Exoskeletons, not cyborgs

Stanislav Yevgrafovich Petrov was an officer in the Soviet Air Defence Forces during the Cold War. In a particularly fraught period in the early 1980s, just three weeks after the Soviet military had shot down a commercial flight operated by Korean Airlines, Petrov, acting as the duty officer at a nuclear early-warning facility, made a decision with staggering ramifications on geopolitical stability and, indeed, the welfare of mankind: Petrov deemed a warning of an impending nuclear missile attack by the United States to be a false alarm, and he dismissed it. Petrov’s decision likely prevented the Soviet army from launching what they would have deemed to be a retaliatory strike. The system warning was, indeed, erroneous; it was caused by an unexpected alignment of the sun with the high-altitude clouds that the Soviet satellites orbited. Petrov was ultimately demoted and suffered a nervous breakdown.

More recently, two developments have animated a newfound enthusiasm for digital advertising automation:

  • The success of automated advertising products from Meta (Advantage+) and Google (PMax), which fully automate campaign optimization on behalf of advertisers (more background on these products in Google’s PMax, Meta’s Advantage+, and the logic of total advertising automation). Meta characterizes Advantage+ as artificial intelligence (AI); I’d classify it as machine-learning-driven automation;
  • The mass proliferation of Generative AI tools, built atop large language models (LLMs) like GPT-3 and text-to-image models like Stable Diffusion, has created an environment of total abundance of assets like advertising creative and copy. Advertising teams can produce these assets quickly, at a low cost, and with minimal human intervention.

It’s not challenging to envision how these two types of tools might be combined to produce a wholly-automated, end-to-end marketing function that requires very little human input or oversight. I described such a fully automated system in The five levels of digital marketing automation, in which I map four features of advertising infrastructure — ad creative production, audience testing and definition, campaign settings experimentation, and ROAS projection and reporting — to a classification system for various degrees of automation. In my rubric, a Level 0 system is totally manually operated and a Level 4 system is not only fully and completely automated, but also entirely integrated. From the piece:

Full marketing automation includes all of the features of level 3, except that the connective tissue between each of the four functional buckets is complete in such a way that all decisions and reactions are implemented automatically: campaigns are automatically created and tuned, creative concepts are automatically defined, produced, and deployed to campaigns; ROAS models and timelines are automatically calculated and used to adjust campaign settings, etc. At this level, theoretically, no marketing team is needed beyond the engineers and analysts that maintain the automation framework.

In the piece, I make the point that most advertising teams that I had interacted with — the post was written in 2020 — were operating at Level 1. This is to say: most advertisers used some automation to assist with repetitive / low-leverage tasks, but the overall process supporting their work was mostly managed through manual effort. On the benefits of Level 1 marketing automation, I wrote:

But the value of even Level 1 marketing automation shouldn’t be discounted: when the tedious, manual work required of most marketing managers is removed, those people are given the opportunity to be much more analytical and strategic. Level 1 marketing automation is a laudable goal, and the robustness of Facebook’s and Google’s APIs makes it fairly straightforward to implement.

I still believe that most marketing teams operate at Level 1. The difference between the marketing environment today and the environment that existed in 2020 is that the ambition to achieve Level 4 marketing automation feels more pronounced and commonplace now as a result of the tools identified above becoming broadly available at low cost*. A team can look to Generate AI platforms, then look to platform machinery like Advantage+ and PMax, and assume that achieving Level 4 automation is as straightforward as making a few API calls.

But like all forms of automation, reaching truly autonomous status with an advertising system is deceptively complex, distant, and, for the vast majority of advertising teams, utterly unnecessary. I believe this is the case for a few reasons:

  • First: as sophisticated as the largest platforms may be, there are components of the digital advertising supply chain that no advertiser should be willing to entrust with them. The most obvious of these is creative production: automated advertising systems will optimize ruthlessly against advertiser-designated conversions, and without advertiser control over brand standards, platform-generated creative will quickly adapt to whatever imagery or aesthetic delivers those outcomes most plentifully in the short term. Most advertisers should want to exert careful control of their brand imagery, and therefore they cannot risk granting creative production control to the same platforms that control campaign delivery and spend, given the potentially-conflicting incentives. An advertising system optimized solely for immediate conversions and controlled entirely by the platform managing campaign spend likely converges around problematic ad creative for every advertiser. The fake mobile gaming ads phenomenon supports this hypothesis;
  • Second: media spend represents a considerable expense for many digital advertisers — in some cases, the largest single expense line item. Real money is at stake with digital advertising, and advertisers should be reluctant to surrender control of that spend to a system that could very well make mistakes. Some automated data products, like recommendation systems, have relatively contained downside risk; advertising systems are fragile and comprised of many components, and an edge case or unforeseen feedback loop in an automated advertising process could cost an advertiser an unacceptable sum of money;
  • Third: from my experience, advertising automation follows the Pareto rule, with 80% of the benefit being derived from the first 20% of the effort of building the machinery. Creating a fully automated advertising system simply won’t be economical for the vast majority of advertisers. The most consequential step change in efficiency with advertising automation is the graduation from Level 0 to Level 1. My sense is that a very small percentage of digital advertisers — maybe a few dozen in total — spend enough money to justify the expense of implementing total, Level 4 advertising automation, and that the vast majority of advertisers would experience very little commercial benefit beyond Level 2 automation.

Advertising and content automation is most productively utilized in the form of exoskeletons, not cyborgs: it’s often more economical and efficient to build tools that unlock additional productivity from members of a team than to build tools that are designed to entirely replace them.

While I’m excited about the prospects of Generative AI for advertising creative production, for reasons that I articulate in Abandoning intuition: using Generative AI for advertising creative, I think it would be a mistake for advertisers to seek to entirely replace their creative production teams with Generative AI tools. Rather, my sense is that advertiser teams are best served by utilizing Generative AI as an exoskeleton: a device that allows the creative team to experiment more rapidly, or more dramatically than they otherwise could with their existing support infrastructure, while still retaining creative and thematic oversight on the finished product.

And the same can be said for tools that automate budget allocation and spend, given the direct contribution of ad spend to revenue for direct response advertisers. Any component of an advertising system would be difficult to fully automate independently; integrating those components into an entirely automated system is not only onerous and expensive, but it presents a risk that should be intolerable to most advertisers.

And while these systems will surely progress over time, the pursuit of totally autonomous technology — across any use case — is redolent of the Philip K. Dick short story, The Indefatigable Frog. In the story, a physics teacher attempts to disprove Zeno’s Dichotomy paradox, which states that sequentially cutting a distance in half will prevent the entire distance from ever being traversed, and he devises an experiment that ends disastrously. Each progressive step forward from the here of automated advertising towards a Level 4 system might be smaller than the last and may flirt with similarly disastrous consequences. While the stakes are lower in building automated advertising systems than in building automated nuclear attack defense systems, the world is grateful that Stanislav Petrov’s judgment could supersede the output of his computer.

*Advantage+ has actually existed since 2020, which is when Meta — then, Facebook — introduced its Automated App Ads, or AAA, campaign strategy. AAA was expanded to include web-destination ads in 2022 and re-branded as Advantage+.