Imagine that you’re thinking of buying a new car. You open Chrome and check out the website of an automotive brand. You stop at the pages showing the specs and the galleries of a model you like. You open a new tab and Google for reviews. You read the reviews and watch some road test videos on YouTube, and then your partners opens Facebook and is presented with an ad from a rival company.
It happens, and it feels both creepy and remarkably clever. Software, most of it owned by Google, has tracked the content you’ve viewed on the Internet, and measured the amount of time you viewed each page to produce an analysis of both your interest and its depth. It’s figured out how close you are to a purchase and it knows what you’re interested in buying. At the same time, other software may have built a map of your relationship to others. Facebook knows who you WhatsApp most often and is able to map your communication patterns to known relationships. It can assume that if you’re thinking of purchasing a new car, your partner will also have influence over that decision too.
As soon as you start looking at the new Ford sedan, Chevrolet could start showing its ads to your spouse.
Marketing has never been more sophisticated. The days when automotive firms depend on their ability to distribute flyers across neighborhoods based on rough demographics are largely over. A Ford outlet might still mail a brochure describing its SUV deals to houses in affluent suburbs but it can get much better results pitching a video to people on Facebook who live in a certain area, are of a certain age, are members of groups about family life and children’s soccer, and who have already visited Ford’s website at least once in the last month.
But that’s only the beginning, software can also track the results of that advertising, using data to automatically adjust the messaging at an individual level and guide leads from curiosity to purchase without the intervention of a salesperson.
That use of artificial intelligence in marketing is happening already. It’s happening now on three levels… and while it’s not yet as intelligent as it appears, it is going to get smarter.
Artificial Unintelligence
We see the simplest kind of artificial intelligence in marketing every day. Whenever we open our email inboxes we see automated messages. The content of those messages is still produced by humans. Copywriters churn out the text in the order confirmations but software determines who receives them, and when. Users of services like MailChimp or Emailchef can create scripts that tell the company to send a particular message if a condition is met. The most basic condition would be a purchase for which the customer needs a receipt or a confirmation but companies can also issue messages to people who have abandoned their shopping carts or who have made a purchase and might need an accessory, or who haven’t returned to the site for a pre-determined amount of time since making a purchase.
At least some of that content can be automated too. The product recommendations that appear below items on Amazon pages use sales data to correlate products. “If you like this then you might like that” is a form of artificial intelligence that draws on purchase data to score the probability of someone who likes one product also wanting another.
The implication of those recommendations is huge. It suggests that businesses can make assumptions about individual tastes, and do it in a way that’s more accurate than we can do for ourselves. But the intelligence, while artificial, isn’t particularly smart. You don’t need the power of a quantum computer or the data available to Google to understand that if someone buys a hammer they might also be interested in nails. The data available to Amazon, however, allows it to make a much broader and accurate range of associations.
The same kind of simple intelligence often underpins chatbot programs. Businesses like Twilio enable firms to replace human sales staff with interactions that are intended to feel human but which usually feel far less. At their simplest, they’re little more than automated FAQs, serving chunks of pre-written text in response to pre-determined keyword triggers. That’s sufficient for the simplest of tasks, and frees up human staff for work that requires judgement and complexity. As Salesforce’s linguist and Lead User Researcher said in an article about the rise of chatbots:
“Out of the pool of problems your customers have, there are some that are best suited for a talk with a human. But that’s not something as common as ‘reset my password.’ Agents’ time is precious so save them for the complex stuff… Let the chatbot take care of the simpler jobs.”
The company cites a study that found that 53 percent of service organizations expected to use chatbots by the end of 2020.
Most of those chatbots will be unintelligent. They’ll guide customers through self-service options and they’ll collect information before passing the call to a human agent. They might help agents themselves as they handle cases. But as long as digital assistants like Alexa and Siri continue to struggle to return information more complex than today’s weather, chatbots powered by a low-level artificial intelligence are going to do little more than take some of the strain off human staff by performing the simplest jobs.
Artificial and Smarter
At the start of 2019, an article in the Australian edition of The Guardian opened like this:
Australian political parties declared donations worth $16.7m in the 2017-18 financial year, according to the latest figures from the Australian Electoral Commission.
This amount is lower than usual, with donations averaging $25.2m a year over the past 11 years.
It looks like a fairly standard lede but it wasn’t written by a reporter. The article was generated using ReporterMate, an experimental automated news reporting system. The software is just one of a number of programs that have moved AI on from delivering content and towards creating that content. In 2016, the Washington Post produced about 850 articles using Heliograf, a rival AI reporting system. They included 500 articles about the election that together generated more than half a million clicks.
