Using AI in PPC Responsibly (Mitigating Risks)
When I started experimenting with AI for analysis, research, creativity and admin, I was really excited! I thought it was so cool to have tools do what I used to have to spend hours doing (e.g., pulling data from different sources, stitching it together, and calculating ratio metrics + PoP % differences). My stomach soon dropped when I realised that by using AI for this, I was opening myself up to risk. Risk that I’d have to manage well to prevent damaging client trust & torpedoing performance.
I thought about the threats AI could introduce & settled on 4, though there are more that I’m not covering.
4 Key Risks:
Hallucinations
Misuse of sensitive data
Decreased ability to learn
Faulty decision-making (due to AI lacking business & account context)
Hallucinations
We know AI is capable of hallucinating. It can confidently answer a question with the incorrect answer & if you don’t know enough about the subject & don’t use critical-thinking, you may just believe it and run with it. I’ve had experience with AI confidently hallucinating PoP figures that I needed for a report, claiming revenue was slightly higher than it actually was. Had I taken that and input it in my client-facing report, we would have had some not so easy to answer questions from the client (bc saying that AI hallucinated isn’t really a valid reason is it?).
It was tempting to input the numbers and just take what AI was saying at face-value, especially as the other 10 times I’d used it it’d been 100% accurate. But, AI having been good historically does not mean it has no risk of being inaccurate in the future. Hence the need to QA the output every time. It only takes one hallucination to fracture your credibility.
Misuse of Sensitive Data
We’ve got droves of data in our accounts and it’s a PPC’ers dream (at least, it’s mine) to pull it, synthesise it, and glean insights. However, if we load the data into the wrong platform (e.g., unapproved use of Gemini for the web), we risk damaging client trust & that data being used for model training. I don’t know enough about model training, so won’t go in depth on that. But, I do know that at some agencies, clients reserve the right to ask about the way their data has been used and which AI platforms it’s been loaded into.
As such, there is a risk that if you use the data in a way a client isn’t comfortable with, you will damage their trust in you and your agency. Working in PPC is as much about winning the client & driving results as it is about retaining the client. Protecting their data & preventing it from landing in the lap of unauthorised parties is part of this.
Decreased Ability to Learn
This is one I hadn’t initially thought of, just because AI can act as a learning accelerant. Got a question? Pop it in Gemini/ChatGPT (whatever you like) and you’ll get an in-depth answer. You can even ask follow-ups. However, the danger here is that you take the information AI gives at face value, without developing practical experience.
Take bid strategies as an example. If an account executive had sat in ChatGPT, asking it to explain how a target ROAS bid strategy worked, the executive would develop a top-level understanding (bids get adjusted up or down, based on estimated value), but not understand how this looks in practice. Of course, they could tell you how a bid strategy works, but they might lack the practical experience that cements that understanding.
For example, I had a high-level conceptual understanding of tROAS through asking questions & reading articles. But I always found myself forgetting exactly how tROAS worked. It didn’t feel intuitive. My knowledge only crystallised when I began pulling keyword data for my accounts & calculating the % of spend being funnelled through keywords without conversion value vs. with.
The data showed me that about 90% or so of spend was going towards keywords with conversion value, so worrying about the bid strategy spending on terms without value was largely redundant. This helped me understand bid strategies in a way that reading text from a prediction machine never could.
Faulty Decision-Making
Finally, I thought a lot about what happens when we make decisions based on AI. Particularly, when it lacks business context. I’ve worked for clients with unconventional setups that couldn’t be changed because of the way the business is structured. I’m talking bid strategies being used in ways that differ from platform best practices. The risk with sharing data about account setups with AI and asking it to make suggestions is that it doesn’t know why your account is set up unconventionally.
It’s even riskier if someone without deep business & account context starts leaning on recommendations from AI. For example, I recently did an account restructure & had to put some keywords on a Target Impression Share. From my business knowledge (emphasis on investing budget where the return/payoff is best) & account knowledge (phrase + t.IS tends to be more expensive & pulls in too many non-converting terms), I’m able to understand that phrase match & T.IS bid strategies aren’t the best pairing.
But when I ask Gemini (screenshot below), it positions phrase + t.IS as a good option. It does explain that there are risks to consider (e.g., loose matching, stronger need to review search terms), but the issue is that the first piece of information might taken as gospel and the caveats will get glossed over.
So, How Do We Manage Risk?
I’ve distilled it down to 2 key questions we can ask ourselves before trusting AI & its outputs:
What’s the cost if the output is wrong & how consequential is it? A hallucination of client figures during a Quarterly Business Review is far more consequential than a hallucination of search terms in a largely clean search term report.
Will offloading prevent deeper learning? Ask yourself whether the answer you’re seeking is something you’ll need to know like the back of your hand, without AI being there to feed the answer to you. Ask yourself, if you needed to recite the definition, or a client asked you how a feature works, would you be able to answer using intuition & experiences or just technical information you’ve pulled from AI.
Wrapping Up
Like all things, AI has both risks and benefits. This article isn’t intended to encourage you to be all doom and gloom. More so to get you thinking about what could go wrong & how to ask questions before they do. Get to know AI & how it works, its failure points, and the impact this has on outcomes (e.g., performance, client relationships, learning) then evaluate from there. You may decide you need more guardrails. Or that some of the risks are worth taking. Either way, you’ll be a little more informed.
Comments