Insight ON From AI Fear to AI Fluency — How One Agency Won Over Their Creative Team

Audio transcript:

From AI Fear to AI Fluency — How One Agency Won Over Their Creative Team

Samuel Archibald (00:02):

They don't want to code themselves out of a dot. I think that's a real fear a lot of people have. I really kind of wanted to push on people. This is not the point of AI. Although a lot of companies are doing it, this is not the point of II. All it is is to take the boring parts of your work away from you. The parts that you're doing is taking an hour long that there's no creativity. It's either right or wrong. You're doing it the same thing over and over again. Those are the things that AI is brilliant at. If it has creativity involved, I don't want AI near it.

Jillian Viner (00:35):

The fear that AI is coming for someone's job is real and Samuel Archibald doesn't dismiss it, but he's also seeing what happens on the other side of it. When the repetitive time consuming parts of someone's work just gets handed off and then suddenly they have the headspace and the tools to explore crave directions that they never could have before. And it was never because they just got better at their craft or they didn't know what to do before. It was because the things that were limiting them weren't about craft at all. It was the noise, right? Welcome back to InsightOn. I'm Jillian Viner and my guest today is Samuel Archubald. He's the director of AI at the Sherlock Company, a creative agency working with entertainment studios across the industry. Samuel built his company's entire AI enablement program from scratch. The training, the platform, the assessments, simply because there was just nothing off the shelf that was going to speak to the specific fears and opportunities that his teammates were facing.

And really what he learned about how people actually change their relationships with this new technology is something that any leader trying to drive adoption really needs to hear. Here's Samuel. Samuel, welcome to the podcast. It's great to have you here.

Samuel (01:49):

Yeah, cheers.

Jillian (01:50):

You're part of the Sherlock Company. You're the head of AI over there. Sherlock Company is not what I think people think it is. It's not a detective.

Samuel (01:57):

So what do you do? So we're kind of a creative solutions agency. We work with a lot of the studios. We take the original English art. We translate it. We resize it for the 50 different platforms that exist now. We also do a little bit of creative work for some companies that are under NDA, but a lot of advertisements, little Instagram reels, stuff like that. Especially anything that needs loads of variations. We're really good at doing those really fast. And obviously five years ago it was all done manually.

Jillian (02:34):

Are you talking about thumbnails,

Samuel (02:35):

Like screenshots? Sunnail screenshots. The little icons, every platform has a slightly different way that they want the exact information. I'm

Jillian (02:44):

Thinking about all the dimensions.

Samuel (02:45):

If

Jillian (02:45):

You log into a streaming

Samuel (02:47):

Platform- You look at Netflix on iPads, different size on the phone, different sizes on the Apple TV. Someone actually has to pick PS5. Yeah. So they actually have to upload computer separate artworks. For each of those, they have to use the exact right pixels. And for the translations, they want the Japanese text to be styled exactly the same as the English was. It's not a font. We can't just type it. We have to recreate that by hand.

Jillian (03:12):

Okay.

Samuel (03:13):

And that's something that AI is still not quite there for at least the scale they need it to be at.

Jillian (03:20):

How long have you been the head of AI?

Samuel (03:22):

Oh, the head of AI, probably only about a year, but I've been kind of leader of AI for three because I've been the only one following it as closely as ...

Jillian (03:32):

What were you before the head of AI?

Samuel (03:33):

I was the developer and the designer. So I started off doing development and then I switched to UI/UX design. So a litle bit of front end dev, mostly working in Figma, building interfaces and how important the user flows are. And I really think that user flows are going to be the one thing differentiating companies because everyone can kind of create anything nowadays. It is really not who's first to market. It doesn't matter. Everyone can create, even if you have a tool that already exists, my first though is, can I create it myself? Not, should I search for the tool? So I think that the ease of use is going to be such a massive importance in how quickly people can pick things up because they can just build it themselves otherwise. So

Jillian (04:23):

You're clearly thinking about just speed to market, creativity, creative problem solving. So you took it upon yourself to really learn as AI evolved, created a use case or a business case for you to come into this role. This role is less than two years old. What has been on your shoulders? Because I think probably a lot of organizations are looking for someone internally to help lead it.

