In this episode, Katie and Chris discuss the different levels of complexity when it comes to AI and machine learning in marketing. What’s the difference between AI and data science, and when should you focus on one or the other? What cautionary tales should marketers understand before positioning a product as powered by AI? Tune in to find out!
Subscribe To This Show!
If you're not already subscribed to In-Ear Insights, get set up now!
- In-Ear Insights on Apple Podcasts
- In-Ear Insights on Google Podcasts
- In-Ear Insights on all other podcasting software
Advertisement: Data Science 101 for Marketers
Do you want to understand data science better as a marketer? Would you like to learn whether it’s the right choice for your career? Do you need to know how to manage data science employees and vendors? Take the Data Science 101 workshop from Trust Insights.
In this 90-minute on-demand workshop, learn what data science is, why it matters to marketers, and how to embark on your marketing data science journey. You’ll learn:
- How to build a KPI map
- How to analyze and explore Google Analytics data
- How to construct a valid hypothesis
- Basics of centrality, distribution, regression, and clustering
- Essential soft skills
- How to hire data science professionals or agencies
The course comes with the video, audio recording, PDF of the slides, automated transcript, example KPI map, and sample workbook with data.
Sponsor This Show!Are you struggling to reach the right audiences? Trust Insights offers sponsorships in our newsletters, podcasts, and media properties to help your brand be seen and heard by the right people. Our media properties reach almost 100,000 people every week, from the In Ear Insights podcast to the Almost Timely and In the Headlights newsletters. Reach out to us today to learn more.
Watch the video here:
Can’t see anything? Watch it on YouTube here.
Listen to the audio here:
- Need help with your company’s data and analytics? Let us know!
- Join our free Slack group for marketers interested in analytics!
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.
Christopher Penn 0:02
This is In-Ear Insights, the Trust Insights podcast.
In this week’s in your insights, we are gearing up for conference season.
We just had a last week a chat with Kathy mcphillips from the marketing AI Institute.
And one of the questions from sort of the q&a period I thought was very interesting was, what are some of the more high level or higher complexity implementations of AI? You’ve seen us in marketing that from Dr.
So Katie, in terms of this is a topic what comes to mind for you in terms of when when someone asks, show me something like really complex activity all the time?
Katie Robbert 0:42
And obviously, I’m not the person to be showing off the complex things.
So that’s a great question.
You know, it’s, it’s funny, it strikes me that a lot of the stuff we do, which to us seems really straightforward, is actually complex in comparison to what kinds of tools marketers can typically get off the shelf or have access to.
So I think our attribution analysis continues to get more complex, especially as we refine the process of bringing in multiple datasets and making it a more advanced model for online and offline channels.
So I think that that’s definitely something I know that natural language processing is definitely moving right along in terms of its ability to write full bodies of content.
And so that I would say is probably an advanced use case, it doesn’t sound like it.
But the methodology to get from A to B to train, the model on how you want things written is probably more work than somebody might realize.
Um, so I think those are the two that stick out to me.
And obviously, you know, I’m, you know, doing analysis on large quantities of data, something that would take someone a very long time to do is something that, Chris, you’ve figured out how to streamline down to mere minutes or hours, depending on how large the data set is.
Christopher Penn 2:15
It’s interesting, when I think about that question, there’s actually a slide that in my mind, MAE Khan talk, it looks like this.
And this is a sheet of paper, and two different ways to put a hole in that sheet of paper.
One is clearly not the right choice.
And the other is the, the normal, sensible choice.
And I bring this up, because I think it’s an interesting question.
But I also think that there’s a lot to be said, for knowing what level of complexity you actually need.
Not just, I’ll be the first to raise my hand, say, like, I love messing around with cool stuff and fun stuff and things right.
But I don’t have a lot of use cases is overkill, right? Like, do you need a neural network to do basic sentiment analysis? Not Not really.
You know, when we look at something like analyzing Twitter BIOS, do you need a sophisticated neural network to do that? No, no, you just counted words, it’s not super complex, I think one of the challenges that we run into, and I know other people run into is, especially that benders, getting, getting getting a solution that is overkill, or not a good fit word made needlessly complex, I guess, is a good way of putting it.
