Sessions is temporarily moving to YouTube, check out all our new videos here.

How Do You Climb A Mountain

Josh Sephton speaking at HydraHack in November, 2017
Great talks, fired to your inbox 👌
No junk, no spam, just great talks. Unsubscribe any time.

About this talk

AI seems like an insurmountable mountain for most people. How do you convince a startup that they need to invest in something which might not pay off for a few years? How do you get people to think of AI as science fact instead of science fiction? How do you demonstrate a clear path to success before you've got budget to build a team? In this talk, Josh explains how to convince stakeholders by taking one step at a time.


So yeah, following on from Holly's talk, mine segues quite nicely, so that's nice little bonus. So, talk is how do you climb a mountain, which essentially there's quite a long journey between like someone deciding we need AI and actually getting to the point where you're deploying stuff, real products live. So, I've just recently been through this journey and hopefully there'll be some tips in here that you might find useful. It's essentially going to be kind of about taking small steps but hopefully along the way I'm also going to convince you that it's not just a game for the big boys, the Facebook's, and the Google's, and whoever else that's got going on. This is stuff that we could all being doing if we really wanted to. So, we're going to start with a story about food. So, about six months ago I wasn't feeling quite myself, I was a little bit ill, so my mom is a Celiac, so she's allergic to gluten, right. So, I kind of recognised all the symptoms, I was like I know what this is, I've got all the same symptoms, all I've got to go and do is go to my GP get a blood test and we'll be on the way. So, I called up my local GP surgery and said, "Hi I need to pick up blood tests." And the receptionist said, "No, you need to see a doctor." I said, "No, it's honestly just a blood test, like my mom's doctor has recommended that all the family get tested. I know what the test is, I just need to see the nurse." And she was absolutely adamant that I needed to see a doctor but there were no appointments for that day and actually I couldn't book one for the future either, so I needed to call up the following day and say, "Hi, are there any appointments today?" And just kind of go through this cycle. So, I did, called the next day spoke to a different receptionist this time, and I said, "Hi, I need to see, I need to get a blood test. I think I need to see a doctor before you'll kind of put me in." And she said, "Oh no, that's not a problem we'll put you through to the duty doctor, you can have a ten second phone call and they'll just kind of write the note to say you need to get the blood test from the nurse." So I was like oh this is a little bit different from the first experience, and it's kind of strange that health care is this way, where you're calling up the same GP surgery and speaking to two different people on two different days and getting different results. Does this sound familiar to anyone? Has anyone had similar experience? So, yeah, health care is a little bit broken in this country. The NHS is great and it does a wonderful job but to steal a phrase from canvas conference a couple of weeks ago NHS is giving sick care, not health care. You can't call up and say I need a preventive blood test can you help me? They're not interested, they just want to deal with you if you're actually dying there and then. So, this brings me to my current job now. So I'm head of AI at Push Doctor. Push Doctor is a private health care provider, paid service but it's all online with kind of GMC registered doctors and you can see a doctor within six minutes on the app, which is pretty cool. The doctors can provide you with prescriptions, and sick notes, and referral letters. And they can also send you blood tests to do at home so if you've got, in my case for instance I could have had that blood test where a little finger prick and send it off. We just closed series b funding in July, so 20 million odd quid to see us through the next year or so which will be pretty cool. And I should probably point out that we're not trying to replace the NHS, it's the equivalent of implementing a taxi service right? You might want to pay to get there slightly quicker. But, you know where we're more than happy to take a little bit of pressure off the NHS as we're kind of going through this process. So, I didn't set out to get this job. I spoke at Hydrahack probably 18 months ago and gave a very very brief introduction to AI and I kind of set out to implement some of these techniques and figure out what I could do with them, and I ended up speaking to Push Doctor about a completely different role but the CTO there is really kind of forward thinking. He said oh you've got a background dealing with data and we kind of think data is probably going to be valuable, do you want to kind of take on this project? And I says yeah that sounds really interesting. Although my talk is convincing a big business that they need to invest in AI, it helps if you start with the CTO who realises that you need to invest in AI. So, face value. Okay, so has anyone used Push Doctor at all? Or anything similar, no? There is a code there so you can have £5 off, give you a second to take a picture or write it down or whatever. Okay, so big data at Push Doctor. This is where we are today. So we're not at scale yet we're kind of doing tens of thousands rather than hundreds of thousands of consultations a month. To compare, the NHS helps a million patients every 36 hours so we've got some way to go yet but if we put everything in place from day one, then hopefully we should be ready to go when we do get to scale. So, this is where we stand today. There's a tiny little bar on the left here, that's the metadata so this is things like when you're consultation was, the diagnosis, the brief doctor's notes that say what was wrong with you, what you discussed, all sorts of things like that. A big thing we're