Generative AI for Marketing: Beyond Simple ChatGPT Prompts

Congratulations on becoming an AI assisted marketing ninja! You managed to whip ChatGPT into shape to create a months worth of social media posts, a dozen unoriginal but somewhat functional blog post headings and even some respectable sounding ads. You and your marketing team are now saving a bunch of time in your workflow and it’s really started to impress your executive team. What’s next?

Are you struggling with this phase? If so, you’re not alone. We call this phase the “Prompting Plateau.” Here’s what it looks like: You started with generative AI and felt incredibly energized and optimistic about the future of marketing. Your first forays into using an AI generator seemed miraculous—and then reality set in. Suddenly, you’re not creating magic. You’re just working slightly differently, with a dash of novelty thrown in. Instead of being ahead of the competition, you’ve managed to slightly improve the status quo. Instead of being new, you’re just using a slightly better word substitute. If you’re looking for help implementing this, talk to our team.

Using generative AI as a command-line intern is low-use. Writing a slightly better one-off prompt to a black box isn’t going to shift your marketing outcomes in a material way. This work is only high-use if you’re designing automated workflows that are powered by the changing dynamics of data and the new abilities of marketing tools. We can only start to think of ourselves as AI marketers when we start to think of ourselves as AI strategy creators.

TL;DR: Your AI Marketing Maturity Path

• The Problem: Relying on manual, one-shot prompts in ChatGPT creates generic, unscalable content and keeps you stuck in a cycle of copy-pasting. This is the “Prompting Plateau.”
• Level 1: Build a Content Engine Chaining API calls to move beyond a single prompt, you can build a content engine that enables the creation of content workflows with multiple steps. Such a workflow can help to take an idea all the way through to a fully formed document in a repeatable fashion, while ensuring that all content is of a consistent standard.
• Level 2: Hyper-Personalize with Your Data. Connect large language models (LLMs) to your first-party customer data (from your CRM, CDP, etc.) to generate really personalized messaging at a scale that was previously impossible.
• Level 3: At the Level 3 advanced, the marketing automation team will learn how to design an artificial intelligence (AI) system that can manage multiple marketing automation tasks autonomously. Marketing automation experts at this level will be capable of designing the architecture that enables an AI system to automatically monitor, assess and take marketing actions, examples of which include determining the sentiment of social media users or qualifying leads without any involvement from marketing personnel.

The ‘Prompting Plateau’: Why Your AI Marketing Efforts Stall

The honeymoon phase of learning how to use a new tool is one of my least favourite phases of anything. I recently asked 219 marketers about their use of generative AI (yes that is a real thing) and here are the key findings: The average marketer is saving 3 hours of work per week and the most common use cases are for content and data analysis. While saving 3 hours of work per week is better than not, it is not a strategy.

When students reach what we call the Prompting Plateau, it means that the efficiency gains that students achieve from using the supported writing strategies and prompt structures we provide, stop growing. There are three reasons this plateau typically forms.

1. Context-Free Content is Generic Ask a public model such as ChatGPT to write about your industry and you’re at the mercy of the open Web and the data set it has learned from – the same open Web your competitors have learned from also. What you’re presented with is Context-Free Content. Generic. Uninspired. Unoriginal and not remotely nuanced to your industry, data or brand voice. This kind of content may at best, exist but at worst is little more than a marketing checkbox that can barely manage to sustain a reader’s interest before they’re likely to nod off.
2. Manual Labor Doesn’t Scale We’re guessing you’re still doing this: copying and pasting from a Google Doc into the ChatGPT interface, copying and pasting the output back into your system of record, and so on. And we know that this flow is not scalable for you or your organization.
3. One-Offs Aren’t Systems Great prompt: great answer. It’s very easy to believe that what looks like an awesome prompt for a particular situation must be the key to opening up an awesome answer. To understand why one-offs and systems are very different things, consider this: one great answer to a few amazing questions does not add up to a conversational AI system. Instead, a good system means being able to come up with thousands of good answers over many years. It’s easy to believe that coming up with that one amazing answer for a certain question will solve the problem, so instead of building something really systemic (something that works at scale across many cases), we treat each one-off instance of the same question of getting that great answer as enough. By doing so, we’re having a solo experience. The experience is not a team effort in which we’re working with a system and the system has “won” the first round of the human experience that we’re participating in. As a result of believing one-off instances will carry us through a whole system, we’re more likely to have an inconsistent and poor app experience, an app that may require only a very specialized kind of knowledge and set of particular behaviors and patterns of engagement, and ultimately, a system that no one human could sustainably enough to really understand, much less maintain.

Its very easy to get stuck here. Realizing that it is a different mindset than instructing an AI to “do something” as opposed to designing a system which “runs a process”.

Level 1: Building a Content Engine with Chained Prompts & APIs

For starters, the first step off a plateau is always the hardest. Primarily, you need to stop thinking about individual input prompts. As a content engine, you are a workflow. A content engine is not one perfect prompt. It’s a series of smaller, more narrowly defined inputs that are connected in a specific order, with each subsequent input being generated from the previous output.

Wow this looks like a ridiculous amount of copy and paste work. Here is the work flow using the API to do the task instead.

Example: The “Topic to Tweet” Engine

In this video we’ll be building a “content cluster” around a blog post. Essentially, a content cluster is a cluster of social media content. In order to make this process easier and less time-consuming, we will be building an automated chain.

• Step 1: The Topic Extractor.
• Input: The full text of your new blog post.

These include potential downgrades to the nation’s financial rating and the fact that a third of the original stimulus was set to expire at the end of this month. In addition, the unemployment rate has reached a milestone that is historically difficult to breach, while an unrelated policy has just gone into effect. The full range of effects from this change remains to be seen, with both positive and negative outcomes possible.

