I think of systems like x.ai’s meeting scheduling service as AI in the wild, meaning that the service actively, even proactively, uses AI to interface with real people in a human-like way. While x.ai employs pattern recognition for parsing e-mails to discover calendar-based language, they do not stop at that recognition and call it AI as Google’s Arts and Culture app does when it matches your face to a face in a famous painting. And while Alexa may do well at putting an appointment on a linked calendar, it cannot coordinate a meeting among a group of people. x.ai sees AI as understanding and action.
x.ai focuses on one area of understanding and action: coordinating meetings. After three years of core research, the company ended up breaking their problem space into three distinct areas of intelligence:
Understanding requests in natural language
Deducing the intent of the request so the right action can be taken
Translating actions back into natural language so people can understand the actions
If the system, as CEO Dennis R. Mortensen emphatically states, doesn’t deliver all three of those intelligences correctly in equal measure and at the right time, then you don’t have a viable solution. “It makes no sense to identify the need for a meeting, for instance, if you can’t take action to help someone set-up that meeting.”
Good AI requires a good model
In its early days, x.ai focused on individuals and teams. This allowed them to grow their AI and human knowledge base. Yes, their human knowledge base. At the core, Amy and Andrew reflect a deep understanding of how people behave when coordinating meetings. The x.ai team builds that understanding into a model and their software reflects that model. Mortensen pointed to one of the myths of modern AI as he described their human-based modeling, “you don’t just download a machine learning algorithm on a Friday, fill it full of data over the weekend, and have results on Monday. It just doesn’t work like that.”
This model has to be as complete as possible. Systems that fail to meet expectations fail in the real world. If users repeatedly need to fix something the AI is doing wrong, they will probably abandon the system.
The x.ai model keeps the company’s knowledge domain clear. Mortensen and his team concentrated their R&D on e-mail-based scheduling dialogs, those found in apps like Microsoft Outlook or the Google Suite. This means they weren’t trying to worry about a lot of integrations as they solved for the core problem. What they did do was work their model so it goes well beyond explicit requests. The service handle ideas like rescheduling, adding people or resources as well as understanding concepts like “as soon as possible” when given a list of meeting date options. As early versions of Amy and Andrew experienced new types of requests, the x.ai team monitored the exceptions to the knowledge base coming in from users and refined their model and enhanced their service to reflect the new use cases.
In the future, Mortensen imagines Amy and Andrew lying in wait on any authorized channel, ready to wake up when a scheduling dialog begins. They would take the metaphorical baton from the people trying to schedule the meeting and coordinate to conclusion.
Eventually, Mortensen sees Amy and Andrew being integrated with other systems that can support meeting activities, like Open Table, Uber or Goto Meeting. Rather than just scheduling a meeting for six and telling the parities involved they are all meeting at 7:30 at the Metropolitan Grill, future versions of the x.ai service might well reach out to Open Table to secure the reservation. It isn’t clear if the team will teach Amy and Andrew to negotiate with each other over their Open Table points. When the dinner is over, Amy might ask its subscribers if they are ready for an Uber and make that request as well. x.ai may well evolve from just meeting coordinating to handling things associated with meetings like starting the coffee maker ahead of a meeting (or ordering it from the local Starbucks).
Currently, Amy and Andrew are constrained by another type of knowledge. They only know how to coordinate meetings in English. The team understands that today’s global economy means meetings anywhere, anytime, and many meeting attendees will need to have their appointments coordinated in a language other than English. That too, is on the roadmap.
Getting AI right requires transparency
Mortensen understands that people are going to poke his product with a stick and try to break it, and that some will be unnerved as they cross into the uncanny valley of human-machine dialog. The key for x.ai is a robust system that delivers on its core expectations. The company is building its AI not as a black box, but as a transparent system that works with people to confirm its understanding before just running off to schedule a meeting that doesn’t meet a users needs.
The other day I was working with Amy as I was evaluating the service and she said she would reach out to connect with my contact, when an earlier thread had already confirmed a time. A quickly cut and pasted a reminder. Amy then handled the details. Not perfect, but I’ve had humans make the same parsing error. And like those human coordinators, Amy’s model will continue to get better the more it is used. Unlike human coordinators, Amy will benefit from thousands of meetings a day to learn from.
Hard work went into making Amy and Andrew just work
Amy and Andrew are effective agents. They just work. Once given permission to access a calendar, they need only be copied on e-mails that might lead to a scheduling conversation. From there the x.ai scheduling bots manage the meeting set-up without further intervention. They do, however, continue to look for explicit instructions and additional data, like the need to change times, the addition of a location, or even the admonition to just stop working on a particular meeting.
While predictions of intelligent agents pervade tech trades, few companies seem to recognize the deep blocking and tackling that x.ai revels in.
AI isn’t magic. AI and data scientists have yet to deliver cognitive algorithms that just figure things out and produce effective human-to-machine interfaces on the fly. I still tell the story of mycin, an early expert system, that, when presented with symptoms derived from a visual examination diagnosed measles. The description, however, was that of a rusty VW bug sitting outside the developers window. Nothing about today’s systems with text or verbal dialog inputs would prevent a sophisticated diagnostic system from making the same mistake. One of the more pervasive Ads, Amazon’s Echo, stops short of offering all of its value because of issues with Alexa’s knowledge integration model. Developers build each Alexa skill for a specific task, but because of Alexa’s inconsistent conversational framework, making the plethora of skills easily accessible remains an interface challenge.(see Amazon Alexa Frailty: Syntax Fragmentation in the Skill Ecosystem for more on Alexa’s challenges). All AI has rough edges beyond which its algorithms cease to work. x.ai understands its boundaries well. By keeping their dialog within collaboration systems where scheduling is a default action, x.ai executes their service in a framework that helps maintain the constraints of their knowledge base.
The x.ai model focuses on capturing how people request a meeting, what they expect when negotiating the parameters of meeting with others, and the delivery of the action: a meeting that takes place where and when it should. While they have ambitions beyond e-mail as a channel, their primary goal remains making the meeting making experience better for all.
x.ai remains one to best examples, literally, of AI at work.
Daniel W. Rasmus
Daniel W. Rasmus, Founder and Principal Analyst of Serious Insights, is an internationally recognized speaker on the future of work and education. He is the author of several books, including Listening to the Future and Management by Design.