missing is the transcription mic. So, we're just putting this in place now to be able to record word for word what you're saying with your doctor. Obviously it's unencrypted in the same way that all your notes are and everything. It never leaves, only the doctor and you can see those notes but it's kind of handy to have them on file. And so this is kind of a graphic there involving the data right. I've deliberately left the scale off because I don't want to give everybody secrets but this is the video that we think we'd be getting every month. So, as you're having an online consultation with the doctor, it's all through the video app, and if we were analysing every frame of the video, then this is the amount of data that we would be having. And we could be getting things like sentiment analysis, like what does your face look like? Are you telling the truth? Are you... whatever. All these things that kind of add in and they're things that we might miss whereas a doctor, as a human would be sensing all these things. So, Push Doctor is in a really great place where to start taking advantage of this data. There's a thing that's getting thrown around the industry at the minute, "data is the new oil". I personally don't like the metaphor very much but I was in a conference in Amsterdam a couple of weeks ago and literally every single talk had this slide in it so I thought I had better put this slide in it as well. Actually, it's flippant but there are a few similarities, data like oil, is kind of useless in its raw form, it requires quite a lot of processing in order for it to be valuable, and actually that's often quite intense processing. Data's dangerous if stored in large quantities in a bad way. Like you know, you can take it quite far if you really wanted to. So, does anyone here have data, does anyone know anything about their users like if you've got any amount of a little bit, yeah? Kind of a little bit? So, this is my other favourite Tweet that's going around. "AI is like teenage sex: everyone talks about it, no one knows how to do it, everyone thinks everyone else is doing it and so claims to do it." So we want to get to the point where we're actually doing it. There's a lot of people talking about doing it, a lot of people theorising it. And there's a conference called the NIPS conference that happens every year. It's the biggest academic conference about AI and the joke about the NIPS conference is that it's the same problem that's been solved over and over again for the last 30 years because they just take the same data sets and they say alright these are petals from a flower like how do we classify them this year? But, we want to get beyond that right. We want to actually use our data and we want to do good stuff with it and it's all going to be kind of building the business making this better and safer. But, where do we start? This is my product road map that I've taken the code names of our products off in case there's anyone from Babylon here that wanted to steal the code names. Obviously, the hardest part of any project. And so there's some easy projects and there's some hard projects, right? The ones at the bottom left hand corner are pretty easy to do, the ones in the top right are harder to do. I'll give you a clue, the one in the top right hand corner is a robot doctor, that's going to be the hardest right? That's going to take us a few years to get there. And we might get there eventually but... Like how'd you pick the first thing, right? Starting in the bottom left hand corner isn't necessarily the right thing because you'll do something but the rest of the business might not see the value in it and they'll go well that's pretty crappy, like it's not a great thing. So, we're like a very customer focused business. One of our top API's for the business that we track every single day is the net promoters score. So, it's kind of a measure of customer loyalty that's used in these big businesses. And that's the number one thing before we do anything, we say is it going to affect our customer loyalty. We don't care about up-time as much, we don't care about any of these other things, it's all about how good an experience are our customers having. To steal a phrase from Steve Jobs he said you got to start with the customer experience and work backwards. It's never going to work if I go into a meeting and say let's do some AI and let's do robot doctors 'cause that's like a massive time scale, it's not going to show any benefit for quite a long time. It's just never going to work. But what we do have is a massive call centre in Manchester that's growing at the same rate as our customer base. Every time we add a new customer we essentially ship in a new customer service agent, and say there you go you answer his questions, you answer her questions. We wanted to know if we could slow down the rate of growth there, we don't want to take anyone's jobs away from them obviously, like they're great doing what they're doing. But this is an area that does have real business impact for us. So, we started to look at the data and it's probably worth noting at this point as well that we wanted to do something that was unincumbent by regulatory issues. Health care is obviously quite regulated, quite rightly we wouldn't want just anything happening so again, another one to just go in and say well we want to do robot doctors, it's like it's never going to work. Because there's a whole bunch of hurdles. So these are some of our actual customers' service queries from few days ago. This is one of them, "Can your doctors prescribe off-label meds?" So you have people, honestly every day coming in and saying can I get some morphine? Can you just like ship it to me? Or can I have a bucket of Xanex because like I need it in order to get through the day. They're quite obvious to see who actually needs it and who doesn't need it, but this is one of the big things can you prescribe my medication? Is this a real service online that I can actually get help with? This is another one, "would your doctor backdate a fit note to January 2016?" So