AI prompting for marketers—lessons on how to best use generative technology

Imagine a super-capable being descends into your workplace, but a language barrier is preventing you from accessing its full potential. This scenario is not all too different from the emergence of AI in the advertising world, which is requiring marketers to learn a new skill in order to leverage the technology’s capabilities: prompting, or crafting an input to task a generative AI platform. Last year, 70% of CMOs provided training on effective use of AI, many of whom are continuing these efforts in 2024, according to Accenture survey data. But upskilling in prompting can be complicated. Not only does an AI model resemble someone who doesn’t speak your language, but its pure regurgitation of training data reflects the learning process of a child, said Dave Harris, principal generative AI author at workforce development company Pluralsight. Marketers are thus using formulas to condense prompt information and saving more technical tasks for specialized staffers, experts told Ad Age. AI marketing blog How brands are using ChatGPT, DALL-E 2 and other tools Read more here The requirement of new AI-related skills comes as the technology transitions to playing a more front-seat role in the ad creation process. Agencies are now using generative AI to help clients create commercial work, and social media platforms are providing the tools to enable promoted posts on their websites. More news: 5 AI trends for marketers to know about right now Marketers heeding these calls are learning to prompt in numerous ways. But according to Harris, the most common approach is through sheer trial and error with the platforms themselves. “We always go in with the idea that AI has the answer and we just have to provide the right information,” he said. When experimenting, Harris suggested that a marketer ask the platform many questions and refine prompts to see how different inputs engender different outputs. And even though the platform may very well contain useful information, a heavy dose of skepticism should be treated for every response, and calls for internal review, Harris said. A growing number of agencies and tech platforms offer AI sandboxes, which allow for prompt experimentation, as well as mixing in proprietary data with AI models, in a way that is secure against any possibility of a data breach. Enterprise clients who don’t pay for AI platforms such as ChatGPT, and thus have access to only the public applications, risk surrendering any input data to the underlying model, which trains on everything with which it comes into contact. Other resources that marketers may find helpful are prompt galleries, which AI providers typically post on their websites, said Paul Roetzer, founder and CEO of Marketing Artificial Intelligence Institute. These libraries offer a number of structured prompts and their corresponding outputs, giving users a peek into which inputs work well and can help them improve their own. “If Anthropic and OpenAI are telling you how to write these prompts, then use that information,” he said. ‘Context is king’ According to Harris, the key to effective prompting has less to do with understanding the AI than understanding the problem at hand. The more one knows about what they’re looking for, the more information they can input into the system and the better the output will likely be. “Context is king,” said Harris. This approach also benefits from cross-platform usage. While some degree of prompting skills will carry over between different bots, Harris said, nuances between how the models are trained will result in different responses, even if the input is the same. The ideal method is to be able to provide sufficient information regardless of these discrepancies. Using a formula can be an effective way to compartmentalize information in a prompt. A creative lead at an ad agency, who spoke on the condition of anonymity, suggested three steps. The first is to narrow the AI’s focus by invoking an area of expertise. For example, prompting a bot to “pretend you’re a Gen Z consumer” will immediately prime it for the kinds of questions that a marketer may want to ask about the demographic. The second step is to provide all of the relevant context. The same marketer would input any background specificities, including their brand and the type of products they sell, as well as their goals for their campaign, their stylistic preferences and the sentiments to which they want to appeal. The third step is to ask for the actual output, such as “generate 10 tweets for the campaign targeted to Gen Z consumers.” More from Ad Age: 5 Gen Z trends for brands to watch Roetzer uses a template of his own when explaining how to prompt, which he borrowed from Wharton School of the University of Pennsylvania Professor Ethan Mollick. After narrowing the model’s focus, the approach articulates step-by-step instructions for the intended output, follow-up directions, constraints for what not to do and a reminder to personalize the information. As a result of its applicability to a variety of topics, the template has helped Roetzer teach prompting to many different audiences, including golf course owners and deans of schools. Another lesson he’s learned is the importance of tamping down the intimidation that can come with communicating with AI. “Just talk to it like a really smart intern or peer,” he said. Does everyone need to know how to prompt? A question that any marketing organization will need to ask itself is whether, and to what extent, AI prompting skills will need to be learned by all employees. Is prompting a specialty skillset, like coding, or is it a new table-stakes expectation, summed up by the popular and somewhat threatening quip “AI won’t take your job, but someone who knows how to use AI will?” The term “prompt engineer” has emerged, but experts draw a distinction that may soothe any worrying marketers. While employees should be encouraged to learn general prompting skills, the technical requirements of a prompt engineer are beyond what is expected of any given copywriter, according to Michael Cohen, global chief data and analytics officer for agency holding network Plus Company. Prompt engineers, said Cohen, are better understood as systems engineers because they do far more than just prompting. They leverage models to generate outputs, and then chain them to other models to establish a workflow. For example, they might tweak one platform to neatly present audience data, which is then fed into another platform that can design a campaign around such data, before being connected to a third platform focused on optimizing campaigns. Plus has developed a team of voluntary lead users to spearhead prompting experiments with AI, cultivate solutions and share these findings with other personnel. The precise information that is disseminated depends on the agency, but once in their hands, these solutions become equivalent to classic software suite tools. “I would expect every employee to use [them],” Cohen said. The anonymous agency lead equated prompt engineers to programmers responsible for building applications and adopting and evaluating new tools. Their agency has a team of five people dedicated to these tasks. General prompting, on the other hand, is basic communication with a pre-existing application, which is encouraged to all the agency’s employees in an effort to trigger for them a moment of realization that AI can succinctly do what they previously labored over themselves, the agency lead said. Learning how to prompt carries significant weight for today’s marketers, said Roetzer, but while employees should experiment and acquire this background in the short term, there could be reasons for hard prompting skills to become less important over time. This is because of the AI advancements happening behind the scenes. When a user prompts an AI platform, it restructures the prompt to enable the most effective interpretation and generation. Roetzer envisions, in even just a year, that AI models will get so good at this that even inputs previously considered incomplete will result in complete outputs. Intent, not information, could become the most important part of tasking an AI. “They’re just going to rewrite the prompt and you’re not going to know it’s happening,” Roetzer said. Sign up for Ad Age newsletters From influencer marketing to agencies, get the latest news and analysis delivered to your inbox Click here

AI prompting for marketers—lessons on how to best use generative technology

Featured Stories