Artificial intelligence / machine learning is all the rage these days and I am constantly bombarded by promises that A.I. based solutions can save the universe. Phrasee is really interesting as it uses machine learning to improve the performance of email marketing campaigns. As a linguist myself, it is fascinating to see how a brutally scientific approach can be used to improve the impact of language.
I really enjoyed chatting to Parry and hope this will convince you that AI is something to embrace rather than fear:
Joe: Hello and welcome to the fifth episode of The Digital Brew. I’m here today with Parry Malm from Phrasee. We’re going to be talking about artificial intelligence and email marketing.
Parry, thanks a lot for your time today, it’s great to meet up. Perhaps you can start just by introducing yourself. Give us a little bit of a background about who you are.
Parry Malm: Sure, my name is Parry Malm. I’m the CEO and one of the co-founders of a company called Phrasee. Phrasee is AI that writes better subject lines, SMS, and app pushes than humans can. If you guys get emails from companies like Domino’s, Wowcher and stuff like that, then you’ve experienced our tech without actually realising that you’ve experienced our tech. What it effectively does is it generates language at scale that’s indistinguishable from that which humans write and gets better results: more opens, more clicks, and more money.
Joe: Fantastic. And how long has Phrasee existed?
Parry Malm: About two and half years. We were founded on February 23rd, 2015. We’re now officially Putney’s fastest-growing AI company.
Joe: How many AI companies are there in Putney?
Parry Malm: That’s not the point.
Joe: Millions. (laughs)
Parry Malm: That is not the point. I didn’t come here to knit pick! (laughs)
Joe: One of the reasons I was interested to speak was machine learning, artificial intelligence, there’s a lot to talk about. I find it fascinating. At the same time sometimes I think there’s a BS factor. I have alarm bells ringing. It looks crazy. It looks really interesting. As an ex-linguist myself, I like the whole notion of language and applying that scientific level to language.
How do you personally see…is AI and machine fundamentally the same thing or are they different labels for similar ending?
Parry Malm: So the term AI itself doesn’t actually mean very much. It’s a catch all phrase which covers various different modern technological approaches to solving specific problems. But machine learning is one form of AI. It’s effectively long run pattern matching with some level of self learning. As recent computing power increases more and more, there’s a new area called deep learning. This is using advanced neural networks that basically mimic the human brain, have hundreds, or even thousands, of neurons firing at the same time, finding deep meaning within large data sets. Hence, it’s called deep learning.
Beyond that there’s even more areas. So we have natural language generation, which is where a machine writes language like a human. You have natural language processing and understanding, which is where machines can understand, at varying levels, the meaning in unstructured human language.
Then you have some really cool areas too, like speech recognition and machine vision for example, which is where the computer sees a human … or we’re seeing some great advances right now, for example, in cancer research where it can show an x-ray of a pair of lungs, for example, to a machine and it can predict with a higher probability if there’s a cancerous tumour.
Joe: It sounds a bit, sort of, sci-fi. Quite scary visions of robots programming themselves. Is that a total myth? How far away are we from taking, machine learning to the extent where you almost lose control because things get so clever that they spot ways of effectively becoming other robots?
Parry Malm: Yeah, that’s really an existential question about humanity itself and what does it mean to be human? There are various conditions of humanity. I think one of the big things is a human’s ability to feel and not just a machine’s ability to mimic feelings. So I think that humans will still react in unpredictable ways. As far as this whole apocalyptic concept…
Joe: It’s quite negative.
Parry Malm: …it certainly makes for great PR angles on the BBC and it makes for great movies in Hollywood. As far as what it’s actually going to be, I’m a bit of a sceptic about that. I think this whole concept of AI coming to take over, and us having to bow down to a robot overlord is, during my lifetime, probably mythological. I don’t have kids, so I don’t really care about their lifetime.
Joe: *Laughs* Yeah I think you’re right, it’s pretty unlikely.
