How Klika Transformed AI CRM Into an Award-Winning Product

Challenges

Although our first goal was to have users do less with automated data entry, a slow transition to AI was introduced. We knew using AI could be a double-edged sword, so every aspect of its integration was carefully thought out.

One of our first challenges was integrating AI intuitively into the existing features. Since AI integrations are a significant technological trend today, we avoided jumping on the wagon without double-thinking about their usage capabilities. We needed to ensure AI capabilities are part of a holistic system where everything is intentional and never an afterthought.

We needed to test the new features to ensure they genuinely help the end user and our client by being cost-effective, easily testable, and editable.

The last challenge was ensuring the features' compatibility and ability to be used in all environments.  

Klika Solution

We tackled the challenges individually, ensuring that the integral part of the client’s original system stayed fully operational as work on new features progressed.

AI features were gradually integrated into an existing feature set, such as generating an email based on user input or recording a previous conversation or email interaction. We have used OpenAI chat completions with functions to achieve this functionality.

Using Twilio VoIP APIs, AWS, Deepgram, and Google Speech-To-Text, Klika implemented a solution where the internal VoIP system calls are recorded, transcribed, and summarised within the app. This allows users to save hours monthly by removing the need to summarize calls manually. With the addition of the system, users quickly remember how a call went and see the next steps they promised or needed to take.

Klika integrated a chatbot on every customer page in the app. Users can ask any question about the customer, and the AI will answer it and provide real context and accurate information from the database.

When we started implementing a chatbot, we envisioned it to serve as a personal assistant for every client. The reason for its implementation was to allow users of our client’s software to have an overhead view of operations happening within their big businesses. By inspecting each business profile, users use chatbots to find answers to general questions that usually take a lot of time to conclude from the data if they were searched and analyzed manually.

This required us to store much of this data using vector storage and handle structured (SQL-type) data within vector storage, which generally works well with unstructured data such as blogs and books. LLMs are generally very good at understanding and recognizing patterns. Combining this with the actual data (context) we provided, users can get better value than trying to understand a bunch of data themselves.

To make things more flexible, if the AI can’t provide the user with an answer, we built a small DSL for the AI to request additional information from our backend. This has allowed us to build up AI’s knowledge data of the customer as time progresses.

By utilizing this process, the AI can now answer questions about the customer without even looking into our standard SQL database, which makes it much faster than a simple querying process. It does not need to look into the SQL because the AI assistant has all the customer data cached in a file. This file is additionally vectorized, so the AI can use vector search instead of requesting the data from our API through an SQL search.

These new features required us to teach the general-purpose AI to be aware of the customer's context, meaning we did a lot of RAG (retrieval augmented generation) to enhance the prompts and tell the AI what’s happening.

Initially, we tried passing a suggested answer based on a deterministic approach, such as a FAQ. We moved on from this approach because users ask so many things, and we wanted higher precision.

We ended up assigning an OpenAI assistant thread to each account. This thread inherits data from the base assistant set up for a more detailed and precise chat. The assistant is loaded with a lot of data about the customer, which is vectorized, with the instruction to the AI to prepare a wide array of questions and answers that the user can potentially ask. If the question is not within the list of AI-generated questions, AI will generate the answer and save it for future knowledge base.  

Results

Users reported that Klika’s system upgrades significantly impacted their work process positively. Teams have reported having a much simpler time understanding tracking client communication, a more straightforward process of onboarding and offboarding employees, and a gentle learning curve.

The Ask AI feature was especially praised for bringing newfound knowledge to teams due to AI’s ability to make sense of a large input of data, which people still struggle with, especially in terms of processing speed and results accuracy.

The benefits of reducing or eliminating various time-consuming activities like writing summaries and other tasks related to CRM administration were also reported. Sales personnel reported a much-needed shift of focus on their clients instead of on tedious administrative tasks, which allowed them to close the deals in a shorter period.  

Technology stack

Ruby on Rails, React, iOS Objective C, Java Android, DOX for auto-documentation, OAuth2 / JWT, Elasticsearch, Redis, Postgres, Heroku, Circle CI, AWS Beanstalk, Azure AD, Azure Intune, Google API, Microsoft Graph API, Google Cloud - Speech to text, Python, Tensor Flow