Samuel (04:48):

Yeah. So luckily I was already internal, otherwise we really would've had to hire someone external. Pretty much everything to do with AI is now on my shoulders. At the beginning of last year, it wasn't that much. It was just exploring and seeing what people can do, but now it's the security, the learning, the onboarding, all this kind of stuff.

Jillian (05:09):

Hello, Atlas.

Samuel (05:10):

Yeah. Don't

Jillian (05:11):

Shrug.

Samuel (05:11):

I'm having to write all these different programs. I've written about 20 different modules.

Jillian (05:17):

Like learning modules?

Samuel (05:19):

Yeah.

Jillian (05:19):

So all the change management and training you're doing.

Samuel (05:21):

Yes, exactly. Yeah. It's all custom. I'm not using any kind of other system, learning management system. It's all custom built using AI. So we use Replit mainly for that. And I use Gemini for the models and stuff like that. I can actually go in through some of the really cool things that we've done in grading the learning platform that are a little bit unique.

So I did a lot of research before building it and I found that retention rates and completion rates are terrible for these systems. Even if they're mandatory, around 60, 50% is what you usually get. So I wanted to change that. I wanted it to be fun. And so I have two different modes that you can choose and switch back and forth at any point. I have basic mode, which is like what you would expect, just the learning platforms. And I have gamified mode. I call it quest mode, which is basically like imagine Pokemon ... Pokemon Go? Pokemon Red from 2002. Oh, old school. Pokemon. Original Pokemon. And you go into different towns, which is different modules. You go to the shop, you can customize your avatar, you gain gold end levels and experience. And people have actually used that a lot more than they have the basic mode.

And we actually had a 93% completion rate of our full 15 mandatory-ish modules. It was all beta testing at the time. So they weren't mandatory at all. But I think surprisingly to me, adults enjoyed the gamification of it a lot more than I thought they would because I think people are kind of sick of just the boring, same old reading. For sure. A long thing about ... And then most people, because I've had a lot of analytics in there, most people kind of scroll to the bottom, go to the quiz and see if they can answer the questions. So one of the changes I made immediately was that once you start the quiz, you can't go back to the learning. So you better make sure you did actually read it.

Jillian (07:23):

I'm sure you made some people a litle aggravated with that feature. That's a really unique approach. I mean, the gamification makes sense. It's fun in this case because it does seem like a lot of people have been intimidated or maybe a little bit fearful of AI. So what's been the response from teammates who have gone through this? What's the feedback?

Samuel (07:41):

So the feedback has been brilliant actually. And exactly what you said, they've been scared mostly of there's two prongs. Where do I go? How do I start? A lot of people see the empty chat box and it's like, what do I do? I don't even know what this can do, let alone what I should be doing. It's almost just the same as Google to then. It's just a box they don't know what to do with it. So getting people to know exactly what they can do, what they're allowed to do and what their role would be good for them to actually use AI for. And then number two thing is they don't want to code themselves out of a job. I think that's a real fear a lot of people have. And I really kind of wanted to push on people. This is not the point of AI, though a lot of companies are doing it.

This is not the point of II. All it is, is to take the boring parts of your work away from you. The parts that you're doing, taking an hour wrong that there's no creativity, it's either right or wrong. You're doing it the same thing over and over again. Those are the things that AI is brilliant at. If it has creativity involved, I don't want AI near it still personally, but or everything else AI is here to just help you do a better job and do things only humans can do because we are still-

Jillian (09:01):

How are you helping people find those use cases?