Even here, though, AI is still low-level. Like chatbots that collect data or return prepared answers in response to keywords in questions, ReporterMate and Heliograf only insert basic facts into predetermined sentence structures. They’re adequate at reporting the result of a college football game and saying who scored how many points and when. They can state who won a local election. But they can’t collect the facts and build an argument to explain the result or its significance.
That means that even this more advanced AI has little use in marketing. A marketing version of ReporterMate can’t put together brochure copy that highlights the benefits of the product and quashes objections. It can deliver statistics but it can’t tell a story, display empathy, or react to a lead’s skepticism. All of those essential elements of the customer journey still require a human touch.
What current levels of artificial intelligence can do in marketing is to better target the content that marketers have prepared. Facebook marketers don’t just pitch ads to carefully defined demographic markets. They also pitch thousands of slightly different ads to slightly different demographics. The old A/B testing in which advertisers would deliver two different ads to the same market then wait a few days to see which performed best has now largely been replaced by continuous automated testing. As soon as it’s clear which ad is winning the most interest, the lower performing ad is stopped and resources moved to the higher performing ad. That ad can then be tested against another in a slightly different demographic so that marketers find that they’re using a constantly adjusting system that is always personalizing content at scale.
For leads and customers, the marketing material that appears in their inboxes and social media streams becomes increasingly relevant. It doesn’t just contain their name pushed into placeholders. AI has made sure that it contains the messages that have the highest odds of triggering the reaction the seller wants.
The Smartest Intelligence
That progress from automated FAQs and product pairings to the constant testing of slightly different messages to ever smaller demographic slices suggests that the future of AI in marketing is going to take three directions.
The first is the continued improvement of message/audience matching. The size of the audiences available to marketers means that results now flow from a test instantaneously. The adjustment of message to audience will grow ever faster so leads will see a higher proportion of well-targeted messages and fewer ads that miss. As the data those results generate continues to grow, the AI software will also be better able to predict which audiences will respond to which categories of message content.
AI will continue to improve targeting, and that will be its main contribution to marketing.
It’s also likely to continue to struggle to produce copy. The failure of cars automated enough to allow drivers to read the newspaper on their morning commute is a relevant warning. Elon Musk predicted that Teslas would be driving themselves in the next three to six months. That was in 2017. At the start of 2019, he promised that we’d be sleeping in the driver’s seat by the end of 2020. It doesn’t look close—and the reason it doesn’t look close is that complete automation is hard. Software like ReporterMate and Heliograf can reel out stats but anything more complex and the sentences soon start to break down. As the Economist recently noted, Google can easily suggest “Birthday” as the word to follow “Happy” in an email. But it can’t figure out the next sentence. It will be a long time before it can, and the complexity of language suggests that at least for the foreseeable future copywriters will still be crafting the words themselves, even if robots are giving them automated hints as they type.
Because AI development in both targeting and messaging has already picked the low-hanging fruit, the big breakthroughs in the near future are likely to come not in better targeting or in automated, more effective copy. It will come in areas that AI has so far barely touched. Sentiment analysis has tended to focus on identifying keywords in text and applying a pre-determined score. Some AI companies are now extending that service into speech. Nexmo, for example, offers “real-time insight into caller sentiments and emotions” to enable “better decision-making, support, and outcomes.” Once AI has learned to understand text, then speech, expect it to move into non-verbal communications next.
Conclusion
In 2012, The New York Times reported a story about a man who had walked into a branch of Target near Minneapolis and demanded to the see the manager. The man was holding coupons that had been sent to his daughter.
“My daughter got this in the mail!” he said. “She’s still in high school, and you’re sending her coupons for baby clothes and cribs? Are you trying to encourage her to get pregnant?”
The manager was confused but he apologized. A few days later, he called the man to say sorry again. This time, the man was apologetic himself. He had spoken to his daughter… and she was pregnant.
Target’s marketing AI software had analyzed her purchases, which might have included a pregnancy test, and assumed that she would now be interested in baby’s clothes. It was an intrusive piece of marketing that Target has since changed. It was also an example of artificial intelligence that required little intelligence: if a woman buys a pregnancy test then stops buying sanitary products, pitching baby clothes seven to eight months after the purchase of the test is an easy marketing move.
Since that incident, AI software has moved on. Its targeting in particular has continued to improve. But as the technology and the data to support it grow, we’re more likely to see ads becoming increasingly relevant and occasionally spooky. But we’ll still be closing the sales with an agent—even when the agent is a car salesman.
Recent Comments