Personal branding, new EU requirements and AI for ad buying #118

A warm welcome to a new episode of Noll till hundra. My name is Fedja Porobic, and as usual, we have Fredrik Ankarsköld with me. In today's episode, we'll jump straight into the main topic. You've started focusing on building a personal brand. We're not just going to talk about your personal brand, but also about why you should build a personal brand at all.

That's a very good question. I've started something I've been thinking about for a very long time. I've built a personal brand over time by sharing stories and best practices on LinkedIn. We've had this podcast, which is completely unsponsored. We don't earn a single krona; we even pay a lot. You picked up loads of equipment for the podcast this morning.

I was completely shocked when I saw all the cables. Now there are microphone covers too, pop filters, perhaps they're called. Why do we do this? Well, we enjoy it and reach lots of people. I know so many listen with dedication, and that makes it extra fun. People ask questions and say hello to you in town. Maybe not in town, but at an event.

It has happened a few times that people I don't know recognise me out in town. But I can imagine it's also nice to get the benefits of being recognised without being recognised by the general public.

In your niche. The big advantage, I usually say, whatever niche you're in or if you're just building something, is that whatever you do in life, you gain an enormous advantage if you have a thousand followers somewhere. Or just a thousand people who like what you do and know what you stand for. If you accumulate a thousand followers on social media, life will go better for you overall.

In what way? In every way. You gain an enormous advantage in working life, whether you're employed, going to build a company or chasing new customers. Whatever you work with or do, you get an enormous advantage if you have just a thousand followers. Cracking the code in today's society: followers are the new… What's the word?

Currency, perhaps? Yes, perhaps. But I'm also thinking about engagement. It's not just about subscribers, but about contributing real value. Of course, you have to have some value, otherwise you get no engagement, nothing. Your videos won't be seen. When we started posting Noll till hundra on TikTok, we immediately got a thousand, five hundred or two thousand views, which is a lot in our little niche.

We post and reach e-commerce nerds specifically. Two thousand views is hopefully two thousand e-commerce companies that have seen this. Then I see in comments, [unclear in the Swedish source: i link] and questions that it's very much the right people. Then I saw another podcast posting, no names mentioned, but they get thirty views.

Obviously, it's also possible to completely flop with your content and get very few views. But thirty views is a lot of people too. If you reach thirty of the right people every time, repeatedly, [unclear in the Swedish source: trettionia kanske titt som tätt]. When you fill a room with thirty people, it's a full room.

I'd like to give a tip there: lots of views don't mean everything. I recently listened to a guy who showed his best videos based on views. Then he showed his best videos based on the revenue they'd generated. Those that generated the most views weren't those that generated the best results.

No, I buy that. But it can also be that those generating lots of views sparked interest in the profile, podcast or whatever it is. Then you go down a deeper rabbit hole.

In his case, the general videos with general advice could apply to more people. That's why more people were interested in the video, but it wasn't his target customer audience. Reasonable. I buy that. We see time and again that when I post things on LinkedIn, going back to 2018 really, it doesn't have to be the posts with the most views. Some posts have had 100,000, 200,000 or 500,000 views on LinkedIn.

They've reached extremely broadly, far outside the core audience, but generated nothing in terms of inbound. Then suddenly we post things on TikTok, YouTube and everywhere about our podcast. People start getting in touch and writing in the comments. We even got comments on Spotify. Then I think it's the right people getting in touch, writing, wanting to work with you, meet or have lunch.

Now, I'm not someone who takes much inbound, because I have no space in the calendar. Literally, I've turned down every single person who's contacted me. But I try to do it kindly and say I've blocked the calendar for the rest of the year, and we'd be happy to talk next year, whatever the case concerns.

Unfortunately. I can say that even for us, it takes quite a bit of planning to make the recordings happen. It does. But back to personal brand: I think it's very important to build a personal brand, whatever your role and what you do, whether you're employed, self-employed or any kind of creator.

As long as you put out things that deliver value, it builds authenticity and personality. People get to know you without being in the same room as you, which is good in every possible way. I usually look far ahead; I play the long game. I share my best tips, hacks and growth strategies. If I do that constantly and frequently over time, people will understand this is the real deal.