It’s just you’re just trying too hard.
Katie Robbert 3:41
Well, it’s funny, we literally maybe 30 minutes ago had this exact conversation about the approach to analyzing some data.
So for one, on behalf of one of our clients, we’re putting together a feedback survey.
And as the survey has evolved, the questions, the set of questions has become shorter and more focused.
So when we originally put it together, we imagined, you know, 4050 questions.
And we would have to write code in order to analyze a lot of text to pull out themes and, you know, sentiment, common words, those types of things.
And now as we’re getting closer to finishing the survey and launching it, we’re realizing that well, if it’s only about 15 questions, and we’re only going to have five to 10 respondents on this thing, we probably don’t need any code at all, we just need a well organized Excel spreadsheet that we can look at and pull out data points.
And so it’s, you’re absolutely right, Chris, I think understanding the purpose, the question that you’re answering can help you determine how complex a solution you need.
And I think that I mean, we’ve been talking about this, you know, six ways to Sunday solutions in search of problems and complex AI and you know, this, that the other and so, it all boils down to the same thing of what’s the question you’re trying to answer?
Christopher Penn 5:06
And when you’re working with AI technologies, in particular, one of the things you have to keep in mind is compute cost.
Right? The more complex you get, the greater the computational burden.
There are some things, you know, when we look at the attribution analysis we do if we were to choose a more complex out with them, what would was typically like a five to seven minute process could be a five to seven hour process, which is fine if it’s just you, but if you’ve got 10 clients, you know, certainly talking 50 to 70 hours of compute time, versus an hour of compute time.
And I think when we’re doing requirements gathering, I don’t think that’s taken into account enough is to say, we have to balance the outcome with the computational cost, the expense, I was trying to put something together this morning and Watson Studio, it’s like, you know, this operation will cost you 20 hours a few time, like, cancel it.
Don’t Don’t do that.
Like, I don’t know what it was.
That was it was saying like, this is a really expensive operation.
But it got me think like, that’s, that’s overkill?
Katie Robbert 6:18
Well, it, it goes back to sort of, you know, getting those business requirements and playing out those scenarios.
So, you know, a lot of times we talk about scenario planning in terms of what happens if you don’t do the thing versus what happens if you do the thing.
In this instance, the scenarios that you’re really describing, Chris, are, you know, what if I want the Toyota tercel? Is that going to get me from A to B? Yeah, absolutely.
What if I want the, you know, Nissan, will that get me to a to b? Yep.
What if I want the, you know, whatever the test latest Tesla is, well, that might get me from A to B, absolutely.
They all get you from A to B, it’s just a matter of how much you’re willing to invest, and how fast you want to get there.
And so thinking about it in that context, you might say, so Chris, you know, you said, Okay, 20 hours of view time, I don’t need that.
So maybe the Toyota tercel is good enough to get you from A to B, and you can continue to, you know, change the tires, and make sure that your engine fluid is, you know, always fresher, whatever cars do, I don’t know, I’m not a mechanic, this probably a terrible example.
So but make sure your brakes are working correctly.
Um, you know, maybe you spring for an automatic car starter to make it a little bit more like up to date.
So there’s things that you can do to build on over time, instead of starting with the Tesla.
Christopher Penn 7:47
Yeah, and now, going back to Ashley’s question about what are some of the more sophisticated things, there are some things where the the models being the pieces of software being used, do you have to have a certain level of sophistication, like if you’re doing natural language generation, there isn’t a cheaper, easier shortcut for a lot of that stuff, there really is, you’ve got to use the big heavy compute power, because it is it has the highest level of accuracy.
So I guess that’s your third trade off is you have your requirements, you have the computational costs, and they have the quality of the outcome.
And you got to kind of push and pull and say, Okay, well, if we really want a quality outcome to be high, it’s going to push the price up.
And it’s going to make the business requirements a lot more challenging, because you have to be very, very clear about what it is that you want this thing to do.
