AI for Email Marketers
Size isn’t important — So say the experts from Phrasee.
Parry Malm, CEO of Phrasee, had the audience eating out of the palm of his hand at last week’s DMA event. The topic was applications of artificial intelligence in email marketing and every marketeer there will have since decided that perhaps they should be working with Parry’s subject line generation company.
Throwing it open to the room, we were presented with a list of subject lines that advertised Domino’s Mexican Fiesta pizza, and were asked to vote for the ones we thought had been written by a human. Despite a relatively equal spread of votes for all the options, Parry revealed that all but one had been computer generated, demonstrating the ability of artificial intelligence to utilise natural language.
Domino’s CRM Manager, Nicola Keane, gave an excellent following presentation demonstrating the impact of Phrasee’s technology. The Pizza Group has seen a 26% uplift in email open rates, a 57% uplift in click through rates, and a 753% return on investment.
So the choice is obvious, right?
The temptation is to go out immediately and grab AI solutions for all your generation and optimisation needs. Tell your email marketing staff to take a sabbatical, throw out the popcorn machine and the other trendy gizmos, and put your feet up. The computers are here to do the work for you.
Well, as clear cut as it may seem, as Parry pointed out, it isn’t that easy yet. Generation is the easy part. If you’ve been to the Barbican’s sci-fi exhibition, Into the Unknown, you’ll have seen Sunspring, a short film written using AI. You have to credit the spectacular acting and direction for any sense of coherence because the script, built by an AI from sci-fi tropes, is complete nonsense. (If you haven’t seen it I absolutely recommend watching it here).
Optimisation is the more complex issue. That is, developing content that achieves it’s purpose. In the case of subject lines, this is looking pretty good. The results from the Domino’s project speak for themselves. The ability to predict success by calibrating the content against linguistic structures that have a proven impact on your audience is incredible. It is, in fact, better than humans at gauging the potential for success.
Below are the true and predicted rankings of the subject lines. The results debunk a common SLO myth, length of the line really does not matter, it only acts as a proxy metric for volume of information. Semantic and syntactic properties are much more important. The secret to success is in dimension reduction, highlighting patterns in ways people cannot, using bidirectional long short-term memory networks. This demonstrates the importance of longitudinal applications of AI — the predictions get more accurate as the network learns over time.
However, it is challenging to produce contextually appropriate, brand compliant, language at scale. There are even greater challenges with images. A generative adversarial network, using discriminatory neural network filters, can identify an existing image that best matches a given description, optimised for a particular purpose, as demonstrated by Expedia’s application of Mechanical Turk to show you the picture of that resort that is most likely to result in a sale.
What the technology cannot yet do is generate original imagery that is fit for brand use. As Parry noted, when that capability is reached it won’t just be photoshop we need to be on the look out for. A picture may be worth a thousand words but those words will almost certainly be fake news.
What will the impact be on my email marketing team?
All the usual fire starter questions were asked after the lecture. Will we all be fired? Is this the end of human creativity in marketing? What will the impact of GDPR be? Etc, etc. The answers, I believe, all offer more opportunities than problems.
Don’t fire your team just yet. Or, as Nicola said, fire anyone who’s only job is subject line generation. If that was all you did you’d be jumping from the roof in no time at all. What you should push your team to do is to beat technology like Phrasee. Have you ever felt smug because you beat your SatNav home? This is the same principle. The computer is there to offer you one answer, but it just means that you have to work harder at creative solutions in order to beat it.
As for GDPR — Phrasee could even turn it into an opportunity by generating the language needed to replace those pre-checked boxes that will result in consumer’s consenting to share their data.
Parry did speak briefly about the limits of personalisation. Working from the premise of a subject line single customer view, he pointed out that language is a weak signal on an individual scale. Due to the number of purely random variants, personalisation moves exponentially towards impossibility. Optimisation is therefore better served by focussing on segmented customer groups.
The free take home was that 2 weeks is the longest an idea should exist in your company before it becomes customer facing, and that all of your teams and your thinking should be aligned around the consumer journey, not the product itself.
I’m looking to connect with individuals and organisations scoping out trends and opportunities in this space. I’d love to hear your thoughts so please get in touch or leave a note in the comments.
Yourzine is open to collaboration, either through exchanging thoughts on the impact for your business, or through discussing possible partnerships.