What are the earlier examples of some form of AI? I suppose its been around for a while. Okay, any commerce recommendation engine presumably is based on, maybe fairly simple rules, but it’s based on what is tailored to the individual and it’s something which is not human thinking, making decisions – hopefully for your benefit – and, let’s look at it positively, a good AI should surely help the end user experience.
Are there any examples which you can think of, perhaps you could share any examples of, early adopters of this technology or is it, as you say, quite hard to spot sometimes?
Parry Malm: So AI itself is actually not a new area. I studied neural networks 20 years ago in uni and they weren’t even new then. In fact, chatbots have been around since the 60s and 70s. A chatbot in 1968 or so, called Eliza, which was trained to be a Freudian psychotherapist. Then some guy at MIT created one called Parry, actually the same as me which is kind of weird, and it was trained to be a paranoid schizophrenic. And one day Parry met Eliza, which was this great, sort of, two chatbots interacting and this took place in 1972.
It’s nothing new, but it hasn’t been accessible for a couple of reasons. Number one is a straight hardware issue. We need a huge amount of computing knowledge to actually access this stuff. And to actually train all these algorithms to do stuff. And secondly, there’s just been much more of a research approach to doing this stuff, where previously, a lot of the computer science research was towards algorithmic optimisation.
As far as people who are doing cool stuff right now, we were speaking about this previously. I’m really a big fan of Spotify and how they personalise stuff because their challenge is quite unique. Yeah. If you think about Amazon, if you bought a toilet seat one day, then you’re probably highly probable to buy a toilet brush the following day. It’s not a hard problem, but it’s a soluble problem.
Joe: There’d be patterns which are very common.
Parry Malm: Exactly, but last night I had dinner with a friend and, on the way home, I was listening to some Scandinavian death metal. I woke up this morning and, walking to work, I was listening to Master P, which is some old school New Orleans Hip-Hop from the 90s, because my taste in music hasn’t move on since the mid-90s – that’s a different issue. So how can Spotify then predict my mood and what kind of music I would like to then listen to? What they found is they’ve basically created this, sort of, vectorisation model of different musical genres, and even songs from individual artists. So think about this huge multi-dimensional space where you have Master P here, and you have related artists to him, let’s say like Silkk the Shocker, Mystikal and C-Murder. And then you have, let’s say, the Backyard Babies who are like, death punk from Finland or something and next to them you have Turbonegro or Generation Graveyard and bands like that. They create clusters around those and then identify within these vector maps what you are likely to like.
Joe: At that time as well. Imagine the time of day, weather, what you’ve been doing, there’s all sorts of – presumably the more data you can throw at it, the more likely it is that you get a better recommendation.
Parry Malm: So, you know what, it’s partially about volume of data, but that’s not actually it. If you have a volume of bad data, then you have a volume of bad data and you’re gonna get bad predictions off the back of it. So this is one of the dirty secrets that AI experts won’t tell you. The algorithms themselves tend to be open-source. They tend to be accessible and these days especially with Google releasing Tensorflow which is a sort of open access wide range of machine learning tools. You actually don’t need a Ph.D. to construct this stuff. The real challenge is accessing data. This is actually a really big opportunity in machine learning. I think we’re gonna start seeing ‘data-as-a-service companies ‘coming out, which are basically big aggregators of data. Not like Experian and stuff like that, but all this unique data that’s inaccessible to the public.
Joe: Yeah, it’s an interesting thought actually.
We as an agency, I mentioned it to you earlier on, I get endless emails almost every day, “AI new platform that does something.” For us, it’s often around paid search management who are basically promising, because it’s machine learning, it is better than humans. And I, I’m a natural skeptic, and I kind of think, “I’m not so sure about that.” And am I wrong to feel like that? Or is a degree of scepticism actually quite good? A good place to be?
Parry Malm: You’re absolutely right. I think it was Einstein that said, “If he had 60 minutes to think about a problem, he’d spend 55 minutes on the problem and five minutes on the solution.” What a lot of these companies do, and I’ve seen them all the time, is they’ll say, “We are AI for this.”