LinkedIn discovered that when they started building with, you know, deep learning models for for their analysis of what members were posting, like, yeah, you need to invest a lot of money and a lot of compute time to make this work.
But when you look at those outcomes, you’re like, Wow, those are really, really good outcomes, like when these language models start writing, like entire press releases all by themselves, like, wow, that’s slightly alarming.
Katie Robbert 9:08
So what’s interesting, Chris, is what you’re describing is the three constraints of project management.
This is a very old theory.
And so basically, your quality is determined by your time, your scope and your budget.
And it’s not that you can’t have all three, but each of those balances out in terms of the quality of the output.
So if you want it, I think you have a different version of this in your talks.
Like if you want it fast and cheap, then you can’t have this or if you want it like you know so it’s it’s along those lines.
So that’s essentially what you’re describing.
Is of these three things you can pick to.
So which of these two things do you want? Do you want it fast and cheap? Do you want it expensive and high quality like what what is the most important to you? terms of the output.
Christopher Penn 10:02
Right? Yeah, it’s fast, cheap, good.
And the unless you’re Google, and even in those constraints to apply to Google, you really can get a situation where you can have all three.
Right? You have to trade something off somewhere.
Right? And I think a lot of folks are under the impression that AI and all these technologies somehow nullify that, that trade off, I guess you can have your cake and eat it too.
Well, no, you can’t.
Katie Robbert 10:35
So what is it about AI? And this, this might be a rat hole question, but what is it about AI? that people think is, is going to suddenly, you know, solve all of those problems? Is it because I, you know, I think about this when I hear these questions come up, and maybe maybe I’m asking the wrong question.
This is, this is me, my brain processing information, live on the podcast, you know, so I, as a CMO, I want to introduce AI, in order to cut down the time it takes to write content, for example.
So I introduced AI, natural language processing, and it writes content.
Well, guess what, someone still has to edit that content, someone still has to post that content, you know, someone still has to tell the AI what to write about.
So there’s still a planning aspect of it.
So did the AI save you any time? Or is it more of a headache to get everything set up and working the way that you want it to? It’s never perfect.
Christopher Penn 11:38
I think it’s three things.
It’s understanding culture and expectations.
So understanding is the first thing people don’t understand the technology, they don’t understand what it can and can’t do.
They kind of assume its magic.
And therefore, because they don’t understand the underpinnings of it, they don’t know what its limitations are.
Second is cultural, you know, everything from Wally and short circuit to Terminator movies and things have created this perception that machines are a infallible and be are much more capable than they actually are.
And third, and I think this is where we see the problem in business is level of expectations.
Everyone who asks questions like that, is looking for a shortcut, they’re looking for disproportionate returns on their investment, it’s the same thing we see in public relations.
You know, when when that was sort of the pitch of a PR firm was you can get disproportionate results for the dollars you invest, we would always say pitches, advertising is linear, you put $1 in you get this right, and you put another dollar in, you get this plus this.
And it scales, so you keep putting more dollars in and keep getting this mission occur.
And the promise of PR, which almost never worked out was that you put $1 in you get $10 out or you know, you get this big story in the Wall Street Journal, and suddenly the world would be the path to your door.
And that was the promise that the PR folks would pitch for years.
And I think the AI thing does the same thing.
It’s like, hey, if you use this machine, you’re going to see disproportion results like winning the lottery.
magic will happen, and everything will be better.
Certainly, that’s what the vendors say.
Katie Robbert 13:13
Well, so you know, that brings up a different aspects of AI is it’s only as good as the data that you feed it.
And so if you knew how to pick the winning lottery numbers every single time you wouldn’t need AI.
So feeding, losing lottery numbers into an AI isn’t going to suddenly predict winning lottery numbers.
And so if we keep going with this example of writing content, that was you said, sort of what I was thinking was, are people just looking for shortcuts because they don’t want to write content or you know, and so if you don’t want to write content, then the content that you’re currently creating is probably not great.
So you put your content into this AI to write your content for you.