I'm not sure quite what scam they've got going on but they need one, that's that far in the past but this is a typical question, like what services do you offer? We write prescription, a fit note, whatever. And then this is the embarrassing bodies brigade that always pops up so, "I'm showing symptoms of claymidia" I don't think claymidia has any symptoms, I'm not sure what this person's problem was. But like the people who are too embarrassed to go to their regular GP or anything like that, like they always come on here. The number of dick pics that we get sent every day is absolutely horrific. Like, don't go into customer services for a medical company that would be my number one take-away from this talk. So we looked and we figured roughly about 50% of our customer queries could be answered like from their FAQ's. If people actually took the time to go and read the FAQ's we'd have 50% less queries. Now we don't blame them for this, like it's an emotional engagement, you only go and see a doctor if you're kind of ill or you already have a heightened sense of emotion, you don't want to bother with FAQ's. You just want to talk to someone. So, this is an achievable goal, this is somewhere we can start. We're not saying like we want to build a chat bot that will deal with everybody's queries in the entire world, all kinds of different aspects of them. What we want to do is take these queries that can be solved by the FAQ's, identify them and give a stock response. And this is like the simplest possible goal for this project, which I think is quite important. And actually we already have the infrastructure to support this AI. When it doesn't work, and it fails we already have a call centre full of agents who can deal with the actual answer. Or actually it's probably worth saying that it's actually humans being supported by the AI in this case, right. We're not going AI first at any point. This is kind of building tools, and building tools are helping the real people in our business and helping them do their jobs better. Again, start with the simplest thing, I know Holly's already given a brief overview of a Bayesian Classifier but honestly I think that everybody in this room can figure out how to make a Bayesian Classifier in about an hour or two reading Wikipedia and just kind of having a play, they really are so simple to the point that it's one of those things that no longer gets clusters artificial intelligence. Every time we figure out something new, and it ceases to be magical then we stop calling it AI. Doesn't mean it's any less actual data science or any AI. And so to be formal about it, Bayesian Classifier gives you the ability to say the probability of a class given a bunch of observations is this. I showed this to the CS team and their eyes glazed over so I went back and I rewrote it and I said what we want to do is find the most important words in a a question and see which class contains the most questions which contain these words. That's far more easy for them to kind of digest. And this is three weeks work, rather than three months or three years, so we're still in that time frame where I can just go and do it, and no one's really bothered by it. Okay, so we're currently, this went live a couple months ago and we're currently handling 15% of all the customer queries through this chat bot, right. So, it's not quite 50%, there's some stuff with our reporting so people come over and say thanks for that, or whatever and that doesn't get attributed to the chat bot. And some people just have complex queries and they say how much does it cost, chat bot gives a response. And then at that point the chat bot kind of signs out, hands everything over to the agents and they say you deal with the rest of it. And then their second response might be like, "Oh great, that's perfect, can I book an appointment for next Tuesday or whatever." All of these things, so we're not quite at 50, but like 15% is still pretty good, right? Again I want to reiterate I'm not trying to take anyone's jobs, but I'm about to start talking in terms of money and people get scared when I compare money to whatever. Our customer service teams work 17 hours a day, seven days a week, 15% is one in seven. So, we could shut down the office on a Sunday for instance. Like if all the queries came in at the right time. I'd say we have 5-8 people on shift. Maybe they're paid eight, nine, ten pounds an hour to answer these questions. That start setting up the real money. We're earning the ability to build something bigger. So now we've got to this point where we've demonstrated value, we've done something really simple and it's taken us a few weeks in order to do it. We're starting to get the rest of the business buying in saying oh that's great, yeah you've already saved us this amount of money, we're like go back and look at it again and see if you can do something more complicated. See if it can handle follow up questions or all of these things. We just like start adding kind of more skills. I think it's really important that we do this. The way to sell AI to a business is to find a business objective that's backed by data, pick a small achievable target, implement quickly to demonstrate value, and then buy time to build something bigger, go back and revisit it. And then you can do bigger and better and more complicated things. So that's how you climb a mountain, one step at a time. There's an entirely different talk that... Ian's my favourite person on the internet at the minute. Convincing a business to invest in AI is one thing, but we still have a massive problem which is that people on the street don't trust AI. So, this was a conversation about seeing a doctor online. Ian is not keen. Ian does not want to see a doctor online. I don't know how he's going to stay in hospital online. But kind of going forward everything I'm doing at work, I've got Ian on my mind. I want to make sure that Ian's comfortable with what we're doing. Like I saoid, a whole nother talk, and I'll leave that for now. But in summary that's how you get business to invest in AI.