Samuel (09:04):

So I've been having little sessions. So I've had a 101 course or 102 course along with the platform to go 102, you have to complete three of them. To go to 201, you have to complete all of them, et cetera. And the cool thing about the tests is that they're not just multiple choice questions. They're also AI created written questions. So the way I have that done is very similar to the presentation I saw this morning where you have different agents with different specific tasks. So I have three different agents. I don't want to be wrong about this. I'll just look up exactly. So I have a three graders, one's a precision analyst, one's a quality reviewer, one's a practical reviewer and then the overseer one looks at everything that they've done, sees if they agree or if they disagree, picks which one and then gives you an overall score if you passed or failed on that place.

And that's been the part that has really helped people because people just kind of fly through multiple choice questions. But these ones, there's no way to fake it.

And so people have actually had to learn how to write good prompts and prompt engineering is the main challenge I think obviously and reminding people that this system is just the output you got is because of the input you put in. That's the only thing that's going on. So you can't get mad at the AI. You have to get mad at yourself, unfortunately. And you have to realize that it was the reason why you got a bad answer was because your prompt wasn't good enough. So I also have a resource, like a prompt auditor for people to put in a prompt they plan and I have again, different agents that look at different things. They can also create little subagents themselves.

Jillian (10:55):

You're teaching people how to do that?

Samuel (10:57):

No, it's a system I built into the learning program. So I could show it on the phone, but I'll do it. And so I use it all the time. I think the tools that you build for yourself always end up being kind of the best tools because you actually use them every day. I ended up first just building it for myself because I wanted something cheaper than ChatGPT Pro and faster. Didn't want to wait 50 minutes for every single time I wanted a good answer. So I built a kind of multi-agent working at the same time for agents looking at my thing and then something to look over the answers that it all gave and give me back one prompt with a proper recipe, roles, goals, out of scope, in scope, what does complete look like, all that kind of stuff. So that's been really helpful for people too, actually.

Jillian (11:50):

The change management and learning process is such a common challenge for organizations. What has been the overall business impact since launching this and getting teammates through it? What are you seeing as a result?

Samuel (12:02):

I mean, we had a workshop last week where people were able to, for the first time, show us what they've done with AI. And I was blown away the amount of people who have never coded before in their lives, writing Photoshop plugins, creating little mini apps in terminal. That was the first time they'd ever opened terminal in their lives and people coding who have no idea really what they're writing. And it's at the stage now where most of the time it gets to write the first try. Whereas a year or two ago it put it wrong the first try pretty much all the time. So we're just enabling people instead of having to go to us developers with an idea. They have an idea, they can try it out themselves. And oftentimes they can do it if it's simple enough. And my thing to say to everyone is, if you do something twice, you should start thinking about automating it.

If you're doing it three times, it needs to be automated. And if you can't, that's fine, but you need to try at least try.

Jillian (13:00):

Yeah. No, that's a good framework.

Samuel (13:01):

Yeah.

Jillian (13:02):

And you're teaching them, you mentioned prompt engineering as part of the courses that you're teaching

Samuel (13:06):

Them.

Jillian (13:06):

What else are you including then there is like fundamentals?

Samuel (13:09):

Yeah, so I can show you. This is all about OpenAI because that was what we were using, but next week is my rewriting this for Gemini. So it'll be in the context of OpenAI, but soon it'll be Gemini. So basically, first of all, they have to go through the UI basics, then prompting fundamentals, then custom settings and output control and then working with files, image generation, code interpreter and data analysis web browsing and how you can make sure that you don't get hallucinations. That's a big one, really important. Canvas, which that's an OpenAI feature, Google has something similar, voice mode. So there's some very specific things that I wanted to drill down on, but breaking up into little segments like that so they don't have to read for half an hour once. It's more like 10 to 15 minutes per segmentation. It's

Jillian (14:06):

A lot of capabilities and we know that the tools are constantly changing. So how are you keeping up with it and how are you helping your teammates keep up with it?

