Can you use ChatGPT as your business chatbot? (and other questions about GPT)

Can you use ChatGPT as your business chatbot? (and other questions about GPT)

An outline for businesses about how to use GPT to get ahead with customer service chatbots.

TL;DR

Using ChatGPT as a business chatbot might seem like a tantalizingly time-saving prospect, but there are caveats. It’s unlikely to have detailed knowledge about your specific business, and adding that knowledge quickly eats up those time savings. It can also create confident but incorrect answers that can mislead your customers. But there is a way to combine the productivity benefits of AI-generated content with user-friendly curation and editing to improve the quality of that content and time to benefit. Now that is magic!

ChatGPT from OpenAI has captured our imaginations overnight. Rightly so, with its amazing ability to answer all sorts of questions on seemingly any topic you can throw at it.

It’s a short leap to think ‘wouldn’t it be great to have this answer questions about our business too?’ Either helping customers or acting as your internal knowledge management system fielding questions about training, onboarding, HR, customer support and more.

Unfortunately, there are a couple of hurdles that make this much trickier than you might think, but there are some cunning solutions too.

What is holding us back from using ChatGPT as a chatbot?

For starters ChatGPT doesn’t know much about your specific business (it isn’t up to date with current affairs after 2021 either, though that might change). It might have had access to your website when it was being trained. But that’s about it. If you try to ask it detailed things about your organization you’ll probably find it lacking. And none of your internal knowledge is in there – probably happily so – else your confidential and proprietary information would be publicly available.

There is a way to get around this by ‘fine-tuning’ the AI model.  Finetuning involves creating at least a couple of hundred “prompts” and “ideal completions” based on your content. It is a lot of work and even then you’re really only ‘nudging’ the model unless you provide even more data.

An alternative to fine tuning is if you provide enough context and content to GPT when you ask it to do something, it will then use that context to answer the question. So say someone asked a question that was in a policy document, you could send a chunk of the policy document to GPT along with the question and GPT could then generate an answer based on the content.  There’s a fair bit of complexity when dealing with bigger volumes of content – but this approach can work and generate reasonable responses.

Which brings us to a problematic word. Generate. The G in GPT stands for generative and is what this AI does phenomenally well. All responses from GPT are generated. i.e. they are not verbatim the content someone wrote, but rather a probabilistic response made up of all the material GPT has been trained on – your material and everything else.  You can dial down how ‘creative’ GPT is with its responses, but it is still based on probability rather than being fully deterministic.

How correct are these answers? Well, mostly correct for widely documented topics, but even then, not always. OpenAI calls these made up, wrong answers “hallucinations” and the challenge  is they come across just as confident and assertive as right answers.  Google, who just announced the upcoming release of Bard, has experienced this problem first hand.  In their demo of “What discoveries from the James Webb Telescope can I tell my 9 year old about”, Bard answered with a completely wrong “fact”, which is kind of embarrassing but also symptomatic of this kind of AI.

Really it is up to the person using ChatGPT/GPT to assess the trustworthiness of the generated response. This is fine if the responses are on a well-known topic. Not so good if this is your customer support chatbot! If the generated answers aren’t correct– how is the customer going to know? If you’re in a sector that has compliance requirements a generated answer made up on the fly might not cut it with your legal counsel!

Curating AI generated Q+A knowledge bases at scale

Curation is a term that comes up time and again around AI. It applies here too. If you could curate and edit the answers GPT generates, you would have more confidence. Editing and saving the edited versions gives the benefit of the AI generation while ensuring that checked and verified answers are provided to users. Curation also is the opportunity to add your brand’s tone of voice – which AI struggles to do a good job of.

Also, if you could give GPT some existing content as a starting point – like a list of web pages from your site, or a PDF manual - and it could automatically create a first pass of conversational content for you, that would still help an enormous amount. This crushes the ‘blank page problem’ because it is always easier to edit some content than create it from scratch.

This turns out to be exactly what our team has been working on. FAQ Wizard turns your existing content into questions and answers using GPT3.5. You can then edit and export the content and use it to kickstart the creation of any knowledge base. And, if you import it into Helpfruit, you can use all the great features to add extra material like video and images to enhance understanding and formatting, tags much more. Plus get full analytics insights on how chat is actually working.

Some other considerations you should be aware of if you’re trying to use ChatGPT directly as your chatbot:

  •  Rich content – GPT does not handle or generate any formatted text. If you want images/screenshots/gifs/emojis or video or even just some bold text and links included in your answers, GPT can’t do that for you
  • Rate and content limiting – you can only send a limited amount of content and at a limited rate to the OpenAI API. This means real time responses for a large audience might not work that well.
  • Privacy – ChatGPT is using all content you enter as training material to improve it. So take care with what private and confidential information is sent.
  • Analytics – If you want to know how well the chatbot is performing and what is being asked, how well conversations actually play out – you won’t get any of this insight from OpenAI.

About Helpfruit

Helpfruit is a customer self-service solution that uses AI to match customer questions with verified answers. Helpfruit includes chatbots, live chat, help pages/articles. FAQ Wizard combines AI-generated content with human curation for faster, more accurate knowledge base development. If you'd like to try out FAQ Wizard, start here.

 

(This blog was originally published by FAQ Bot. FAQ Bot rebranded to Helpfruit in May 2024. This blog has been updated accordingly.)