Editor’s note: This story originally appeared in the Unleashing Digital Value issue of Workflow Quarterly.
From their base in Montreal, ServiceNow’s VP of research Nicolas Chapados and his team are developing new applications for AI in the enterprise. Many of these applications sit in the earlier stages of the so-called AI value chain, where basic AI components scale up and help construct more complex solutions.
In a recent conversation with Workflow Quarterly, Chapados shared his team’s latest work on AI-authored workflows, cognitive buddies, and humans in the loop.
Our group recently prototyped a “text-to-workflow” solution. We can state in plain English a business problem that we want solved, and the model outputs a workflow that solves the problem we’ve described. What we want is to be able to state in a couple of sentences the job that needs to be done and to have that converted into a sequence of actions that incorporate relevant best practices, error checking, and other checks to make sure the workflow operates successfully and can be deployed robustly.
Your cognitive buddy
Today we’re still using the classical, mouse-based user experience that we’ve had for four decades. Imagine that in addition you can have a conversation with the system. Imagine you want to do a scenario analysis. You can just ask the system, “Hey, what if that variable becomes this? What are the consequences?” And boom—you get the equivalent of 100 clicks in one verbal request. Conversation becomes an additional modality to interact with the software.
Or imagine you have different bots that are all part of this conversation. One is specialized in financial planning and analysis, another in customer experience, and so on. These AI partners can propose next best actions given a current situation. You can ask your AI buddy, who is constantly looking over your shoulder, “What do you suggest I do here?” It’s still the very early days for these kinds of interaction paradigms, but I think they could be quite powerful.
Humans in the loop
I’m a big believer in the potential for AI to amplify human thought rather than substituting for it. To get the most satisfaction out of these models, I think a human operator should train the system, starting small and not trying to do too much initially. We’re seeing early examples, such as GitHub’s Copilot, which takes natural language text and turns it into usable code. If you don’t like the suggestion, you just get rid of it.
These kinds of interactions are very productive. But we still need to figure out which user experience patterns are most successful when humans and machines work together on tasks. ServiceNow is partnering with the company Hugging Face on the BigCode project, which seeks to responsibly develop large language models that will serve as the basis for future code-generation applications.
The ability for the user to get quick feedback from the machine and the machine to learn from the human in real time will be key for the success of such systems.