Parry Malm: And a couple of years ago it was, “We were Uber for this.” Before that it was, “We were Tindr for this. We were Google for this.” And it’s just one more of these trends brought on and fuelled by over-valued VCs who are causing this. What I would really say is … and one area where Phrasee specialises is, we’ve defined one specific business challenge. We’ve looked at a different suite of solutions that could be applied to it and we landed on some permutations of deep-learning and natural language generation as a solution to our problem.
Are we are pure play AI company? Yes and no. We use various elements of AI and various aspects of stuff, but we also have numerous facets to the business itself.
Joe: Yeah that’s interesting.
So it’s specific AI for email marketing. Now, I, we do some email marketing, but not a lot. I’d like to do more, to be honest and I’ve seen some great results with email. Again, my natural scepticism and glass-half-emptiness … email, to me, strikes me as an area where there are challenges. Some people are just fed up with spam and unsubscribe to everything, and that’s fine there’s no point in sending them an email if they don’t want it. And even Google, the way they’ve changed the Gmail inbox and adding automated, sort of, hiding stuff really. That, I suppose, that’s an extension for affecting email overall.
So AI within email. What was it that drove you to found Phrasee originally? What was the challenge that you wanted to overcome?
Parry Malm: Yeah so prior to Phrasee, so two jobs before that, I was working client side for a large media company. I did a lot of email marketing, a lot of testing of subject lines and stuff and I used to always wonder, “Why does this one work and why does that one not work?” I then moved to a email technology company where I led the account management team, so I dealt with hundreds and hundreds, if not thousands, of email marketers throughout the world. And every one of them used to ask me, “What’s a good subject line?”
Joe: Yeah, “what works?”
Parry Malm: Yeah, my answer was, “Well, just test a bunch of stuff out and just pray that that one works.” That always felt like a total cop out. So one night three or four years ago, a bit longer now I guess, me and one of my co-founders, Dr Neil Yagger, who I studied computer science with in Canada 20 odd years ago, got talking, we were going to a death metal gig in Camden Town that night…
Joe: There’s a common theme here I think.
Parry Malm: …yeah, exactly, and we were at the BrewDog pub in Camden Town, knocking back a couple of Dead Pony Clubs and we just got chatting. We were talking about this whole problem. He’s like, “Well, I wonder if the solution there could have and approach to it” using his skill set. You see he’s got a Ph.D. in AI and this stuff and from there we just, sort of, came up with the idea. So yeah.
Joe: And what’s been the key challenges? Development? Product? What’s been the bigger hurdles you’ve faced?
Parry Malm: So we’ve gone through a lot of the permutations in the actual optimisation models. So we initially thought, for example, that it was all about individual keywords because we had done a bunch of research on paid search and whatnot. And in paid search, individual words make a huge difference. In email subject lines it’s not the case because people don’t read it like a search. They read it like a sentence.
Joe: Yeah, bit more of an emotional response, isn’t it?
Parry Malm: Yeah, yeah exactly and so we basically learned that individual words don’t really affect end result. What do, is the context in which the words are used. What we’ve built actually is quite a multi-pronged sophisticated deep-learning model that can predict, at scale, which subject line is going to win a split test before it’s sent out to the wild.
Joe: So is that … about the testing, does that happen across the board or is it individual testing? So is it per user or do you have, sort of, buckets and silos of people who will be the test bed for each different message?
Parry Malm: Yeah, it’s tested in buckets. I’ve been approached numerous times by people who go, “Oh, but I want the one perfect subject line, per person, every time.” And that is just … it’s a computation that’s infeasible.
Joe: Yeah, too much to do and it’s … I expect you’ll be more likely to get it wrong as well. You’re getting too freaky for everyone. You’ll probably get it wrong for everyone. Whereas if you do the bucket approach, if you get it right for 80% of people within that bucket, that’s success and the overall campaign metrics, no doubt will be approved.
Parry Malm: Well yeah, yeah. I think the problem is two fold. First of all, there’s that whole issue of false positivity, which you were just speaking about and when you get it wrong it looks really bad. But secondly, to have a technically robust solution, you need a bunch of good data points, and since companies keep their first party CRM data close to their hearts, they don’t have this aggregated data upon which they can make choices. So it’s really, really high risk and I’ve seen some people play around with it. I’ve not seen anybody who’s actually made money from it.
