights to arrive at the final state.

To explain how the algorithm works, for each customer the algorithm creates a unique variable to understand visitor’s intent by computing total interactions over total sessions.

The resulting output is used to find the first level intent state. For example: If the resulting output is 2.1, then it would push the customer to state-medium based on the following table

If the variable is <2 – Assign State – Low
If the variable is >2 and < 3 – Assign State – Medium
If the variable is >3 – High
The algorithm holds this state and checks for unique interaction count. If the unique interaction count is lower than the desired threshold, there might be a change in the intended state.
If State is Medium and Unique Interaction Count is < 2, change state to Low
If State is High and Unique Interaction Count is < 2, change state to Low
If State is High and Unique Interaction Count is < 3, change state to Medium

These updated states are checked for Recency count. If the last interaction of the visitor has been more than 8 days, then there is a possibility that the states will be updated again.

If State is Medium and Recency Count is >8 then change state to Low
If State is High and Recency Count>8 and <14, then change state to Medium
If State is High and Recency Count>14, then change state to Low

With this, the final intent state is arrived and updated against the user.

Though the algorithm is currently configured for web and app behavior data, with the integration of data containing written feedback or voice data, we can gather the sentiment associated with the latest interaction and use that to change states as well to bring in more context to personalized campaigns.

The algorithm derives the range scale by using the min-max range of the given parameter. The min-max range is arrived by comparing against all user interaction counts. For example, for a particular user, Unique interaction count is 2 and for another user, the interaction count can be 155.  Using least count and highest count, it creates a range scale that is used in comparing individual interaction count for state derivation.

The states can be configured based on the requirements of the marketing team. The algorithm uses a standard three-stage logic (High-Medium-Low) to break the range scale to states. The states can be configured by just specifying the total states needed for your requirement and the algorithm automatically classifies the range scale into the specified state ranges.

This allows in computing intent states in real-time and allows marketers to automate targeting campaigns with higher effectiveness and greater customer satisfaction.

Everybody knows silos are bad for business

It is evident that 99% of businesses, irrespective of big or small, is bleeding heavily due to data silos. The current environment demands huge dollars to be spent in order to maintain data, which covers paying for redundant data and isolated data due to data sources that don’t provide a common unique parameter for unification. Data grows on a daily basis and there is no escape from the cost incurred on growing redundancies and data silos. This sub-optimal approach not only costs a lot of money but also create gaps in achieving the desired results.

Problems associated with data silos can be many. For example, if you are facing the problem of higher customer acquisition costs, you can see that this is due to (a) inaccurate attribution models leading to ineffective ad spends and (b) lack of customer engagement or irrelevant targeting. Both these problems are resultants of data available across many data sources and lack of unique tags to merge them. So, the analysis carried out over individual data-sets results in wrong analysis, which results in low effectiveness and wasted resources. Continue reading →