AI in Advertising Part 1 of 3: How to transform agency workflows with AI
This month, in Part 1 of a jam-packed AI in Advertising VML strategic director Antonio Petra gives a practical guide to using AI to improve how your agency operates with your clients and internally.
Then look out for Part 2: Vincent Maher's seven life-changing AI tools that you should be checking out.
We round up with Part 3: Key AI in Advertising news and highlights from the past few weeks.
Transform agency workflows with AI
I recently posted on LinkedIn about how AI's most significant impact on advertising will enhance operational clarity and efficiency.
One of the comments I got back was, “I think everyone understands there are benefits, but in my experience, what people are struggling with is the how".
Sometimes, in the fog of putting the I into LinkedIn, we lose sight of reality.
In this article, I will explain a way to approach probably the costliest pain point in agency communications:
The intersection of communication-related to initiating or enhancing our product – Briefing and Reverts.
Without clarity of thought and specificity in these two areas, work will be more expensive and cause more incredible frustration for both agency and client, with the ultimate cost being the relationship.
This is not an exhaustive instruction manual, and for learning, I encourage you to lean into any issues you might have with your prompt as an opportunity to learn.
I also intentionally focus on this one area in detail with practical tips to inspire you to explore other places where you can explore your AI solutions.
I am also steering away from jargon and use the terms LLM and GPT to describe tools like ChatGPT, Claude, Gemini, and Microsoft Co-Pilot.
These instructions can be easily reconfigured for the following kinds of “co-pilots”:
- Brief enhancement – Help me write a better brief.
- Brief Analysis and Conversion – Is my brief precise, and can you adapt it into (i.e. a creative brief)?
- Brief completion – Help me finish this brief.
- Revert consolidation – Can you consolidate feedback from multiple people into a single checklist?
- Revert clarification – Does this revert make sense, and are the instructions specific and clear?
Getting started
The first thing you need to do is pick your AI tool. Your first prize is a company-sanctioned tool that promises data security. Your other option is a mature LLM with file upload ability and access to the Internet.
Your primary concern is that the data you upload won’t appear in response to another chat. The information in the brief on the new product launch won’t appear on a list of products in the wild when another user requests that information.
In my view, every agency should already have given all their employees access to a paid platform with higher usage limits and security. This is essential for evolving your workforce for AI, and it’s the right thing to do from a data security point of view.
You have few options if you do not have access to a company-approved paid solution.
For many publicly available services, data is only private when you have a paid account (i.e. OpenAI). Services like Claude promise data security, but this has some limitations. Your options for data security are either a paid service or Microsoft Co-Pilot, or you can roll your own AI (Llama).
You can also follow this guide, which shows how to opt-out of certain AI tools using your data for training.
Now that you have a tool, it’s time to fire it up.
You can start by telling the GPT precisely what you are doing, and it’s good practice to keep informing it so that it can give you feedback.
Example first step: “I am setting you up as a GPT to help me analyse the brief for completeness and provide recommendations on completing gaps. I am going to start by including some background information and rules. Please ingest this content and use it throughout any conversation I have with you”.
Setting Up
You may need to enter a reference depending on how you want your output to read or look. A reference could be a template you use or a document that defines excellent output, like this document from the IPA on How to Brief.
It’s also helpful to input some operating rules before you get started. These rules are important as they save you time later in correcting and checking.
Here are some example rules (some with explanations):
- Please always use British English spelling.
- Don’t confidently guess if you don’t know the answer. Say, “I don’t know” or “This is an approximation.” – Note: This prompt helps prevent the system from making up an answer (daydreaming).
- Cite your sources if you’re providing research.
- Feel free to contradict me and point out where my logic or thinking is flawed or factually incorrect. Note: AI can sometimes try too hard to please you, giving it a bit of an edge and telling you when you are wrong.
What you also want to think about is whether you are just going to use this AI once off or if you’re going to come back with other briefs.
If you are going to do the latter, you need to give it kind of a reset instruction; for example, I am going to be using you to refine multiple briefs, so anytime I come into the conversation and say, “New Brief”, it means that we are starting the process of analysing the brief again or analysing a new brief”.
Writing your prompt
Now it’s time to write your prompt, which is a good prompt for a complex operation and will have a few components.
CO-STAR is one way of structuring a prompt, and to be clear, I tend to mess with it a lot, but it is helpful to describe the facets of what a prompt needs to do:
- Context or Background: Tell the AI what it is and be specific. This helps it provide more contextually relevant answers. For example, “You are a highly experienced marketing strategist in the FMCG area.”
- Objective: What objective do we want to achieve? For example, “Your task is to assess briefs for clarity and completeness. You will point out where you think there are gaps in a brief and propose amendments or ask questions to help fill these gaps.”
- Style: What response style should it use? For example, “Respond in plain English, without jargon, focusing on brevity and clarity”.
- Tone: (Optional) What should its response attitude be, for example: “Respond like Steve Jobs” or “Respond in a professional but warm business manner”.
- Audience: (Optional) Who is the desired audience? For example: Creative – “make it inspirational”, Client – “keep it professional”, CFO – “keep it silent”).
- Response Format: This is where you will ask the AI to apply templates or specific formats you would have given in the setup (Brief Template, Best Practice Guide, etc).
Pro-tip: If this is overwhelming and you don’t know where to start, fire up a second GPT, tell it you are trying to write a prompt to do X and ask it to help you write it.
Fine tuning
By now, you have a working co-pilot. It has background documents; it knows how it should behave and what to do.
It's time to start feeding it some tests.
There are two things to consider in this process:
- Prompt writing is descriptive coding or programming, and like any programming, the GIGO (Garbage In, Garbage Out) rule applies; a poorly written prompt will output less helpful content.
- Your prompt most likely will require re-writes to parts of it; I recommend keeping a copy in a text editor and copying it across, changing only one thing at a time. This gives you a backup of your prompt that you can undo if you break something.
Keep on testing and iterating until your output is exactly what you need. Don’t be discouraged if the first output is not what you were expecting.
There is a study that shows that humans have an inherent bias toward AI in that if the initial result from a newly tested AI is not up to expectation, they will abandon that AI.
Don’t fall for this bias; persevere until you get an acceptable output.
Assessing your output
No AI model is perfect, and assessing your prompt will always be a balance between several factors. However, by now, you should have something that gives you a reliable output, your tests have worked for an ideal situation, and you’re confident in the results.
This does not mean you can go into autopilot and cut and paste for the rest of your day. You need to check the output for two reasons:
- LLMs can drift over time, their algorithms can be rewritten, and the output can change. So, your output is not guaranteed to be perfect every time, and that one time might end your career.
- If all you’re doing is cutting and pasting, you should start polishing your CV because it means an AI can do what you do for free. You need to add your human experience and creativity to the output to elevate it beyond an excellent output and make it a great production.
Sorry, that’s a brutal ending, but it is the truth.
The power of AI is human + AI. Studies have shown that AI is much more effective as a collaboration partner.
Future-proofing your career depends on how you develop your skills in enabling AI as your collaboration partner.
About the ACA
We are the official industry body for advertising agencies and professionals in South Africa, counting most major agencies among our members.
Find out more about the ACA.
*ACA Future Industry committee comprises Jarred Cinman, Vincent Maher, Musa Kalenga, Haydn Townsend, Matthew Arnold and Antonio Petra.