Joe: Yeah, well it’s an interesting challenge to, well, not guess. The whole point is to scientifically understand their mindset at that particular time, and understand what triggers would be most effective. It’s a huge challenge. But a fascinating one with it.
Where do you think Phrasee is particularly strong? Are there areas – I presume it’s constantly evolving and developing as all products are – are there some particular strands that you are most proud of? You think they’re absolute, real excellent piece of work?
Parry Malm: Yeah, so we focus on a few main industry verticals. We’ve found great success in the retail e-commerce space, great success in travel and holidays, great success in financial services. We’ve also found news cases that we’re starting to work on now, that we didn’t know existed. So, we work with a large telecoms company and we learned from them, one of their big challenges was making sure that people were at home when the engineers came to service their lines. And missed appointments were costing them something like £100 million a year. So what we’ve been working on with them through both subject line and SMS channels, is optimising language such that it increases, at scale, the probability that people being at home when the appointments are made. Which is quite a neat outcome.
Joe: Yeah. It’s very good, and it’s very easy selling technology because most decisions are made on finance and if you can say, “We’ve saved a fortune by, sort of, being a bit more clever about it.”
What percentage is email and what percentage is SMS? Is it, sort of, more email you’re doing?
Parry Malm: Correct, yeah.
Joe: Is SMS growing?
Parry Malm: Yeah, it’s growing-ish. We’ve found that that market fluctuates a fair amount. It falls in favour and out of favour almost daily. Even within individual CRM teams.
Joe: Yeah, yeah. I often wonder about the future of SMS, the likes of ‘Whatsapp.’ I was a pretty late adopter to Whatsapp, to be honest, my brother’s going, “Yeah, yeah you have to use it.” And, “It’s like SMS, but better.” And I was like, “Well, what does it do that SMS doesn’t do?” But I do love it actually. It’s very rare that I actually get text messages. It’s very rare, personally, that I will give businesses my mobile phone number so they can’t text a message. So it’s quite interesting.
But then email. Again, I’m a pretty active unsubscriber. If I feel that, that a) I haven’t asked for it and b) I’m just bored and it’s just clogging up. I get so many emails every day, I’m actually fairly aggressive. But I think that’s opportunity because the good emails, they’re lovely, they’re great, they really stand out. And you think, “Finally, something actually does appeal.” And it works. And I think … I was reading a statistic the other day about how effective email marketing remains and it’s arguably the most effective form of digital marketing. And we’re basically search. We don’t do enough email. But it does work.
Parry Malm: Well, there’s basically a few ways you can look at it as a channel but one important way is this: you remember a few years ago in about 2008, 2009 where everybody jumped on the Facebook band wagon?
Parry Malm: And they all, threw money on building up Facebook groups and Facebook followings and they had no even listing probability next to your friends and whatnot. And then Facebook went public in, what was it, 2011?
Joe: I think so.
Parry Malm: And they brought on Sheryl Sandberg, I think, who was a great commercializer of online audience. Then low and behold, companies then found when they were posting stuff to their organic following, hey weren’t getting as much feedback, as much P and page reach. And that’s because, lo and behold, Facebook said, “Well, hang on, it’s great being this company but we also want to make money.” So they started charging people for it. We found the same thing with Twitter. Where we found the same thing with all these online things. The great thing about email, is that data is first party and you control when it gets pushed out. That’s why, I think, it’s going through a bit of a renaissance.
Joe: Yeah, I agree with that and I think people who when you subscribe to the email and, again it’s very hard to avoid it through personal experiences, but I am now ironically more likely to subscribe to a newsletter on the basis that most unsubscribe processes work pretty well. So, I’ll give it a few, and if I like it great, carry on reading it and if it’s annoying then I’ll unsubscribe quite quickly.
I think this is the nicest part is you can do it on your own time, you know, ringing up is irritating whereas with email, you can park it, come back, it’s good.