Samuel (14:13):

So yeah, I'm kind of resetting everyone's progress pretty much every month and a half and making them redo most of the courses because they've completely changed. I mean, I look at my course now and I was still mentioning GPT 4.0, which is months ago in some of my things. I find most of the courses I have to write, rewrite pretty much every month. This

Jillian (14:38):

Is a

Samuel (14:38):

Full-time

Jillian (14:38):

Job.

Samuel (14:39):

It is a full-time job just doing that. I also have to make these apps myself and worry about security, make sure no one's putting public queues and JSON files, stuff like that.

Jillian (14:51):

And for the teammates who are going through this, how much time are they dedicated to the learning?

Samuel (14:56):

So we still have a lot of work going on. So it is, I would say probably an hour or two a day that we allow them to have time. A lot of the more dedicated people are doing on the weekend in their own time because it's not just useful of the work, it's also useful for personal life if you understand what it can do and also most importantly, what it can't do and where to avoid using it for being a psychiatrist for example, maybe don't use it for really, really important clinical things, stuff like that. But yeah, kind of teaching people to use it in their personal life I've found really helps them understand better where to use it in their professional life as well.

Jillian (15:40):

So you mentioned you had a meeting recently, everyone kind of showcased things that they've been trying. Has any of it resulted in a new way of working or a new offering or a new tool that the company is now using?

Samuel (15:52):

Yes. So a bunch of Photodrop plugins we are kind of merging into one tool and although for now it's just internal, it's something that we think could be a SaaS product kind of thing. So we are looking at actually putting it on the Google Marketplace because it kind of encompasses so many things that we do. It saves us about an hour a day per person and there's a lot of people that do similar work to us for smaller studios or even the studios themselves, if they often want to take the work away from us and do it internally, maybe it's an option for us to still have some work with that.

Jillian (16:28):

Let me ask you something else because you talked about people who have never coded, never opened terminal or now doing these things. One of the things we'd often talk about AI for people who are embracing it is that it does unlock new capabilities. It's blurring roles in organizations.

Samuel (16:45):

Really is, really is. Yeah.

Jillian (16:47):

What's your take on that? Is there any fear? Is there any discomfort or friction that you're seeing by empowering someone to do things that normally would've just gone to the engineering team?

Samuel (16:58):

So there is some friction, especially between the normal non-coders and the engineering team, especially because the engineering team has been not as fast as I have been and I'm not in charge of the engineering team immediately. I technically am, but they have an official boss who's not as gung-ho about AI as I am. So they almost feel a little bit like, "Hang on, why are people doing this when we should be doing it? " And I've had some talks with them, introduce them to the agentic system and how it's not just you copying a pasting code anymore, it's actually letting the AI look inside your own code base, understand the context window. We have a million tokens now, very few systems have more than a million that you'd need for it to understand. But I think having one of my rules is that if you're going to write some code, you have to have a read me file for every new folder you create in order for humans and AI to understand what on earth you're doing.

So if someone does want to take a look at it, it is explained exactly what's going on in each file and there's none of this kind of stuff that no one really knows what it's for. AI wrote it, didn't document it, think I'll come back. You often see in the thoughts, I'll come back to that and document it later. It's like, "You can't do that, mate." I think the empowerment is better than the fear. It empowers people more than it does give fear to people. There's a little bit of pushback, but there's more excitement than pushback, I would say, generally.

Jillian (18:44):

Is it starting to change the scope of what people are responsible for?

Samuel (18:49):

Definitely. And especially because in my opinion, in five, 10 years, if this keeps on this trajectory, most people's jobs will be prompt engineering and that's it. They'll be talking to some kind of AI system and telling you what to do. And so it'll be very specific prompt engineering for analytics for finance because when the system can do it all, all you need to do is tell it what to do. And if it's slower to do it by hand and less accurate, it's kind of a waste to do it by hand. So teaching people that prompt engineering is not just important for what I'm trying to teach you now, it's so important for your future a job anywhere really. It's going to be like being able to type is going to be just as important, I think, in the future. Yeah.