One big reason I started posting on TikTok, Instagram and everywhere was that I wanted to get the message about the podcast out and get more people listening. That was the first thing. Then I thought perhaps I'd mix a few everyday situations into the feed so it doesn't just look like I'm promoting a podcast.

I did that too, and those have also gained good traction. Then I thought: what's the goal of posting things from everyday life? I've called my newsletter ‘Inside an entrepreneur's head’, but what's the goal of it? Now I've figured it out.

What is it? The goal of posting more frequently is to reach beyond the e-commerce audience. I want to reach kids who are starting their e-commerce careers, are a little lost in life or are trying to get out of the suburb where I was. I don't know why I want that, but I feel a calling in life. Growing up, I had many international role models.

There was nobody who'd done what I wanted to do in Sweden. I had to look to Neil Patel, Sean Ellis and David Arnoux in the Netherlands. By the way, a bloody good guru, brilliant at growth hacking. These were some of the leading people in their niche fifteen, sixteen years ago. Obviously, some of the world's leading people even today.

They were people I contacted and reached out to. I even sent letters, literally, to get in touch with them and learn more from them. It was hard to contact these people, but regardless, I learned a great deal. As I learned more, I also got better at what I do.

I got better at e-commerce, could perform better at work and found a calling, a mission: to become really great at this. If I can inspire just one person to choose the right path instead of another, I feel I've succeeded in the mission of sharing more of my life and stories.

This is in contrast to there being so many course sellers today who only sell online courses. That's their sole goal with their personal brand. Why do you post on TikTok and Instagram? Well, to sell an online course for 10,000, 20,000 or 40,000 kronor. Really, these are scam courses. In 98 per cent of cases, the buyers are young people who feel lost and uncertain in life.

It's set up in contrast to a higher vocational education programme that might take two years. This is a course that's quick, but the information is dated as soon as it's created. The courses don't teach how e-commerce psychology or marketing psychology works, or how people act in certain situations: the basic, fundamental things. Instead, they teach tactics that work here and now.

Usually quick-fix tactics: how to find a product, create a site, do this and that or set up Google Analytics. Fine, you can go through it, but you can also ask your AI. It's not that hard to set up Google Analytics, Google Tag Manager or your Shopify store today. I'm frustrated that so many course sellers are talking so much rubbish out there.

So I wanted to share the real deal: real life behind entrepreneurship, investing in small brands and zero-to-a-hundred journeys with different brands. How we do it and why. I want to put out as much relevant content as possible so people can follow and think: ‘This is the shit. This is what I should do to succeed with my brand.’

If they do it and succeed with their brand: applause. It makes me fucking happy and warms my heart. Lots of people get in touch, and I meet many at events who say: ‘I've listened to your content for many years. I've done exactly how you've structured your Meta account, or exactly this, and it's worked so well for me.’

‘Thank you so much.’ That's exactly the feeling I want to get out of it. For me, it's become a personal calling. I don't want to crush every course seller, but I'll be damned if more young people are going to be drawn into that black hole, buy the wrong courses and think they're going to… Do you see all courses as bad? No, actually not.

There are good online courses in specific niches that are practically free through Coursera, other [unclear in the Swedish source: e-commerce sites] or Khan Academy. They're pretty okay. There are paid ones. ‘Hate’ is a strong word, but what I dislike is the fake lifestyle gurus showing flashy watches: ‘Look here, I have an AP. It costs this and that.’

‘Here's my car. Drive the same car as me. Here are luxury houses.’ I don't like that kind of content. If you have an AP or a Rollie, wear the shit, but you don't need to show off so much beyond that. I love watches and the horology behind them. I've been to exhibitions in Switzerland and Milan.

For me, it's almost like an art exhibition. I love hearing the story behind it and seeing the craftsmanship when a watch is built. It's amazing. I recently met a guy at a dinner who worked on the gears in some of the watches. He could spend forty hours working on one gear by hand.