• Output: A structured list of key ideas.
• Step 2: The Angle Generator.
• Input: One of the sub-topics from Step 1.
• Prompt: Take the following topic: "[sub-topic]". Generate 3 different angles for a Twitter thread: one controversial, one instructional, and one that asks a question to the audience.
• Output: Three distinct creative directions.
• Step 3: The Thread Writer.
• Input: The “instructional” angle from Step 2.

Building on the concept of instructional angles, this next post in our series focuses on write angles. Write angles are clearly demonstrated by labeling each vertex of the shape, with the word at each point telling you which side it describes. For example, if you looked at the vertex at the top, you would see the word ‘top’ written there. What’s key to understanding write angles is realizing that the order in which the sides are named matters. Work your way around the polygon and name each side from left to right, making sure to write in the correct direction as you go. Check out the full blog post to see these concepts in action and understand more about write angles! #writeangles #geometrykeys #mathstrainingresource #supportingmath achievement Read the full blog post here: [link to blog post]

• Output: A ready-to-schedule Twitter thread.

You connected all the workflow steps through an API. Now you have a machine: feed it a blog post and it will yield 15 social media headline options and several pre-filled social media posts, without having to visit a single web page. Automate and simplify the work, to free up time for more strategic and creative input. Prompting the tool is now a matter of-course, while your focus shifts to running the machine.

Choosing an API for your project is the first step towards putting AI to use. It’s not super complicated, but it does mean learning a bit about how APIs work. (Don’t worry – you’ll be rewarded with a massive pay off for only a few minutes of relatively simple work. If you’re just getting started, consider starting with our guide for Choosing the Right LLM API.

Level 2: Hyper-Personalization Using Your First-Party Data

The moat you defend, is the moat you create. While everyone else is busy scraping data, left and right, ad-infinitum, in the hope of having that piece of content here, and there, occasionally (a few days, a week) you could hook your LLM (large language model) to first-party customer data, that is more than “fresh enough” to the task at hand, creating output of far better quality.

McKinsey estimates that companies that are good at personalization earn 40% more than those that are not. With all the hype about personalization, it’s understandable to wonder what’s in it for your business. We’ll get to that, but first—why is personalization so hard to achieve in the first place? The answer is because achieving true 1-to-1 personalization has historically required an enormous amount of work. And that’s why Tendril, for the first time ever, is making it possible.

Use your CRM, e-commerce site or data warehouse to feed data into your AI. Personalise your AI by creating hyper-personalised content for each individual using their data as input to the prompt.

Example: The AI-Powered E-commerce Cross-Sell Engine

I sell some fairly extravagance level kitchen gadgets and tools in my online store and on Friday Jane came in and bought one of the most expensive espresso machines we have for sale. It’s a $900 machine! So as ridiculous as a generic confirmation email is and needs, I thought up-selling with a generated AI email that might actually work in this case. Love to know if it will generate any sales.

Here’s the system:

1. The Trigger A purchase is made. A purchase is made and your store’s e-commerce platform (like Shopify) sends a purchase notification to Fulfillio as a webhook, which includes details about the purchase and information about the buyer.
2. So after adding Jane as a contact to our intent and identified a few of her purchased items we now take it to the next level by adding the data enrichment capability. We are now sourcing her purchase history through integration with our CRM. So here we can see Jane has purchased a Prima Elite coffee grinder, some glassware, specifically double walled glasses.
3. The Personalized Prompt: The data for this field will automatically populate in the Master Prompt Template.

Here are some details about this order: * Customer Name: Jane. * Item(s) purchased: La Marzocco Linea Mini Espresso Machine. * Additional items purchased with the espresso machine: Baratza Sette 270 Grinder Fellow Double-Walled Tasting Glasses Our customer service voice is always professional, yet approachable. Use the provided details to create a more natural & human email that is also within the suggested word count.

1. Congrats on your new coffee machine You coffee snob us about how many awesome espresso shots you’re going to be able to make with this thing!
2. We suggest the Acacia Lunar Scale and an Artisan Roaster’s Choice coffee subscription to go with her new machine.
3. For every suggestion put the sentence that explains why the suggestion is relevant to her situation — this should be a small sentence that describes how her situation links to the suggestion that has been made.
4. Action: This takes about 4 seconds in total. Our LLM generates the copy for our email, sends it to the email service provider (for example Klaviyo) and the email is sent to Jane.

As I looked at the end result it clearly looked as though it was written by some sales person who had done his homework and could tell an awful lot about Jane’s tastes and preferences. And I am getting this email from a seller to each and every person on my buying list, on the occasion of each of their purchases. Possible marketing even from a few years ago? Not really! Possible marketing, even as I write this, as a totally automated and even as a manual process? I hardly think so. You know as well as I do that such marketing as I’m describing above, after hundreds of purchases with hundreds of different buyers — yes you know it can’t be human or even automated for any length of time without breaking the bank — has got to be the work of “Artificial Retail Intelligence” I’m going to call the technology behind such marketing, “custom AI“ — the technology that will change the potential of marketing from a seldom and seldom effective process to being a source of profit, significant and dependable for a multitude of businesses across the world.

Level 3: Creating Autonomous Agents for Marketing Tasks

If Level 2 is about giving data to the AI, Level 3 is about assigning a task, and giving all the necessary means to complete it. So, moving from an artificial intelligence that you only give a few instructions, to a really autonomous agent: not just a “do this” — “that’s the answer” type of AI, but really an entity that will carry out a process of any length, using any means that have been assigned, and which may choose the route it takes and even modify its trajectory during the execution of the task.

According to a Salesforce report, Marketing

Ready to build your own custom AI marketing engine and move beyond the prompting plateau? Talk to our team of experts today for a free consultation and discover what’s possible.