Simple and proven ways to implement AI in your CX function right now

Data security review

Almost all the advanced and widely used AI models like GPT-4, Claude, and Gemini are hosted in the cloud, so most fear a dreadful data leak or unauthorized use of their company’s internal data for training these models. These concerns are mostly unwarranted, like spotting a ghost in a countryside graveyard. None of these big three players use our data to train their models as they are already pre-trained with public internet data. In addition, also look out for vendors who are SoC 2 certified and provide data encryption to eliminate any sort of data security concerns.

So the first step is to invite your infosec team/CTO to review AI vendor’s data privacy declaration, infra designs, SoC 2 certifications, and data encryption to give them confidence about data security and most importantly get their vote of confidence in return to get started.

Custom knowledge training

All these AI models are pre-trained on varied and extensive datasets- books, internet data, and other hundreds of gigabytes of data, and are extremely capable of understanding patterns and nuances of our general language. But here is the catch, they literally possess zero contextual knowledge about your own company’s product/services. Think of an intelligent outer space Alien entering Earth with zero understanding of our Earth and its creatures.

So the next step is to custom train these general-purpose AI models using your own company knowledge that exists in help center articles, websites, PDFs, excels, past customer conversations etc. So that you leverage the intelligence of these general-purpose models while imparting memory of your own company data also called as ‘RAG’ technically.

One important thing to remember here is “Garbage in is Garbage out”, so make sure you train only on updated and authorized data sources to start with to get accurate and helpful answers from your custom-trained AI.

‍

AI quality testing

Let me share an open secret about AI: Current AI isn't a creative, logical, rational thinking system like Einstein's brain. It's more of a probabilistic, pattern recognition, non-thinking system, only as good as the data it's trained on. So more and better the training data set, the better the AI performance.

Now, once the AI is custom-trained, test the AI quality by collating a list of product/service queries/ past customer queries, the more and varied the list is, the better the testing is. Somewhere around 50-100 queries is good enough for the evaluation stage.

As this step is the difference between a demo project and a production-ready project, try to recruit a child-like AI enthusiast as well who also has a Steve Jobs like eye-for-detail in your CX team to lead this evaluation step. Finally, encourage the users who are evaluating to upvote/downvote the AI responses helping to generate an objective data report of AI accuracy and training quality score(% queries for which data is available in the training dataset) at the end of the evaluation. Accuracy rate of ~ 90% and training quality score of ~ 80% is a good indication for production readiness.

Deploying AI to your team and customers

Big congrats to everyone who has reached this step! This is definitely the most exciting and anxious part as the real quality and value of AI will be tested, similar to a kid taking final exams. It helps to keep anxiety at bay and start small by deploying AI within your teams internally, first in Slack or MS Teams channels. This way, the team can interact with AI, ask queries, and validate answers.

Once the team is confident about the accuracy and utility, you can add the AI as a web widget to your website or product. AI agents can serve as intelligent, instant, 24/7 assistants for your customers and team, performing 10 times better than traditional rule-based bots. Finally, since AI isn't 100% accurate, ensure there is a way to hand off to a real human with flesh and blood for handling complex queries when a user downvotes or explicitly requests to connect to a human.

‍

Continuous monitoring

Just like a parent who won’t leave a teenager unattended however intelligent and ambitious they are, AI cannot be deployed and forgotten. Have periodic biweekly/monthly checks on its performance and adoption - unique users, queries answered, accuracy, training gap, etc. Identify teammates who haven’t adopted AI yet nudging them to start using it, make the AI Agent chat widget discoverable on your website, find topics in which AI is underperforming to update the training data to continuously keep improving the AI performance and adoption.

Evaluate Impact

Unless you are a cute-looking toddler, no one gives a damn for just trying (wish our human civilization gets better at this 🤗 ), any new tech is as good as the real value it produces.

Now is the time to measure AI’s impact by measuring outcomes like queries solved, users engaged, time saved, etc. to validate the real AI impact in supporting and managing customers and also continuously communicate to the entire company about the value AI has added especially in terms of instant, intelligent, and contextual support to our customers, unlike traditional scripted, boring, and rule-based bots.

‍

Final thoughts

Companies who want to implement AI > # Companies who are evaluating AI > # Companies who have successfully implemented AI - is the current state right now. My main aim in writing this long article with amateurish humor is to demystify applying AI in CX functions. AI has definitely turned out to be a once-in-a-generation technology and there is a huge opportunity for businesses to serve their customers in a much better way than ever before.

Curious to know more about how to implement AI agents in your CX function- Customer Support/Success/ Onboarding/Presales/ PS, Book A Demo with us, we certainly promise an interesting and passionate conversation about implementing AI at your company without sounding like a buttoned-up salesperson :)