How to design an AI-powered chatbot
One of the most tangible ways of seeing AI in action is to with a customer-facing chatbot.
In this post I’ll help you think about how this could provide meaningful value to your customers, and in turn grow your pipeline and revenue.
If you prefer to watch, here is a video overview of this post.
Stop building support chatbots
If you ask someone to imagine a chatbot, they immediately think of a support chatbot, “Ooh our customers could speak to it and we can save money on human support agents”
But for a support chatbot to be great it needs integration with all your other systems - billing, product, contracts. These then become very complex and expensive projects.
Instead for this exercise, I encourage you to think of other use cases where you can help your prospects and customers with other jobs they are trying to do.
As an example, on my website I know that visitors are often trying to use one of my templates or plan a workshop. The chatbot on my home page recognises that and helps them.
Put yourself in your customer’s shoes
The jobs-to-be-done framework is well known in product development circles. It makes the point that we only buy products or services to do a job for us.
We don’t want an umbrella - we ‘hire’ it to prevent rain getting on us.
So when designing your chatbot use case think of the jobs that your prospect or customer is trying to do at that point in the customer journey.
Early on in a buying process they are trying to learn about the category and to build a business case for change.
After signing an agreement they are trying to build their internal team and processes.
Design your chatbot(s) around helping the customer with their job, instead of focusing it totally on your job (getting them to buy something from you, or reducing your cost of serve).
Who are your ‘gurus’?
In any company I have worked in there have been the ‘gurus’. The experts that everyone wants to have on a customer call because they have the depth of experience, the right questions, the templates that guide the customer and give them the confidence that we’ve done this before.
Often the information in these people’s heads is unique to them - you won’t find it in your standard operating procedure or intranet.
This is the knowledge and experience that your customers want access to, and you should think through how to capture that information and build it into your chatbot.
Take Infosecurity as an example. Often these individuals are brought in at the end of a sales process. Maybe they also bring along a long document that explains all of the security credentials and certificates the company has.
Imagine an Infosec chatbot where the customer’s Infosec team can quickly ask their questions while they are in their internal meeting instead of having to book a call and wait two weeks.
Prepare your chatbot ingredients
Having defined your use case, now you can start to prepare the ingredients for your chatbot.
Interaction flow
Here you should consider the user experience based on the jobs to be done we looked at earlier.
If it is a simple job - “Confirm company address” then that is a simple FAQ.
But if it is more complex “Help me define the typical roles and responsibilities that we will need to hire for to roll this product out”, then the interaction will be more of a coaching scenario with questions to gather more context.
Think back to your ‘gurus’ - how would they interact with the customer, what questions would they ask?
Document sources
Secondly, we can train your chatbot on the domain knowledge that your customers are after.
What documents, templates, processes or policies do your ‘gurus’ use?
Often this is an opportunity to bring “in-house” some of the most valuable content that sits on one persons laptop.
Avoid integrations
This goes back to our point at the start about support chatbots. Adding in integrations adds complexity, cost, time and risk of hallucinations that send your chatbot project into the land of perpetual POC.
Instead, focus on building a chatbot that stands on its own right.
This is another good reason for focusing on prospects, as there it little to no customer data to integrate with.
Next steps
Having gathered all your ingredients, now you can build.
I will not go into the details in this article but you’ll have two main decisions to make:
Picking the backend LLM by testing your use case and evaluating the responses you get back. (Also note, OpenAI’s new Responses API makes it easier to search files without having to build a RAG architecture)
Deciding where your chatbot will go. It could go on your website for prospects to access, but you might choose to only show it to authenticated users in your support site, community or product.
Consider also timebound use cases, such as a chatbot to allow attendees to prepare their attendance at one of your events.
I hope that inspires you to come up with some innovative ideas for your own chatbots that truly help your customers with their own jobs-to-be-done, and brings to life the experience of your best people without having to wait weeks for a meeting.
If you provide value to your customers throughout their buying process, you will end up with more pipeline and ultimately more revenue.
Get Started
Whenever you are ready there are three ways I can be helpful:
Model 101 Playlist: 20 minute guides to using ChatGPT, Claude, NotebookLM for work
AI Inspiration Briefing: Show your people the path forward in this 90 minute live session
Kowalah: Buying platform to help you pick the right AI tools for your business