How does effective use of AI tools benefit your organization? 

It can process and create enormous amounts of text very quickly – AI algorithms and tools give us the ability to create better quality content in less time. And they’re improving all the time; OpenAI just announced (summer 2023) that they’ve increased the context window for the GPT-3.5-Turbo model that powers ChatGPT from 4,000 to 16,000 tokens (essentially 20 pages of text) for the prompt. 

AI is far lower cost than humans – the different types of artificial intelligence tools are substantially less expensive than human beings that have salaries, wages, and healthcare costs. The per-token price for the GPT-3.5-Turbo model that powers ChatGPT, if you use the web interface, is $0.002 per 1000 tokens—ridiculously cheap. This is what everyone (marketers included) is looking at when thinking about how these tools should be leveraged and the benefits they bring to the organization. They’re about making your work better, faster, and cheaper.

How will AI change the jobs of marketers?

Marketers will become the orchestrators of AI tools – In the dynamic world of marketing, we are no longer merely practitioners. We are evolving to become the maestros of AI tools. Just imagine the power and precision of an orchestra, where every instrument has a role to play. As marketers, we direct this symphony, reducing the need for a multitude of players. Our newfound ability to harness AI means an organization may require only a single, knowledgeable conductor instead of a forty-piece ensemble. By learning to wield these tools effectively, you can amplify your impact and become a transformative figure within your organization.

Marketers will become programmers – Marketing professionals are evolving into a new role – that of programmers. Using a generative AI tool equates to coding, albeit in plain English. This might seem strange; it’s programming that strikingly mirrors the memos you’ve been circulating around the office. As marketers, we need to slip into the shoes of developers, shaping a streamlined version of the software development lifecycle. And in this process, we must weave in vital considerations like business and technical requirements for models. By doing so, you unlock a new dimension to your marketing prowess, enhancing your strategic vision and enabling a more refined approach to problem-solving.

Marketers need to think about how to get around AI in their channels – An increasing number of marketing channels, particularly in the digital space, are seeing AI act as an intermediary between marketers and their audiences. It’s as if a digital wall has been erected, separating us from our target demographics. Search engines, Facebook, LinkedIn – they all play gatekeeper, and if you’re looking for a reliable, unobstructed reach, none of these platforms offer a guaranteed solution. The exceptions may be Slack, Discord, or Telegram, but only to a certain extent. It’s imperative for us, as marketers navigating through the AI landscape, to devise ways to circumnavigate these AI gatekeepers to ensure our message reaches the intended audience. By outsmarting the AI machinery of others, you can seize control, ensuring your voice remains loud and clear, cutting through the digital noise.

Tools and Use Cases 

What are the different categories of AI tools available for marketers? How should you think about choosing tools?

There are two fundamental architectures for AI tools:

• Transformer model – used in natural language processing tasks, these work on a principle called attention; they are large indices of probabilities trained on text that can take a string of words and predict the next logical word. Transformers have had a significant impact on NLP, powering state-of-the-art language models and achieving impressive language understanding and generation capabilities.
  • Examples: GPT-3.5, GPT-4, PaLM 2, Vicuna

• Diffuser model – a machine learning model designed to generate diverse and varied outputs from a given input. Unlike traditional models that produce a single deterministic output, a diffuser model introduces randomness and uncertainty to generate multiple plausible outputs. Many of these are based on indexes of captions on images programmed from art museums, news stories, etc. Then, you can give them a prompt, the model will compare the language in the prompt with words related to images, and create an image or video that mirrors the language in the prompt.
  • Examples: Stable Diffusion, Midjourney, DALL-E

Some transformer models have been wired to APIs – you’re starting to see a hybrid model now, e.g. with Microsoft CoPilot or Adobe Photoshop, where a tool allows generative prompts that feed into an LLM—so transformer architecture is wired to the internal programming language of a piece of software. Adobe hooks it into their programming language for Photoshop and Microsoft wired ChatGPT to the Microsoft Knowledge Graph so it’s able to control other parts of Microsoft Office.