Jillian (19:43):

I'm imagining that, especially in more of a creative space, you probably had a group of people or some individuals who were fearful, maybe resistant to the change before- Yes, very much so. ... like in the before. Do you have a story, you don't have to name them, we won't do that, but do you have a story of somebody who was in that camp and has come to the other side? Yes.What is their life before and after?

Samuel (20:05):

Yeah. So there's a common misconception that the audio, video and image and text models are the same thing. Not a lot of people realize that they are completely ... One's an LLM, one is they're completely different, the other's also completely different. They're only related and they're all AI. The LLM, although it says that it's generating it for you, it's not. It's sending that over to a different model and then getting it back. So letting people know that the audio and the image is separate to what I'm talking about, which is mostly just LLM stuff, helps people understand that although the image models are the things people are most scared of, I would say image and being creative. And that's not what I think is the breakthrough. I think that the image models, although they're brilliant at making new images, what we need it for, editing, which is still not quite at a hundred percent level, which is what we need.

So I like to kind of remind people that learning and embracing the LLM is not the same thing as embracing these image models, not the same thing as embracing the audio models that create music. I don't really like those. I'm a musician, I know I'm a photographer. I think that you can be pro AI and still not enjoy those things, which kind of helps people comp up, analyze things and realize that, okay, if I just talk to ChatGPT and help it out, I'm not helping out eventually make my job obsolete because it's just text. So teaching people that those are three completely separate systems and that the text LLM there is the breakthrough. And my personal opinion is that until very recent, until NanoBanana really, the image models kind of just piggybacked on AI's amazingness, that's what I think helps a lot of people. And then also showing them how bad it is at doing some things, how it hallucinates, how we are still very much needed and that there's very few systems that can work with no humans in the loop and that any technology is going to make it so there's less jobs that's not exclusive to AI and that making sure that your job is not taken away.

The only way to do that is to learn AI. It's not to ignore it. You have to embrace it so that way when you can say, I can do this 10 times faster, not it can do it 10 that faster. Yeah.

Jillian (22:50):

That's a great reframe. I'm going to leave our listeners with one actionable item. You can't say, "Just go out and do it. " I'm not lucky you don't want to. So what's the advice that you would give a leader today who is trying to get broad scale adoption of AI in their organization?

Samuel (23:09):

I would encourage people to show failures as much as they do successes. So if you have a meeting like a workshop, things people show don't have to be things people succeeded in. It's just as useful in my opinion to show things that didn't work using AI that things that did. And that I think takes a lot of the pressure off people who aren't willing to say, "Well, I've tried loads of things, but none of them worked." If you tell them that's fine, chose the things that didn't work, then they're much less nervous I found to speak out about it.

Jillian (23:43):

Yeah. That's a very simple and practical tip.

Samuel (23:46):

Yeah. Just do it. I'm kidding.

Jillian (23:49):

Thank

Speaker 3 (23:50):

You so

Samuel (23:50):

Much for checking with us

Speaker 3 (23:51):

Today.

Samuel (23:51):

This

Speaker 3 (23:51):

Is

Samuel (23:51):

Fascinating.

Speaker 3 (23:52):

Cheers.

Samuel (23:52):

Yeah. Yeah. Thank you so much.

Speaker 3 (23:54):

Thanks for listening to this episode of Insight On. If today's conversation sparked an idea or raised a challenge you're facing, head to insight.com. You'll find the resources, case studies, and real world solutions to help you lead with clarity. If you found this episode to be helpful, be sure to follow InsightOn, leave a review and share it with a colleague. It's how we grow the conversation and help more leaders make better tech decisions. Discover more at insight.com. The views and opinions expressed in this podcast are of those of the hosts and the guests and do not necessarily reflect on the official policy or position of Insight or its affiliates. This content is for informational purposes only, should not be considered as professional or legal advice.