Guess what, you’re still gonna have not great content because the AI can only write what you give it
Christopher Penn 14:04
to that’s where I think there is some potential for a slightly just proportion results because if you hate writing content and you’re bad at it, you just crank out crap.
And I use a a naive pre trained model like GPT j for example, and you have it right into the Create mediocre content.
It’s not gonna create crap like you face rolling against your keyboard is going to create garbage it will create you know, like a white bread and mayo sandwich no one you won’t starve to death but no one’s asked me for that.
You know, shape a nice.
Katie Robbert 14:36
I like those by the way.
So I would ask for that.
Christopher Penn 14:42
Artists, it’ll have bad days.
Katie Robbert 14:43
Oh, I’m sorry, the artists in the one feels.
Christopher Penn 14:47
What? If your contents crap and the AI content is mediocre? That’s already an improvement, right? It’s an improvement over the garbage you’re putting out and in all honesty, When you get a lot of the marketing communications we get from other companies it is garbage.
So AI producing mediocre would actually be a disproportionate return on investment because you’re no longer handing out pure garbage now, is that good enough to stand out? No.
Is it good enough? You know, it’s like AI creating music, I use the eight, Eva tool for creating the music for the pre roll for the talk I did with Kathy, I use AI for its intended purpose.
Was it great music? Was it gonna win a Grammy? No.
Is it good enough? for like a one time broadcast? Sure.
It was it was inoffensive?
Katie Robbert 15:39
Well, and I guess that sort of back to where we started was, you know, if you think about the different use cases, or the different versions, how complex to how simple, you know, I think it’s always probably a good option to start with, is it good enough? Are you going to get good enough data to do something with because you can always improve upon it, I think the trap that a lot of us fall into, and we’ve we’ve been guilty of it.
So we’re not pointing fingers is we try to perfect something.
So we spend much more time much more energy, much more budget and resources trying to perfect something before just Okay, now we just need to start using it.
And I think the same is true of AI.
You know, we try to make it perfect and do exactly every little bit piece that we want it to do needs to write our content and fold our laundry and make our coffee, and balance our checkbook and hire people like you could maybe just start with one of those things.
Christopher Penn 16:46
And the other thing that computation I think is very interesting.
And I forget where I heard this is that AI is bad at what we’re good at.
And AI is good at what we’re bad at.
Machine tasks like computer vision, natural language processing suffer from very intense compute intensive processes that are very difficult for machines to learn.
On the other hand, doing something like glue, logistic regression, super easy and fast for a machine, but you and I are like, I can’t do that by hand.
I mean, it’s, it’s, it’s been a while.
You know, you look at the formula for logistic regression.
You’re like, what goes where are you? But the machine is like, Okay, cool.
You need a million record data set, you know, logistic regression across it here, done.
And so from, you know, to Ashley’s question about, you know, where what are the higher level applications, in some ways you don’t actually, you may not actually need them for some things you kind of do.
And, to your point, there may be some easy wins in the tasks that we’re not good at, that the machines could crunch, you know, instantly.
And to say it was a whole bunch of time.
Katie Robbert 17:56
So what, you know, what strikes me is one of the misunderstandings of how AI works, I think there’s this, you know, idea that AI produces things instantly.
And that’s where the time savings is gonna happen.
Like, I press a button, and the AI automatically shoots something out at you, and you have the thing.
And you know, what you’re describing Chris is, you know, 510 20 hours of computation time.
And so is it machine learning? Yes, it absolutely is.
But it’s just running in the background, it’s not doing anything exciting or interesting.
You’re just watching your terminal screen, just like turn three line after line after line.
But that’s, that’s what it is.
That’s how it works.
And it’s not, depending on what it is, it may not be faster, it may be faster than someone trying to hand compute all of that stuff.
But it still takes time to run all of all of that information through to create what may be a single line output.
And I think that that’s a misunderstanding, in terms of, you know, how it works, but also when you factor in the complexity.
And so obviously, the more data that you have, the more things you want it to do, you know, go down the list, the longer the AI is going to take to give you back a result.