Models are split into public and private:

• Public models – like GPT-3.5 or GPT-4. You don’t have control of them, they’re built and administered by their respective companies. They’re very large models that are jacks of all trades and masters of none. They are not superior at any one thing. These are good for general, safe, and public use cases.
  • Examples: GPT-3.5, GPT-4, PaLM 2

• Private models – these are models built off of publicly-released open-sourced data (for example, Facebook’s LLaMa) and the community takes that and tunes them for specific use cases. The open-source models tend to be tuned to be very good at one specific task. 
  • Examples: GPT4ALL, Cobalt, Karen the Editor, StableLM, LLaMa

How can you evaluate and choose a model?

Consider the TrustInsights.ai 5P Framework™:

• Purpose – what is the purpose, what are you trying to do?
• People – who are the people who will be using the tool and how technical are they?
• Process – what are the processes you use now? Document these extensively.
• Platform – map the platforms so you can see where different tools will slot in. 
• Performance – evaluate whether you did what you wanted to accomplish. Was time, money, or headcount saved?

Don’t think about tools before scoping what you want to accomplish – companies that go with a tool-first approach end up overspending without results. They don’t get the result they want because they weren’t clear to begin with. It’s like going into the kitchen and starting with the blender before you even know what you’re going to eat. 

What are the different categories of generative AI use cases within the marketing function? What tools can help with each?

Prompt Management

What are the four stages of leveraging AI? How should you progress through the stages? 

The four stages of leveraging AI are:

1. Prompt Engineering – designing and refining prompts to get the desired output from AI models.
2. Prompt Deployment – implementing prompts in real-world applications and monitoring their performance.
3. Fine Tuning Public Models – adjusting the probabilities within the libraries of public AI models to better suit your specific needs.
4. Fine Tuning Private Models – customizing private AI models to sound more like your unique voice or style.

Fine-tuning public and private models involves changing the probabilities within the libraries of the models – think of a large language model as a pizza with various toppings. What’s shipped out of the box is a pizza with a specific combination of toppings. Fine tuning allows you to adjust the toppings to your preference, essentially putting your thumb on the scale of the language you want the model to sound like. For example, if you want the model to sound more like your blog, you can feed it a dataset of your blog posts and rewrite the weights in the model to capture your tone of voice and unique style. 

Determining which stage you should be at within the four stages depends on your use case and requirements – factors such as data security, compliance, legal, and regulatory requirements all play a role in deciding which tools and approaches to use. For instance, if you work in healthcare, you should never put other people’s healthcare data into any public model. Instead, you should use a local AI model that runs on your computer and keeps the data secure.

Why is prompt design important? What can you do to create/engineer better prompts? 

Prompting is programming – when writing prompts, remember that they are essentially programs or software. They require specific instructions to produce the desired output. A prompt is a creative brief where you can’t revise or talk to the designer after writing it. Be as specific as possible to ensure the desired outcome.

Consider the expectations and context windows of your model – each model has a context window, which is the amount of data it can ingest. For example, the basic LLaMa model has a context window of 2048 tokens (~1,400 words), while the GPT 3.5 Turbo model has a context window of 16,000 tokens (10,000-11,000 words). To create better prompts, figure out what the model is expecting. For example, GPT 3.5 Turbo looks for a system prompt, a role prompt, and an instruction, while LLaMa looks for instruction and response, in Vicuna and Alpaca, the model is looking for background, instruction and response.

Tailor your prompt to the specific system and get detailed – to get the maximum performance, understand the underlying system architecture and tailor your prompt accordingly. The more detailed a prompt is with relevant keywords, the better it will generally perform.

See the TrustInsights guide on effective CHATGPT prompts here.

What are useful prompt management tactics? What tools might help you manage prompts?

Establish a knowledge base and program for managing prompts  – just like using a git repository for code, create a centralized location for storing and managing prompts within your organization. Create a program (rules, governance, etc.) for managing prompts. Implement governance and rules and set up a formal process for managing prompts, including guidelines, rules, and best practices.

Adapt to existing internal processes and use your tools – align your prompt management strategy with your organization’s current processes and tools, such as SharePoint or other content management systems.

Integrating AI into Marketing Function

What vendors or third parties can help you integrate AI into your marketing function? How do you evaluate these vendors?