So where do you see the future? We’ve agreed it’s not the Terminator, it’s not cyborgs killing us. Do you have a clear vision, a road map for the next few years or is it learning everyday and you’re not too sure? What’s the big vision?
Parry Malm: Yeah, so we certainly learned a lot in the last two-and-a-half years since we commercially founded. We were working on the tech itself quite a bit longer obviously. There’s not a day that goes by that somebody who is not commercially connected to Phrasee whatsoever gives us some great advice about things we’re doing wrong, things we’re missing and so on and so forth. But, I personally think the real value is really using these short forms of advertising text of first party data more effectively. See, one of the challenges with AdWords … first of all, there’s been about 15 years of research and commercial effort put into optimising AdWords. Could an AI approach improve on that? Maybe, but then you’re actually trying to improve something that there’s hundreds of companies who do currently, and Google are doing also. And trying to beat Google at AI or Facebook at AI is really, really challenging. Yeah, so we’re quite happy to focus on improving the business model we have but look at expanding out both geographically and into further direct marketing channels.
Joe: Yeah, where’s your audience currently? Is it UK? Is this English language or are you operating across all languages?
Parry Malm: No, we’ve got customers from Australia to California. We operate in French, English, German, Spanish, Italian. I’m probably missing one or two.
Joe: Yeah and as a customer of yours … I hope it’s not just literally a question of pressing a button and then everything happens? They must have some input. Is there a steep learning curve? Or is it something you can work hard to try and make as easy as possible for the adopters?
Parry Malm: Yeah, we focus really hard on ease of use. Because we know that CRM teams, especially, they’re dealing with 20 different systems on a daily basis. We don’t want to-
Joe: Be another one.
Parry Malm: -exactly. So we really, really focus on … and it usually takes people about five minutes per campaign. Usually, what they expect is it to be a really convoluted, complicated process, but we’ve purposefully built it like that.
Joe: Yeah and how long, realistically – I don’t know if you’re doing weekly campaigns – presumably their learning gets better and better and better, the more campaigns they do, I know we’ve talked about the quality of data, but let’s assume the data’s quite good. Is it right to assume the performance will get better and better and better. Probably the gains will start to slow up over time, but is their a standard learning curve for the software itself that you’d normally recommend?
Parry Malm: So we generally find after the first couple of months we benchmark on an average 20 percent uplift as far as opens, and then it increase as we look down the funnel. That fluctuates on a client-to-client basis because there are finite limits of how good one single open rate can be. So if you send out a message to 100,000 people, well first of all there’s going to be 50,000 who are busy and don’t see it. So let’s throw them out. So what we basically do each time is inch towards a theoretical maximum on a campaign-to-campaign basis.
Joe: Okay. Yeah, that’s sensible. I think, again my cynicism here, lots of people will set expectations that are unrealistic expectations then clients inevitably get frustrated. I’ve seen that a lot with big management software, paid service space where it costs quite a lot to use this technology and, the gain it’s not just the cost, it just doesn’t stack up and then you just think well no…it just doesn’t – it’s illogical.
Parry Malm: And we’re seeing that more too, with the increase market players who are doing AI for ‘X’ where markers often like to follow trends which is laudable it’s nice to be up on trends and whatnot, but a lot of them don’t necessarily get the purpose of AI and it’s a toolkit to solve problems.
Joe: Yeah, Marvtech is the new Thintech. So basically they get something, shorten it, then stick “tech” on the end and a lot of it is about their own self fulfilment and I think.
Cool. All right. It’s great speaking to you. Good luck with Phrasee and I’ll be watching with interest and watch this space.
Thanks very much. Hope you enjoyed that. Keep tuned for more. Thank you.
A big thank you to Parry – it was a great trip to Putney and I feel relieved that my apocalyptic vision of AI driven cyborgs threatening to take over is not likely in the (near) future!
We have a few more episodes in the pipeline for The Digital Brew but are always keen to hear from anyone who would like to get involved. If you have anything that you would like to discuss, please feel free to get in touch.