Editorial
The Gist
- Scaling limitations. Generative AI struggles to efficiently create highly personalized emails without significant manual intervention, limiting its practicality at scale.
- Traditional superiority. Existing personalization methods and machine learning models are more effective, reliable and risk-averse than current generative AI approaches for email personalization.
- Unproven potential. The performance and return on investment of generative AI personalization in emails remain uncertain, with unclear benefits and possibly diminishing novelty.
Let me be blunt: We are so far away from generative AI writing personalized emails in any meaningful way. So very far. Yet, I keep hearing people suggest that generative AI will be writing highly personalized emails to individuals in the not-too-distant future.
I don’t think this is a realistic expectation at any significant scale, with any degree of automation and with a reliable expectation of increased performance — much less a positive return on investment. Let me explain why.
Not at Scale
The best example I’ve seen of generative artificial intelligence (AI) writing a one-to-one email was a meeting request for a salesperson. First, that’s a pretty generic request — and, in the example, amounted to 10 words. Not a big time-saver. Second, the salesperson is chaperoning the AI, and if they don’t like what’s written, they either have to edit it or write a followup prompt for changes, which further reduces the time-savings. In that example, the degree to which it was personalized was achieved through manual intervention and not automatically driven by the recipient’s past behaviors, demographics or other criteria — which is the traditional definition of personalization.
While generative AI will get much better in the years ahead, that example is a very realistic example of how generative AI would be used in a “blank page” email situation right now, given generative AI’s many faults and limitations, including its propensity to “hallucinate.” That use case appropriately minimizes the risk by (1) asking for a routine request, (2) keeping the text short and (3) having the output monitored and approved by an employee before it’s sent.
Related Article: The 5 Biggest Changes From a Decade of Email Marketing Change
Not Automated
To date, I’m not aware of any brands using ChatGPT or any other large language model (LLM) to incorporate personalized content into an email. And there’s a good reason for that: It’s unnecessary.
Traditional methods of personalization are highly effective and are far from fully utilized — due to siloed and unreliable data. And machine learning models for product recommendations and send time optimization are even less utilized. These existing tools are tried and true and offer solid returns with much more control — which is to say, with far less risk — than generative AI.
In fact, when generative AI is ultimately used for personalization, I predict it will be fed content that’s been personalized using traditional methods. Put another way, any personalization done by generative AI will be done on top of traditionally personalized content. That approach would minimize generative AI’s opportunities to introduce inaccuracies, biases, plagiarized material and other problematic content while simultaneously focusing generative AI on the kind of personalization that only it can do.
Just what are these new forms of personalization that only generative AI enables? Here are a few:
- Tone. Some subscribers are receptive to more aggressive pitches while others are much less so. And, of course, there’s a whole spectrum of potential tones that could be used. Generative AI could rewrite copy on the fly to better align with what subscribers respond to best.
- Vernacular. Language varies in significant ways by geographic region, education level, religious beliefs and other factors. Especially if it’s informed by sales and support correspondence, generative AI could adapt a brand’s messages to match the recipient’s language usage.
- Image backgrounds. Generative AI for images can enable brands to create highly personalized images based on the subscriber’s location, industry and more. For example, an outdoors retailer could place a model in any number of national parks depending on the location of the subscriber or knowledge of where they like to hike or camp.
As with all personalization, a big part of the challenge will be securing enough data to make sound decisions. But even if that hurdle is cleared, there’s the question of performance and generating a return on investment.
Learning Opportunities
Webinar
May
9
Driving Customer-Centricity in the Financial Services Industry
Discover how financial organizations are leveraging generative AI technologies.
Register
Conference
May
9
An Irish Roundtable in the Valley: The Economic Impact of AI on the World of Work Stanford 2024
Exploring the Impact of AI On the World of Work with Irish diaspora leaders and tech business champions from the North of Ireland
Register
Related Article: 5 Ways to Make Your Marketing Emails More Personal — Without Personalization
Unproven Performance
Oracle Marketing Consulting has identified more than 170 segmentation and personalization criteria that can be used in digital marketing campaigns. However, just because you can personalize a message using a particular data point doesn’t mean your subscribers will respond positively to it. Through experience and testing, each brand must discover which factors truly move the needle for them. The same is true for generative AI personalization.
Even if you have the data to try to personalize the vernacular of your emails, for example, it’s currently unclear if this would be a winning approach. It’s likely that at least some subscribers would find this kind of personalization creepy and manipulative.
It’s also likely that some performance boosts likely wouldn’t be sustainable as the novelty wears off. The history of first-name personalization likely provides a good example. A decade and a half ago when it was new, first-name mail merges moved the needle for a while. But it quickly became a hollow gesture and a gimmick because it generally didn’t signal that the email’s body content was any more relevant to subscribers. First-name personalization got brief bumps when brands gained the ability to personalize images with a subscriber’s name and again when they could personalize videos with their name. But subscribers still view these as technological stunts that don’t signal anything meaningful.
Generative AI personalization may have the same effect, drawing attention away from your message and to the technological stunt of personalizing an image with the skyline of the person’s home city, for example. Sure, it will be cool the first time you encounter it, but the novelty will fade quickly because it’s based on relatively easy-to-obtain geographic data that doesn’t speak to the person’s needs or wants. Plus, attention-grabbing tactics often don't translate into better performance.
Uncertain ROI
Even if generative AI personalization can boost performance, a positive return on investment is unlikely — definitely not at current “per token” price levels. And given email volumes, it’s possible that generative AI personalization may never make financial sense except for highly targeted campaigns.
It’s worth noting that many of our clients have seen lackluster returns when using predictive subject line writing tools like Phrasee and Persado. These tools have been around for many years and are primarily driven by machine learning models where wording recommendations are based on the historical performance of a brand’s subject lines and other copy. If models that are trained on historical performance can’t reliably generate a strong ROI, marketers should be deeply skeptical that generative AI tools with no access to performance intelligence can.
Muddied Terminology
A big contributor to this premature idea of generative AI personalization is that some people are using the terms machine learning, AI and generative AI almost interchangeably. While they’re all loosely related, they’re far from the same. They operate using different algorithms and models, have different goals, are built on different datasets and are appropriate for different use cases.
With generative AI and AI in general being so hot right now, brands need to work extra hard to be clear-eyed about what’s possible now and what may perhaps be possible at some point in the future. Just like in the late ’90s when adding “.com” to company names was all the rage, now .ai domains are super popular. Just like then, in some cases these are just aspirational or opportunistic marketing moves.
Generative AI will ultimately be huge, and brands should absolutely experiment with and get comfortable with it. However, we’re in the hype-iest part of the hype cycle, so proceed with caution and ask lots of questions.
Learn how you can join our contributor community.