First, gather requirements and determine the specific needs of your AI integration – this will help you find vendors that fit those needs and align with your budget. You should also consider whether you should rent AI services or build your own, depending on the importance of the function to your business. If the function is integral to your business, it is better to own it rather than rely on a third party.

Third-party advisors can be valuable in helping you evaluate and customize AI tooling for your marketing function – these advisors can assist with tasks such as evaluating vendors, and potential employees, and transitioning to new tools like Google Analytics. They can also help you get more out of your data, by building custom tools or fine-tuning models for marketers if needed.

To assess a vendor’s credibility, request to speak with an engineer unaccompanied by a salesperson – this will allow you to ask technical questions and determine if the vendor is knowledgeable and trustworthy. Be wary of vendors that refuse this request, as they may have something to hide.

When should you use a general LLM, and when should you look to train the model on your own data?

You want to own and train your own model if the task is integral to your business – decide whether to rent AI services or build your own, depending on the importance of the function to your business.

Training your own model can provide cost-effectiveness  – using an API for AI services can be expensive, especially when your app has a large user base making numerous API calls. Owning your own model allows you to pay only for compute power, reducing costs in the long run.

Owning your model provides you more customization and control over usage – when AI is integral to your company, having a customized model ensures that it is tailored to your specific needs and requirements (e.g. if you have sensitive data that you need to protect). Owning your own model also gives you more control over how it is used and deployed, such as running it locally on a device or server.

Which of your assets might you train a language model on? What model should you use to train your own AI? 

You can fine-tune with public and open-source models including: 

• GPT-3 (DaVinci) – the most recent version of GPT-3 from OpenAI that can be fine-tuned. It has the least amount of architecture overhead since OpenAI hosts it. You only need to do the fine-tuning and pay the bill.
• LLaMa (for research) – although not commercially licensable, LLaMa version 1 can be used for research purposes. It was trained at a cost of about $100 million and took Facebook around nine months with a server farm of 100,000 GPUs to process.
• Mosaic ML model – a commercially licensable model that can be fine-tuned for various tasks.
• EleutherAI GPT-J-6B – another commercially licensable model, the GPT-6J from Eleuther.AI can be fine-tuned for specific applications.

Keep in mind that fine-tuning these models is different from training them from scratch – which can be prohibitively expensive and time-consuming. Fine-tuning involves adjusting the pre-trained models to better suit your specific needs and tasks.

How can you leverage AI to facilitate a move to marketing mix modeling?

Marketing mix modeling and media mix modeling are types of regression-based AI that help direct the resources of the marketing function – these models analyze various data points to determine which variables, alone or in combination, have the highest statistical relationship with the desired outcome. They rely on regression- algorithms like LASSO regression, ridge regression, and gradient boosting. Tools like IBM Watson Studio, R Studio or Python can help you find which channels and activities have the most valuable statistical outcome.

Security, Privacy, and Human Involvement

What security and privacy considerations should be taken into account when leveraging AI in marketing?

Results of AI are not copyrightable – the US Copyright Office does not recognize copyright production for works created by non-humans, including machines.

Anything you put into a public model might not be secure – for instance, if you work in healthcare, you should never put other people’s healthcare data into any public model. 

Uncensored models have more linguistic diversity but create more risk – if you remove words or concepts out of safety, it will damage the probabilities of the words around it. An uncensored model is more creative and has more linguistic diversity, but it also has more problematic text and more bias.

What are the most important things to get right?

Determine the purpose of your use of AI – before leveraging AI in marketing, it’s crucial to understand why you’re doing it and what use cases you need the tool for. Implementing AI just to follow trends or to be one of the “cool kids” is not a valid reason. 

Evaluate whether or not AI is the right solution for your problem – not all challenges or tasks are suitable for AI. It’s essential to assess whether any of the AI solutions available are suitable for your specific needs. There are situations where AI might not be the best choice, and other methods should be considered. Just because you can use AI, doesn’t mean you should.

What are common pitfalls?

Believing AI is magic – AI is not magic, it’s mathematics. It is not going to cause the end of employment or bring about a new renaissance. It will change the marketing landscape, but it’s essential to understand that it’s a tool, not a solution to all problems. It’s important to guard against this assumption with stakeholders and emphasize that AI is an accelerant, not an instant, fast, and free solution.