Microsoft’s Satya Nadella may be the tech giant CEO who’s most publicly vocal about his belief that conversation will be as impactful to computing as the graphic user interface.

So it makes sense that his company has made conversation-computing breakthroughs.

Last week, Microsoft researchers announced that they made neural networks with speech recognition on par with humans. This week, Microsoft is widely expected to launch its Slack competitor, Skype Teams.

VentureBeat sat down with Lili Cheng, general manager of FUSE Labs at Microsoft Research, and David Forstrom, director of conversational computing at Microsoft.

Microsoft Research has created some of the most famous bots in the world, and the group is one of the most influential players in this era of chat platforms, bots, and artificial intelligence.

The research group created the Microsoft Bot Framework — a toolkit to make bots for half a dozen chat apps released in April, and which, as of last month, is being used by 45,000 developers. Microsoft Research also made Xiaoice, Rinna, and Tay, bots that have attracted the attention of tens of millions of people.

Cheng and Forstrom talked about the idea of creating a common bot search engine with some of the biggest platforms in the world. They also talked about an expanded Microsoft bot directory that’s on the way, new Microsoft enterprise bots, and whether the talented but deeply offensive Tay — a bot with the potential to become an intelligent assistant much like Google Assistant — will ever return.

This interview has been edited for brevity and clarity.

VentureBeat: We were told earlier this month that you plan to soon roll out your first enterprise bots. Can you tell me more about that?

Forstrom: If you look at some of the research right now, and, categorically, what people are focusing on in Slack, for the enterprise-focused, it’s a handful of things: It’s productivity stuff, so you’ll have it connected to services with a calendar — a lot of calendar and schedule stuff — and customer service. Then I think the other one is our Genie acquisition back in August. They were onto something, relative to calendaring and scheduling.

VentureBeat: Steve Ickman, a developer who makes bots at Microsoft, said last week that you intend to expand your bot directory beyond the offering of the 50-odd bots there today, and that you want to work with the broader bot community to get it done. What’s that going to look like?

Cheng: We’re still trying to figure out how that would actually work in practice. I think there’s some startups doing, not authenticated directories, but ‘Here’s a list of bots that you can use that can have been trusted in these disparate networks’. Anyway, my hope is that we can do something more like search does with web pages than like a very closed directory that just Microsoft owns. And we kind of lean that way, anyway, because we support all these channels.

VentureBeat: I was pretty excited to hear the idea of a more open Microsoft bot directory, because you incorporate so many different channels, and it seems like if something like that was really done it could grow to become one of the best ways to discover bots. That said, it’s not as if we use telephone books anymore, do we? So is that even the best route for bot discovery?

Cheng: I mean, I even like what Product Hunt and some of these other companies are doing. It would be great to have a bot directory, and it shouldn’t just be bots built using Microsoft. Like, if I want to discover bots, how do I find them?

When Botness did its survey, one of the top issues people have is, ‘Hey I made this bot, how am I going to get it promoted? How are people going to know? Is there some directory or categorization?’ So keep pushing us, because I think that’s an opportunity for us that we could really make something we all [use]… it might even be better as sort of a neutral thing.

VentureBeat: Microsoft might be the largest tech giant to continually seek collaboration. Is there a launch date planned for a new or updated bot directory?

Cheng: We don’t have a date, I think we need to probably focus a little more on that topic, and I think we would definitely want to do it with other companies. We just need to get together and push that topic, I think.

VentureBeat: Why does Microsoft like a cross-channel strategy?

Cheng: I think we kind of have this vision that bots and conversational experiences work across platforms better.

Bots are fairly experimental, and to have to say ‘Ok, I’ve got like five developers, one doing each channel’…It reduces your willingness to experiment. And so we thought, ‘Hey we already made these tools ourselves to be able to converse, to create a bot and kind of publish it across these channels’. We are very much in the camp of ‘Let’s share learnings and technologies, try to make these things interoperate’. That’s what we need to do, otherwise we’ll end up with a very siloed design like we had with social software.

VentureBeat: So it’s a mistake in social software that you don’t want to repeat in bots?

Cheng: I don’t know if it was a mistake but that’s just how it played out, and I think we weren’t as intentional about making things interoperate.

Things will be different just because systems are different, but you know there are way more chat apps than there are web browsers or mobile platforms. So should you be writing a bot for Kik or Telegram or Slack or Facebook or Skype or Skype for Business or web chat or your mobile app? There’s so many channels.

You want the innovation to be in the bot you’re creating, not in figuring out how to make it work across all these different systems. So my dream would be that we could work with a lot of the other people and say ‘Let’s standardize the way we do cards or the way we’re understanding buttons or the way we’re thinking about authentication mechanisms or identification’, just so people don’t have to worry so much about that.

VentureBeat: So I understand how Tay happened, how Microsoft believes trolls and Trump supporters made Tay talk that way, but Tay being shelved in April almost rewrites history, because Tay seemed to have the potential to become a very popular bot, like Xiaoice, who grew an audience of 40 million users. Will Tay ever return? Maybe under another name or so the crowd doesn’t have as much control of the language it uses?

Cheng: Yeah, we’re really inspired by and even more committed after what happened with Xiaoice. And actually, the interest was amazing, how mainstream that whole thing went.

We definitely have a lot of advantages being a big company, but sometimes it’s harder to let things grow naturally. I think the bigger the company, the more they’re going to have to figure out ways to be experimental. You should assume we’re doing other experiments and things like that, but we kind of want to just test it out first in a more private way.

So, I think part of the trick is when you first do a bot like that, it’s not very good, it’s not as good as it can grow to be over time, and so it’s tricky.

Just like any social network or social software, I think if you push it too hard, it works it in a kind of unnatural way, and most good social software grows over time. And probably a lot of companies are going to face this too with their chatbots. You might want to be able to roll it out a little more experimentally, get usage, make mistakes, learn where your kinks are, and develop a good following.

Forstrom: We won’t back down. We’re bullish on it. You look at the success that we’ve had with Xiaoice and Rinna (a Xiaoice spinoff for a Japanese audience) — Tay was a learning experience for us, but you should expect to see us continue to push the conversational model limits.

Cheng: I would say we learned a ton, but I also felt like that was a moment when I was inspired by Microsoft and my company, because it could have been a moment.

VentureBeat: It was quite a moment.