Agentic AI is the latest buzzword doing the rounds in the world of Artificial Intelligence. This is a guide to what it means, and what the implications are for customer service leaders.
If you haven’t come across “agentic” as a word, it essentially means the same as autonomous. Wiktionary defines it as: that behaves like an agent: able to express or expressing agency or control on one's own behalf or on the behalf of another.
Therefore agentic AI means an artificial intelligence system that can act on its own and can control its own actions. When you hear agentic AI, think autonomous AI.
Obviously this sounds pretty futuristic even today with generative AI taking more and more of a role in our day to day lives. But there are more short term implications of it. We’ll use examples from customer service AI to illustrate.
How agentic AI would deal with customer service requests
Imagine that a customer asked “What materials are your trainers made from?”. In the past, if a chatbot was asked that it would have required training to recognise that specific question, and then a templated response would be given. This could be one of several templated responses. The point being that the AI would ‘understand’ the question and then trigger a programmed response.
Currently, with generative AI, it can be a bit more sophisticated. The AI still has to recognise the question, but it can then be trained to look for the answer to the question in the knowledge base. This means that the answer can always be “live” - i.e. if the knowledge base is updated, the chatbot does not need to be re-programmed. But crucially the AI needs to have been trained to only look in certain places to find the answer, so it doesn’t hallucinate.
So far, so good, but these are only helping with relatively simple question and answer responses.
What about if a customer asks the chatbot to do something, such as amend an order? This is possible currently if you have the deep integrations to the order management system required. Once the request is understood, then the AI can look up the order, and then overwrite it with the new changes.
This is already the beginnings of agentic AI, and it’s something that DigitalGenius customers such as Organic Basics will be familiar with.
True agentic AI would be that you could instruct an AI to build the flow and find the information for you, without having to build it yourself.
Agentic AI becomes self-improving
The example outlined still requires some flow mapping in most cases. Creating an if-then series of steps allows the retailer to put the AI down strictly defined paths.
The next step of AI is to stop strictly defining paths and allowing the AI to be guided by what it thinks is the best step.
The key mechanism here will be allowing the AI to understand what success is, so that it has a measure of what is better. That might be a CSAT score, or a similar metric which the AI can use to rank different actions. Actions and end-results that create a higher CSAT would be favoured.
The idea of an autonomous AI running wild with little oversight, making its own judgements and deciding things for itself is a little bit terrifying. It’s HAL from 2001: A Space Odyssey-level stuff.
But just as with current generative AI systems there would need to be guardrails, and lines that an agentic AI would not cross.
For example, the thing that would make customers happiest is probably that they always get a refund, even if there hasn’t been a problem. An AI that issues refunds to all customers would be a disaster – so this could be somewhere where strict rules are put in place. For example, only issue a refund to a loyal customer who has never had an issue before, otherwise pass to a human.
Current instances of agentic AI involve a blend of autonomy combined with strict rules, logic and performance. Knowing when to rely on the flexible and dynamic nature of LLMs and when to use strict rules is the key skill in building successful agentic AI.
Agentic AI - AI acting like an agent
Another way of thinking about Agentic AI is like a layer of AI that sits across all of your systems and data and is able to access and pull the information that it needs to perform its functions.
So, if you have a CRM, an Order Management System, ERP, Ecommerce platform, payments system and so on, this AI layer could sit on top of these technologies and orchestrate the right response.
This is basically the way that an agent would work now, looking to find the right answer from within your systems. An AI could do the same tasks quickly, the question is would it be able to make the right judgement calls? This is where training, guardrails and all the other limitations would have a role to play.
Getting started with agentic AI
Truly autonomous and self-learning agentic AI for customer service is still in its infancy, but there are ways you can get started with the building blocks. These are deep integrations to your systems, generative AI, real time data and process flows.
All of which is combined within the DigitalGenius platform. Talk to us today to find out how you could start with agentic AI.
