Anyone who has worked in an office will know how much email admin they have to go through before starting their actual work in the morning.
Looking to remedy the tedious task of scheduling meetings, US-based AI startup X.ai has created Amy, an intelligent agent that lives in your email and manages your schedule.
“Amy exists in the same way as a colleague, but exists in your email,” X.ai CEO and founder Dennis Mortensen tells Techworld during Web Summit in Lisbon.
The idea is pretty simple. You receive and email asking to set up a meeting. Next, you cc in Amy to the email chain and it will sort the scheduling for you.
You won't need to download any software, or select anything from a drop down menu. Amy lives in your inbox.
“When someone sends me an email saying “Hey, I’m going to be in Manhattan in December have you got any time for a meeting?”, he explains, “I can reply back, in natural language saying “That sounds like a good idea, I’ve cc’d in Amy and she will help put something on my calendar.”
Read next: Should robots have rights?
For Mortensen, the concept is not to help you set up a meeting, but to actually take that task totally from you.
“Most software today will augment you to be able to do a job a little bit faster but few software packages will do the job for you. Amy isn’t helping you. She is saying, I understand what you want to do and I will do it for you," Mortensen explains.
“In fact, you’re not going to be able to see this [process], you’re going to be excluded from it.
“Think in jobs, not in tasks,” he adds.
How human should we make AI agents?
X.ai have two agents, Amy and Andrew, both of which can be cc'd into your email chain and complete the meeting scheduling task for you.
Obviously, for Amy and Andrew to be a success, they need to know, or seem to know the best way to set up business meetings in a professional and effective way.
The question of 'how human shall we make them?' crops up time and time again. But for Dennis, it isn't the case of making them totally indistinguishable from humans.
“I don’t think that it’s a game of whether I can fool you into believing it’s a human, because that’s not the point," he says. "I think it is extremely valuable to know it’s a machine, but in the end, it’s done such a good job, you say thank you."
Mortensen made it clear that Amy and Andrew aren't here to replace humans, but to be part of the workforce with us.
“AI agents like Amy and Andrew will exist alongside human agents. For these agents to exist they need to work in an environment where they co-exist with other human agents.
“Just like with self-driving cars, you can’t say on Mondays all cars must be machine driven. That’s not going to fly.
“We also have human drivers, or human agents in this case. And that’s why we humanise it. Because the agents need to exist amongst other human agents.
"We’ve already seen agents being able to repeat back sentiments that fit the purpose of the text or talk, such as an apologetic tone if certain keywords are used or an upbeat one to match the purpose of the agent."
For Mortensen, humanising Amy and Andrew is vital in ensuring business relationships are upheld.
“Let’s say Amy emails you because I ask her to reschedule a meeting with you and because she works for me so she’ll do what I tell her to. Then two hours later I ask her to reschedule it again. Then later for a third time, I ask her to reschedule it," explains Mortensen.
“At some point I move from being a nice guy, to a bit of an arsehole in the eyes of the person I’m trying to set up a meeting with.
“Normally, people would say “Listen, just forget it, let’s reschedule for a later date”. But if we introduce a level of empathy, we have more of a chance making sure the meeting is kept in tact.
“The agent assumes some sort of empathy, which of course it does not have, but seeing that in an email makes you slightly more okay with being messed around.”
What are the challenges facing Amy and Andrew?
The challenges facing X.ai and other intelligent agent makers will centre around the same things. How can we create an agent that can evaluate situations and take the right and appropriate actions each time?
For Mortensen, creating an agent that has a higher level of reasoning and decision-making is important, but is also one of the biggest roadblocks.
“There’s a natural language understanding challenge for Amy," says Mortensen. "So say you email Amy saying you’re going to be five minutes late for a meeting, if Amy does understand this in full, then she needs a reasoning engine that asks ‘what do I do with that information?’
It's these complex and very common email scenarios that are currently proving to be difficult for Mortensen.
"Understanding is the biggest challenge, because people are fucking crazy! They will say things that are not true, or be super ambiguous [over email].
"For example, say last night at 1am I asked Amy to set a meeting up with someone 'first thing tomorrow'. That is not true. What I mean is today. But as humans we often use the term tomorrow for meaning 'once we have slept'.
“The agent now needs to do two things; one have very high confidence in having detected and extracted the right date and time, and be even more confident to say, “I know what you said but this is what you mean”. And there is a thousand of those examples.”
Most of us will email the same people quite often and change our email style depending on that person. We might take a more conversational tone or a formal one given the type of email and its importance.
“One of the questions we get asked is “I’d like to pick between a few personas”, but actually that’s not really what you want," he says.
“What you allude to is the tone and how Amy needs to pick up on it. And we don’t do any visual tone change to the person because the data is very sparse."
Mortensen highlighted how humans provide unnecessary information to AI agents, adding an extra layer of navigation.
“Let's say you say to Amy, I can't do the meeting tomorrow at 1pm because I have to go to my children's school at 2pm, but I think it should end at 3pm. So how about we do 4pm?
"The reason for changing the meeting, Amy will have to know that it doesn't matter. The reasoning, isn't important," says Mortensen.
Another challenge for Mortensen is the way we write emails. We very often lack the direct nature that AI agents prefer and 'understand'.
"Nice people find it difficult to just cancel meetings outright,' says Mortensen. "Instead we say "I don't think that meeting at 3pm is going to happen. But we should get a coffee sometime.
"What he is really saying is that he wants to just cancel the meeting. And that is a real challenge that we're dealing with," he adds.
Mortensen doesn't shy away from the obstacles thrown up when creating an AI agent. Instead he aims to master a smaller section, within natural language.
We are not trying to solve natural language, it's an unsolved science," he explains. "But what we can do is pick a sliver and potentially solve meeting schedule language. And that is what we are trying to do here."