It could be, you know, three days.
Christopher Penn 19:20
It’s interesting to see that too, because I think the the AI and machine learning community have not done a real good job of explaining, like the difference between data science and AI.
And I think that distinction is important for marketers, because many times you’re probably asking for something, that’s the wrong thing.
So the fundamental output of data science is an answer.
I have a question.
Here’s an answer.
The fundamental output of AI is a piece of software, hey, I need a machine to do this thing.
Machine creates its own software.
If you’re looking for an answer, you’re not looking for AI you’re looking for data science.
You’re looking for it like it.
You know what channels worked best.
You may use machine learning techniques, you know, like Markov chains for that, or for example, but fundamentally, you’re not getting a piece of software you put into production, you’re getting an answer to your question like, oh, let’s spend less money on Facebook.
Whereas if you are looking for a piece of software to do the thing a lot, then you’re, then you’re looking for AI.
And I think that distinction is kind of lost on people.
Katie Robbert 20:28
I would agree with that I’ve heard, you know, machine learning, data, science, AI, all of those things used interchangeably.
And I know when we talk about artificial intelligence, we talk about the hierarchy of statistics, machine learning, deep learning those things and how they all are different layers of artificial intelligence.
And so you might just be doing statistics, you might just be using a piece of code to run a linear regression model.
Is that machine learning? Yes.
Is it artificial intelligence? sort of not really, you’re really just expediting the mathematical equation, you know, the, it’s not giving you any information outside of what you’ve asked it to do.
It’s not learning anything.
It’s literally just computing quickly.
And so I think you’re absolutely okay.
And that might be okay.
You know, it’s funny, I look at this image.
If you’re listening to us, then you can see the image up on our YouTube channel for this.
And we can post this image along with our social posts of this podcast.
But basically, what I see is the Chris Penn tool, and the Katie robear.
So if you asked, and I mean that lightheartedly, like if you ask both of us to make a hole in a piece of paper, I’m going to grab the hole punch, and Chris is going to grab the power tool.
And I think that it speaks volumes to our approach of how to solve the same problem.
And that’s, you know, something that you can look at within your own organization, with your own teams.
If you ask everybody the same question, how do they approach solving the problem? And that will give you a good understanding of, you know, how complex the solution might need to be?
Christopher Penn 22:20
And and one of the things that I think really, really emphasizes that is, we’ve seen it in Google’s handbook for developers around machine learning.
The first rule in their machine learning handbook is if you don’t have to use machine learning, don’t use it, because it’s incredibly compute expensive.
And you know, for that to be the first rule in the machine learning handbook.
It’s like, Oh, that’s interesting, instead of you the answer to expect is make everything a machine learning problem solve everything on machine learning.
Katie Robbert 22:50
Do you think that people try to use machine learning just to say that they’re using it to sound advanced to, you know, give this perception of things that aren’t actually true?
Christopher Penn 23:04
Without a doubt, in 2018, the Financial Times investigated 100 companies that said, the US machine learning in 35 of them were lying, just not a scrap of machine learning anywhere they were lying, just for marketing purposes.
say that, you know, they were, they were trending Oh, they were one of the cool kids do.
And they I, that’s very much the case where, because there is this perception.
Now we’re going back to our three factors, that machine learning can give you disproportionate returns, it then in vendors mind says that we can charge more for this, oh, if we’re using machine learning, I can, I can add a zero to the price tag.
And then on the other side of the table, the CMO is saying, oh, they’re using machine learning, I got 10x better results out of this.
So justifies increased price tag, and you know, the realest straw saying all of you are lying to yourselves and each other.
Katie Robbert 24:02
I find that psychology really fascinating, I can understand it from the wanting to, you know, create this perceived value of if you’re using AI, then it must be better.
Um, but the realist in me is like, well, who cares how it gets done, as long as the results are good, like, you should be signing up for the tool that gives you the best results based on what your needs are, be garlis of whether or not they use AI or not.
And I get that that’s a very, you know, difficult thing for people to stand behind, because there’s a lot of pressure to be doing the latest and greatest and staying ahead.
