Forethought aims to build more accurate chatbots with constrained generative AI models

Forethought has been building chatbots since 2017 with increasing levels of sophistication, intelligence and automation. Today, the startup announced the next phase in their development. It’s bringing generative AI to the platform with a beta release of a new tool called SupportGPT.

The product is designed to deliver auto-generated customer service responses without the need for human intervention. In spite of the blinding hype associated with generative AI, it’s still early days for the technology and there are limitations. CEO and co-founder Deon Nicholas says that his company recognizes this and has designed the generative AI in SupportGPT to use a narrower set of data than more generalized GPT applications, which he says should help deliver more accurate answers.

“With SupportGPT, our customers can start to access more focused answers to their customer’s questions,” Nicholas told TechCrunch. While it uses OpenAI technology under the hood, it has been modified and enhanced with Forethought’s engineering spin on the concept.

He recognizes that one of the big issues and challenges with generative AI in its current state is blatantly wrong answers, but he believes by limiting the set of answers the model can access, it can reduce these kinds of “hallucinations” we’ve been seeing where the AI confidently answers incorrectly.

“Hallucinations, where the AI goes off the rails, is the main problem of generative AI, and so we’ve developed a few clever algorithms, while leveraging the existing infrastructure that people have been building,” he said. “Forethought feeds this to the generative model as a prompt or even as what you’d call a guide, and it ends up being a lot more tightly coupled to the customer’s actual workflow, customers’ actual business.”

Being able to constrain the AI in this manner means it’s more likely to respond in a reasonable way. In a demo from early customer Upwork, he showed how if you asked a question out of scope like the weather, the tool would recognize it and try to steer the conversation back to subjects it knows about. By programming the AI to understand that there is a limited set of responses it can answer, it can tell you in an intelligent way that you’re veering away from that. In the demo example, when asked about the weather, the bot responded that it wasn’t a meteorologist, certainly a reasonable response in the context of the question, and provided examples of the kinds of questions that it could answer.

One of the keys here is making this work for each industry and company, so Forethought also announced a beta of SupportGPT Playground, a sandbox where companies can experiment with SupportGPT using their own data.

Forethought won the TechCrunch Disrupt Startup Battlefield in 2018, and has raised $92 million, per Crunchbase, including a $65 million Series C at the end of 2021.

It’s just one of many companies taking advantage of generative AI for business. Salesforce also announced a pilot this week of Einstein GPT, which adds generative AI capabilities across the Salesforce platform. We can expect to see many similar announcements in the coming months.

But for customer service, which has been using chatbots for years now, this is a potentially big leap forward toward more accurate, and less frustrating, interactions with automated bots, helping them provide more accurate and meaningful answers more of the time.

“I think you have to marry the LLM (large language model) layer, which has changed the game, with being able to actually understand the company’s policies…and if you do that well, and do that at scale, you will have this brand new paradigm of customer service,” Nicholas said.