Monthly Archives: May 2018

Having established that there’s a real, valuable use case for Google’s phone-call-making AI (making outgoing calls which have to be routed via complex menus, lengthy queues, or security gatekeepers) I got thinking.

When I was in my early 20s and worked in a real office with doors and a bit of peace and quiet, I had access to a much valued colleague who’s function has almost entirely disappeared from modern life, unless you are enormously rich and powerful. She was called "a secretary".

One of the secretary’s functions was handling incoming phone calls: blocking the nuisances, re-routing the misdirected, taking understandable messages if I was not available, or putting the caller through with a clear announcement if I was. Where "a secretary" scored enormously over "a telephonist" was in knowing a bit about my business and me personally and being able to make some decisions on her (it was usually a her) own. She could also recognise regular accepted callers by their voice and deal with them much more quickly than strangers.

I’d like a computer which can do that.

Now this is definitely a step harder than just placing outgoing calls, but only a step. We don’t have to create a full-blown JARVIS (Iron Man’s AI butler) to get a lot of value.

Recognising known contacts by their voiceprint and incoming line details should be pretty straightforward, and it should be easy to make the list manageable, adding rules about how to deal with different people at different times. Taking messages can be a hybrid of two technologies. Because the caller is talking to a computer the call audio can be recorded, but the automated secretary could run through a simple script to get a direct call-back number ("now you are sure that’s direct and he’s not going to have to go through some horrible menu to get back to you"), spell out the caller’s name and company if it’s not recognised, and get an identifying account number or similar so I can verify the call’s veracity and quickly get my case recognised on call back. These could all be popped into an email or text to me, so I can see them written down rather than having to listen to them and write them down myself.

Those capabilities alone would get rid of a lot of nuisance callers. Scammers who want to offer to move my money to their own accounts are not going to want to leave verifiable contact details, or will not be able to provide valid authentication. Sales calls are a bit different. Most "spam" callers don’t waste their time with answering machines, so if we make the AI secretary recognisable as such that will get rid of most. Any who are really persistent can then be recognised by "trigger" words, such as "PPI", or "double glazing", or "the security department of Microsoft Windows" (yeah, right), plus non-verbal cues like the double-ring of a connection from Asia, or the chattering background in an Indian call centre, just like I do it. That would be a really powerful application of machine learning technology. I could choose how my secretary deals with identified nuisance callers: just hang up, choose a random insult from a list and hang up, keep them talking until they get bored, or redirect the call to an 0898 number where I’m sure the young ladies will be happy to listen to them all day, for a fee.

While we’re at it, let’s make the voice and personality programmable. I had Joanna Lumley’s voice (Purdey rather than Patsy) on my satnav for a while, and that would tick a lot of boxes for me, as a 50-something male. But I can also see the charm of recreating some famous fictional assistants: JARVIS, or how about Chris Hemsworth’s character from Ghostbusters 3, ladies?

OK Google. How about this?

Since Google’s demo of an AI bot making a phone call a few weeks ago, the reaction I have read seems to be completely polarised. About half the reviewers are blown away, believing it to be unleashing AI wonders/horrors which are half a step away from SkyNet going live. The other half are nonplussed, seeing no potential value.

They are all wrong.

Let’s deal with the "this is the advent of true AI" bunch first. Google have demonstrated a realistic sounding voice which can currently deal with a few, very limited scenarios, and I suspect will rapidly fail if the other party goes significantly off track. Sure, it’s a step forward, but just a step. If you want to see a much more convincing demo, catch up with the program "How to Build a Human" from about 18 months ago, in which the makers of the Channel 4 Sci-Fi program "Humans" got a mix of British experts to build a robot Gemma Chan, who (which?) was then interviewed over Skype by a bunch of entertainment journalists. About half the reviewers didn’t realise they weren’t talking to the real Gemma. That’s much closer to a Turing test pass.

At the other end of the scale we’ve got those who don’t see any advance or value to a machine which can help make a phone call. To those, I have a simple question: "how did you get on, the last time you rang your bank / utility / travel company / <insert other large organisation here>?"

I completely agree that it’s a waste, and maybe a bit sinister, to task a robot with making a call to a local restaurant or hairdresser. But when was the last time you rang anything other than a small local business, and got straight through to talk to a human being? We all waste far too much of our time sitting on the phone, trying to navigate endless menus, trying to avoid the dead end where all you can do is hang up and try again, or listening to "Greensleeves" being played on a stylophone with a reminder every 20s that the recipient values your call. Yeah, right.

If I want to deal with a computer, I’ll go onto the website. I’m very happy doing that, and if I can do my business that way I will. The reason I have picked up the phone is one of the following:

• The website doesn’t support the transaction I want to execute, or the information I need. I need to speak to a human being.
• The website has a problem. I need to speak to a human being.
• The website has instructed me to phone and speak to a human being.

Spot the common thread?

So I have the ideal use case for Google’s new technology. It makes the phone call. It navigates the endless menus, referring to a machine learning database of how to get to a human being as quickly as possible, and how to avoid dead ends in that organisation’s phone system. It provides simple responses to authentication prompts if it can, or prompts me for just the required information. If the call drops or dead ends it starts again. And it listens to "Greensleeves" or equivalent, silently in the background, until it’s sure it’s speaking to a human being. At that point, it says, like a good secretary would, "please hold, I have Mr Andrew Johnston for you", gets my attention and I pick up the call.

In the meantime, I get on with my life.

In some ways, this is actually easier than what Google have already done, because most of the interaction is computer-to-computer, and actively doesn’t need a human-like voice or understanding. It’s certainly a better use of the technology than pestering the local hairdresser.

OK Google. Build this, please.