But quite honestly, sometimes, the latest technology isn’t going to give you the answer that you need.
Christopher Penn 24:46
Or it’s going to give you it in a way that’s not sensible.
We ran into this two weeks ago when I was cleaning up our Instagram influencer list.
I had a list of 14,000 Instagram accounts, and I have the URLs of those accounts and so much Choice Where could I write a scraper and image recognition? image recognition algorithm that look, do facial detection? Like, is there a face in this image? Yes or no.
And I think it was going to take me about 80 hours of work to code this thing to make it run.
Or I could set up a process on mechanical, Amazon Mechanical Turk, make a bounty of one penny, for every, every account, you you verified.
So you click on the link, go to the Instagram profile.
And then, you know, go back to Turk and say, is there a picture of a human in this profile? Yes or no.
It was done in an hour and a half.
From the time I lost it, and it costs $230 to screen all 14,000 accounts, the error rate was about 1%.
The machine error rate and the first time around is going to be around 70 75%.
So 2025 25 to 30%.
Wrong as the error rate.
So I had the choice of using artificial intelligence, which would cost 80 hours of time, no dollars, 80 hours of time, and a 30% error rate, or using human intelligence, or, you know, a couple 100 bucks, but a 1% error rate, and and done in an hour and a half set 80 hours.
Just because there’s an AI solution doesn’t mean it was the right choice in that example, was the wrong choice.
Katie Robbert 26:17
Yeah, I think that’s hard for people to come to that conclusion.
Because a lot of it is you don’t know what you don’t know.
And so they’ll reach for the AI solution.
Because they don’t know what the other option is.
They don’t know how to solve the problem themselves.
Otherwise, they may not know that something like Mechanical Turk exists to solve this problem.
So I think that a lot of it is around just education.
And so you know, that Google Developer Handbook, saying, if you don’t need to use machine learning, don’t, I think it’s a good place to start.
But what it doesn’t do is it’s not their job to do it.
But what it doesn’t do is educate like, here’s the other things you could be doing instead.
Christopher Penn 27:00
Exactly, here’s all the different options.
And you’re right, a lot of people may not know that tools like Mechanical Turk and Upwork.
And all these places exist that can use human intelligence, which is still for the most part, you know,
Katie Robbert 27:15
I could see how painful it is for you to say this.
Christopher Penn 27:18
It is because, you know, I look at some of the things people say in the news and like, wow, how did we ever survive as a species with people that dumb? But you know, that’s, that’s for another show?
Katie Robbert 27:28
Yeah, probably not the show.
Christopher Penn 27:34
So, but I think though, that raises an interesting point.
to round out to Ashley’s question about one of the most sophisticated forms of AI I actually think at the bottom of the top of the pyramid is human intelligence to be able to say like, yeah, there’s a bunch of things machines are good at, there’s a whole bunch of things machines are bad at.
And this really sophisticated marketer knows when to use which
Katie Robbert 27:58
I would agree with that.
Because I think once again, all the AI in the world is great, but at the end of the day, a human still needs to create it, build it, maintain it.
Christopher Penn 28:07
It’s it’s no different than your blender.
It’s just not as tasty when it’s done.
Katie Robbert 28:14
And on that note,
Christopher Penn 28:16
and on that note, if you’ve got comments or questions about anything in today’s episode, please hop on over to our free slack group go to Trust insights.ai slash analytics for marketers where you have over 1900 other folks can talk about analytics data science to your marketing challenge of the week or AI, to your heart’s content.
And wherever it is you’re watching or listening to a show.
If there’s a channel you prefer it to be on, go to Trust insights.ai slash ti podcast where you can see all the other options.
Thanks for tuning in, and we’ll talk to you next time.
Need help making your marketing platforms processes and people work smarter.
Visit Trust insights.ai today and learn how we can help you deliver more impact
Need help with your marketing data and analytics?
You might also enjoy:
Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, Data in the Headlights. Subscribe now for free; new issues every Wednesday!
Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new 10-minute or less episodes every week.