We Built an Artificial Intelligence and it’s Scary A.F.
March 15, 2017
Scary good I mean. I’ve stayed schtum on the topic until now because – well because we’ve only gone ahead and actually done it. We built a marketing A.I.
HERE’S THE (RATHER LONG) STORY.
About a year ago – at this precise show in fact – I bumped into a guy who had just watched me give a keynote about the modern B2B buyer’s journey – more specifically our research showing that large chunks of the widely accepted theories, frameworks and narratives were – well frankly bullshit.
Anyhoo – so this guy was interested and started asking odd questions and hypothesis and long story short (jumping eight weeks into the future) I was sat with a sizeable group of Machine Learning and A.I. coders looking to operationalise our theories. And I was now co-founder of a startup tech company – Soto.
BUILDING A VIRTUAL MARKETING HAL 9000.
Our goal was to build out a digital assistant, a virtual marketer who could think critically and creatively in the same way marketing strategists do. Not Marketing Automation – no that’s pretty dumb, relying on humans to predict and plot journey’s. No we’re talking Autonomous Marketing.
THE PROBLEM WITH A.I.
There were two immediate challenges:
- We were going to try and teach the A.I. all of Protocol’s theories, frameworks and templates – and honestly many of these were human designed processes that hadn’t had validation beyond lots of humans using it. What if… our Protocol protocols were bullshit?
- Just how good was the tech? We are now dealing with huge chunks of theoretical, highly subjective and non-structured data. Would the machine be smart enough to make sense of it?
More importantly would I design something that a) might make myself redundant or b) try and kill me in my sleep?
HOW TO TRAIN YOUR A.I.
So yes my biggest fear was that the A.I. wouldn’t be able to understand the ways humans work. The nuance, the subjectivity and all of those hard to explain networked thoughts and reasonings that go into making a piece of content, writing a copy headline etc.
Without giving away the secret sauce (which in reality is a year long cooking experiment where we through out the recipe and started again like, 12 times) we started with something simple : Protocols Purchase Decision Model. The PDM is a buyer’s journey value proposition hybrid thing that our research has helped to shape. It looks at key messaging types and stages and is what we use to build our Content Marketing Strategies. It even goes as far as to recommend %portions of content types by ownership and stage.
WE’VE READ IT HUNDREDS OF BEDTIME STORIES
Our A.I. is a software tool to help professional marketers to get the job done better, faster and cheaper – which means it needs to think like a seasoned, educated and experienced marketer. So, just like a child we had to teach it the framework fundamentals then let it read. The A.I. understood natural language structure and nuance in the way that Google Assistant, Alexa or Siri does and even sentiment.
We let it read a lot – it’s goal to break down sentences into one of the categories above. At the time of typing it’s read, analysed and categorised something like 20,000 sentences – all guided by a human helping hand.
WHAT WAS THE POINT?
Well – we wanted to see if the machine could learn to find, digest, score and categorise content – like a human would do during a content audit. Effectively ‘curating’ on the fly.
What we know is, most large B2B organisations don’t really struggle with a lack of content. Sure they might struggle with a lack of awesome content – but generally they have hundreds if not thousands of pdfs and web pages floating about.
A human being might take days or weeks to trawl, read and index it all – and that’s what we were trying to solve – and have.
After it’s training the machine was able to mimic a veteran content marketer and match or exceed their accuracy in certain situations.
Well so what two things:
- Our Protocol frameworks have been validated by machine learning (which I can’t tell you what a massive relief that was), and
- Yeah that means it can make the same ‘decision’ as a human counterpart 84% of the time. Reading complex marketing documents and analysing their contents – cool right? Scary cool. Our line in the sand was to hit 75% accuracy (which is within the bounds of human to human consistency in many cases) so 84% is insanely cool.
WHAT ARE THE ETHICAL CHALLENGES?
I imagine there are some reading this with excitement – and a few with horror. The latter group worried it will replace human jobs. Which – honestly it probably will however we believe this will help many firms who don’t have the budget, time or resource to plan and design their marketing strategies as well as they’d like. Agencies too could benefit from leveraging a trained A.I. to do much of the heavy-lifting.
On a more existential level technology has always automated processed and workforces continue to adapt and evolve – for me this technology does much of the donkey work and drudgery that many highly educated and experienced people have to do – allowing them more time to work with the results and apply their critical and creative eye to the next stage.
Do you have any thoughts, feedback or comment? We would love to hear from you.
Share this Post