Want to AI-upgrade your customer service? Read this.

A year ago, as GenAI Lead for EMEA at Infosys, my team and I implemented a GPT-4 AI system to handle complex client emails for a global corporation.

It used to take more than 4 hours for:

It now takes less than 1 min. by implementing AI to:

CR agrees to the answer and sends the email to the client.

If needed, the CR can edit the AI's response.

Under the hood, much more has happened.

A retrieval-augmented Generatio (RAG) architecture connects assets (knowledge, databases, systems) to retrieve relevant information.

To make it work, you need also …

• … an LLM (e.g., GPT-4, Claude 3.5 Sonnet) that we interact with to answer the emails for us. We give it the aforementioned relevant context.

• … prompt engineering. The LLM needs to know what we want from it.

• … a cloud (e.g., Azure, AWS) connecting all components and making them run.

But there is nothing impossible, not done before, nor black magic.

Since that early project, I have accompanied many more.

My All-star Team

I have cherry-picked a superstar AI implementation team, and you can book us so we can implement your AI customer service system.

We don’t have managerial overhead or a whole research arm to finance. Value for money only + only high-quality outcomes matter.

What are the right use cases to start? Not all are juicy, low-hanging fruit.

For instance, product feedback from clients. You can use AI to summarize, sort, and slice & dice, but you want a person to be aware of the feedback.

The ideation and evaluation of use cases are multifaceted.

You want to look at ROI calculations (resources saved), scalability, user experience enhancement, alignment with strategic goals, and more.

Because it can be tricky to decide which step to do first (or next) with AI, we have built the GenAI Accelerator Workshop.

Here is a snippet from a former interview with Steven Ramirez in preparation for the GAIAS conference.

My team and I have seen/worked on many different use cases in CS

• Chatbots via LLMs can do many things for your organization - from updating information to sending client forms.

• Inform the internal churn score by scanning email communication and customer behavior thoroughly.

• Instant multilingual support.

• Visual analysis of documents sent in attachments, scanning contract numbers via LLM-internal OCR.

• Proactive customer support: analyze data to remind customers of renewals, maintenance, or upgrades.

Most underrated use case? AI customer calls.

Not many know this, but AI voices are getting fast. Groq, the fastest LLM, and Deepgram, the fastest text-to-speech model, are on par with human reaction time <250 ms.

AI voices are also getting authentic, sounding 98.6% human.

Done well, you can outsource up to 90-95% of all calls with your customers to an AI. The other 5-10% can be taken by a customer representative.

By now, I have probably evaluated all of the text-to-speech models. While it depends (e.g., what language? Speed, or quality?), I think Cartesia is doing an outstanding job. It's extremely fast and high-quality. Try it. (No affiliation, just my opinion.)

And today, we have the weakest AI in the future. 

An AI voice will be so good that you can speak in a customer’s accent and even dialect to sound familiar and help customers understand better.

Don't be misled; here's my advice

Building these solutions requires an initial architecture - the base architecture.

Having that, you can build various use cases on top. The ratio is roughly 80-90% of the costs for the base architecture and 10-20% for the use case, such as starting a claims process at an insurance company or assisting with transactions at a bank.

⚠️ Don’t get tricked by other consultancies that try to sell each use case as a separate project! ⚠️

After the base architecture is built, you can throw many relevant use cases on top.

If unsure about AI in CS, consider that AI has:

The AI strategy should be such that it moves up the value chain over time.

Identifying your AI strategy is part of our GenAI Accelerator Workshop.

Not everything is perfect out of the box; it has to be engineered. However, if done well, AI speaks in your brand voice and can even have a ‘hidden agenda’ of upselling or cross-selling.

To do this, you need to master prompt engineering. We teach professional prompt engineering techniques (not what you can find on every second YouTube channel about AI) as a standard project handover.

Speaking of interaction with clients

Of course, you don’t want the AI to step out of line. No bias, no discrimination. 

• We have a proven and structured approach to thoroughly stress-test the AI over weeks.

You might also ask, How do you ensure the security and privacy of our sensitive data and clients' personal information when using cloud services and AI models?

• There is a lot more to this. However, we would set up a secure cloud environment, respecting Robust Security, Data Encryption, Compliance with Regulations, Identity and Access Management (IAM), and Secure Configuration and Best Practices.

• Further, the AI model is chosen so that its source is trustworthy, one can get an own AI model instance, your data is not being used for training the model, and other non-technical and technical requirements.

Lastly, what about liability? Who’s fault is it when things go south?

• Do you remember that in my initial anecdote, the generated answer needed to be cross-checked by someone?

• This is designed so that the liability sits with the person sending the email.

We suggest a 3-step rollout of the AI system in CS (depending on the use case):

1. Build a PoC and test it on historical data. Have a confidence score of the AI. KPIs reached? Next step.

2. Pilot Phase: Let the AI answer customer requests but write them in a document. An employee scores the answers. Analyze data to identify AI rollout patterns and verify confidence score reliability.

3. Production: Let AI answer each use case until the pattern is fully automated or it can reliably be forwarded to a colleague.

We have many more thoughts on the confidence score, but that is for another post.

Or refer us and get a lucrative bonus. 😉