Team AI is not Personal AI
You want to make the workers in your office happy? Give every one of them an assistant.
Oh, sorry, you can’t afford that. That’s too bad. Fortunately, there is a technological solution.
We just completed a survey on worker attitudes about AI assistants. It shows that people are ready, and eager, to work with artificially intelligent assistants: 77% expect them to make meetings more productive. 78% think they will become personally more productive.
I was surprised by this result. I thought people would find the idea of having AI-based assistants at work disturbing or creepy. We actually asked about that. And a small percentage do, but the majority think it would be “productive” or “awesome” to have AI that recognizes them and knows their schedule, and helps them, for example, start meetings.
I believe we have Alexa and Siri to thank for that. We’re getting accustomed to having AIs in our lives. These bots are getting good at helping us as individuals.
Team AI is Different
But now it’s time to bring AI assistants into the workplace, and have them help our teams. Which, as it turns out, presents new challenges for AI engineers. These are some of the things a team AI has to do, that a personal AI doesn’t:
1. Identify the speaker
A personal AI can generally assume that, when it’s asked a question, it should answer it for its primary owner. In a team setting, even a simple command, like “Start my meeting,” means that the AI has to first figure out who the “my” refers to.
Currently, our Cisco Proximity technology can identify who’s in a room; over time, we will combine it with AI technologies like voice and face recognition to determine who, exactly, is speaking at a given time. This is key to making team AI work, but it hasn’t been a priority for personal AI developers.
2. Understand business context
It’s one thing to ask your smartphone for the weather. It’s quite another to ask a team AI assistant to bring up the latest sales figures from Des Moines. Team assistants need access to deep data sets, and they also have to understand their particular company’s lingo: code names, acronyms, the jargon and technology of the business they serve.
And they have to keep this information private, inside their company. A team AI can’t rely on the vast data set of all users worldwide to fine-tune its understanding. That smaller data set means the AI must be better at seeing connections and tuning its learning algorithms.
2. Be much better at keeping quiet
In a personal setting, like your kitchen, the accidental activation of an assistant is sometimes annoying, and often amusing. In a team setting, having an assistant interrupt because it thinks it has been activated is not as easily forgiven. The false-postive firing rate for a team assistant must be much lower than what we accept for today’s personal assistants. A meeting AI has to mind its manners — or it won’t be invited back.
Furthermore, in a group setting, continuous voice recognition is much more complicated. There are interruptions, there’s side chatter. We have been spending a lot of time training our voice assistants to be rigorous about rejecting invocations — in addition to making sure it can recognize its intentional phrase, “Hey, Spark,” even in noisy conference rooms, no matter who says it.
The meeting AIs will understand group dynamics. They will behave differently in a casual weekly team meeting than in a high-level long-term strategy session of senior executives and board members.
3. Keep private data private
Personal assistants have access to our most private data. A lot is personal, like financial data and personal emails, and our devices also contain business chats, confidential data, and perhaps files like performance reviews. When a team AI is helping to coordinate things between group members, it has to know what data it can use, what it can’t, and what it can use but only if it obfuscates. As our automated assistants get more powerful, we’re going to have to be careful to make sure they know what data is private to the individual, to the group, and to the business.
Workgroup products like corporate email and chat systems generally have access to rich contact lists, and often directories of files attached to projects. We can use that data to make the team AIs smarter, but sometimes we shouldn’t, for corporate privacy reasons. Our team AIs need to understand much more about data and document security than a personal assistant.
5. Learn to take smart notes
Eventually, we will have AIs in our meetings, and they’ll be listening in all the time, so they can record, transcribe, and summarize what’s happening in a meeting. Will they interrupt if they can’t make out what someone is saying, or who’s saying it? What if their transcription is perfect but the meaning itself is vague? Building AI team members is about much more than voice recognition. We have a lot of work to do to encode correct behavior.
6. Work on specialized hardware
Personal AIs run on small, dedicated devices, as well as personal gear like laptops and smartphones. Team-focused AIs will work in acoustically complicated environments, where several microphones might be picking up the conversations of multiple people at once. Good meeting-room AIs will require hardware and interfaces designed for teams, not just individuals.
Towards happier teams
We are just getting started with team assistants at Cisco: Early in November we announced our first AI meeting bot, Cisco Spark Assistant. I think its ability to start, join and end meetings, and place outbound calls, will make teamwork easier. But there is a lot more that we will do beyond that. The challenges I listed here are some of the ones we’re working on now. They will make teams work better — and I believe they will make us happier at work, too.
