How to Craft an AI Knowledge Base

Active Agent Support Through an AI Knowledge Base

Your knowledge base exists to support agents. AI is set to take agent support to new levels. It makes sense to bring the two together and deliver tailored agent support. Generative AI excels at processing and summarizing huge amounts of text, and you can leverage this ability to provide tailored support to agents during interactions.

Let’s explore how that might work:

• First, you have to train a Large Language Model (LLM) on your business documentation, to give it all the information it needs to answer common customer questions. Then, you need to rigorously test guardrails. The AI doesn’t know what information is relevant and what’s not; what’s true and what’s not. You need to minimize the risk of ‘hallucinations’: that is, your AI making statements that are simply untrue.

• Once your AI has been trained and properly restricted, you need to make it accessible to agents. This requires a single, unified interface. If your agent has to tab out of an interaction to access the AI, it’s going to waste more time than it saves. AI suggestions should be provided in the form of screen-pops alongside an interaction, letting agents access the information seamlessly.

• And generative AI can do more than just offer summaries of articles. Equipped with transcription tools, it can analyze the content of an interaction as it unfolds. From here, it can automatically surface the knowledge article most relevant to the customer’s situation, without the agent having to ask. That’s another step removed from the customer support process, and Average Handling Time reduced that little bit more.

It’s clear that generative AI is going to transform the way your agents access your knowledge base, forever. The AI knowledge base will bring massive efficiencies and elevate your Customer Experience. To take it that one step further, you can add intelligent decision trees.

Structuring the AI Knowledge Base through Intelligent Decision Trees

Customer journeys are complex; very rarely is one customer journey identical to another. That’s why we need human agents; only human intelligence can offer that mix of empathy, instinct, and knowledge that leads to a speedy and satisfying customer interaction. But developing that skill takes time, and AI can support agents before they reach that point.

This begins with intelligent decision trees:

• Decision trees are the skeletal structure of your knowledge base. In essence, they offer next steps. Once an agent has gotten to the end of the article, where do they go? What questions might the customer ask? What related topics might be of use to the customer? Are there opportunities for upsell? The agent needs to know.

• Linking articles together manually can be a hassle. Your knowledge management team might miss subtle connections, or fail to update connections when new information arises. Even in the best cases, the needs of every customer are unique; no two journeys are alike.

• An LLM can identify thematic and content similarities across your knowledge base, and build decision trees automatically from there. Not only does this save hours of pains-taking labor, it can be done on the fly, during an interaction. Depending on the customer’s intent and speech, the LLM can make suggestions that update in real-time, to ensure that the most relevant next step is always surfaced first.

The result is a seamless experience for the agent. They don’t have to spend time navigating decision trees that might not link to the most accurate information, or that might not be relevant to a particular customer inquiry. Once again, the AI knowledge base enables us to cut average handling time.

The Future of AI Knowledge Base

Generative AI is an important tool in any technical writer’s arsenal. Let’s face it; no one enjoys writing long explainers of company-specific issues. It can take a long time, and trying to rush the process will lead to poor copy and confused agents. Luckily, the AI knowledge base can help.

Generating an explainer based on company data should be easy, once you’ve trained an LLM on relevant sources. But that battle doesn’t end there:

• LLM can only respond to the prompts you give them. Generating a good knowledge article requires some skilled prompting: you need to balance the general and the specific, meet a particular format, and focus only on specific issues. Here are some prompting tips to get you started:

• “Act as a topic expert…” – Asking the LLM to respond as if it were a subject matter expert can lead to more focused, technical responses.

• “Produce a knowledge article, using a three-point structure…” - Being specific about formatting is important, otherwise you’ll have to spend time reformatting the generated text in a way that works for you.

• “Explain jargon in simple terms, comprehensible to the average reader…” – Remember your audience. Agents joining the contact center for the first time will struggle if presented with a mess of jargon. Make your knowledge articles simple and accessible, and you’ll make your support more effective.

• Even when you perfect your prompting, LLMs can only do so much. The LLM doesn’t know what’s in your brain – only you know that. They tend toward generalization, even when trained on company specifics. Remember that an LLM is a probability model – it doesn’t actually know the solutions to customer problems; it’s guessing, based on mathematical probability. Sometimes, all your agents need is a clear statement of verifiable fact, presented in the same way, every time.

• LLMs also tend to ramble, not to mention hallucinate. If you’re going to use generative AI, you’ll want a good editor. Letting an LLM run loose and failing to catch its mistakes is a recipe for disaster as misinformation becomes entrenched, both in the model, and in your organization.

An AI knowledge base is powerful, but you don’t want to outsource your entire knowledge base to AI immediately. The technology still has a long way to go. Focus on empowering agents in meaningful ways, delivering real value for customers, and using AI to take small steps that lead to big savings.