Exactly, it's the same thing. I've seen it live: just putting two parts together takes an hour. The precision and care required. I enjoy it, love the craftsmanship behind it and can easily pay for it. It has personal value to me.

But it's not something I'll then go out into the world and say: ‘Do this, follow your dream, buy my course, and you'll live the same life as me too.’ Fuck no. That's not what you want to encourage. I recently saw a person I've followed and watched now and then for many years. Now he shared: ‘I've sold courses all these years.’

‘At the same time, I've worked with real products.’ Then he shared a graph: he started a new store, it went bloody well until it didn't, and then it went down. The course business kept the finances afloat so he could have his sports cars and things like that.

Then it went up again: he started a new store that went bloody well. He kept doing that year in, year out. Then it was: ‘Now I'm only going to build real products.’ He closed the course business, but still didn't succeed. There was a dip, and suddenly he was making a loss for a period.

Now he shared it and said: ‘That's how it was all the years before. But now I've come up with a new course that has nothing to do with products, because physical products are so hard. Now you can do white labelling with vitamins. Buy my course about it, and you can succeed with that instead.’

Then I thought: you haven't succeeded all those other years with all the other courses you've sold. Why would vitamins suddenly be different? This is exactly the counter-trend and reaction I publish content for. I don't care whether I have two hundred or two hundred thousand followers tomorrow. I literally don't care.

Could not care less. I want to create a counter-reaction for the young people who often fall victim to that kind of content: fast sports cars, cool watches and nice houses, with a course behind it you're supposed to buy. The goal is to become like me: ‘Look what I've achieved.’ A lot of it is fake too.

Borrowed cars, rental cars or leased cars, rented houses perhaps in Marbella. They've rented a house for a month and make a hell of a lot of content so it looks as if they've lived there for ever. Or a flashy watch they've borrowed, or made content with and then sold on. That's what you want to stop, because these courses are often pure rubbish and dated as soon as they're written.

There is an idea and a point to some of the courses or parts of them. We're talking about fundamental truths and things that might also work right now.

Okay, those exist. But it's still not the path that will… It's the mentality around how you get people to buy the course that bothers me.

Okay, so it's not really courses themselves, but people selling a flashy lifestyle funded by the course business that bother you. Yes. Indirectly, it becomes: ‘Look, I've built this e-commerce business or this dropshipping company.’ But really, you're living off poor young people, usually guys, who buy your courses.

Enough said about that. Perhaps we should talk more about Swedish e-commerce. Swedish commerce declined in August by… Nine per cent, I think. Exactly. Nine per cent is still a lot. Last year, in 2025 versus 2024, Swedish e-commerce increased quite significantly in August. I think it was double digits. This year, it's falling by almost double digits, nine per cent.

Then it was mainly foreign transactions that decreased. Exactly. It wasn't as bad as it looks. In the data, we see foreign transactions decreasing from 23 to 16 per cent, or 23 to 14 per cent. That's quite a significant decrease in foreign parcels.

Yes, this is based on what we talked about two episodes ago regarding Shein and Temu. If you remember that episode, we discussed there now being a three-euro penalty fee on all foreign Asian parcels coming to Sweden with a value under 150 euros.

Really, a little the reverse of the American de minimis rule when we were shipping parcels to the US. If we exceed a certain value, we have to pay duty. Now it's if you fall below a certain value and import a parcel under 150. It's aimed directly at Shein, Temu and the large Chinese platforms selling rubbish products to Europe. This is at EU level, nothing Sweden could have regulated anyway.

I think it's fantastic. What do you think? I think it's good. One reason is that there's a lot of junk that's often low on safety. They have no chemical controls, which is a risk to users. In studies and measurements, it often turns out nine out of ten products tested, for example Labubus, are bad.

It was on TV4 News now about those squishies that are super popular and have gone viral. Nine out of ten contain harmful substances. Children have even put them in the microwave, and there have been catastrophic injuries. Some have even… Yes.

That was the craziest thing. God forbid. You need to keep an eye on your kids. Some products are so directly harmful and corrosive that you really should stay away from them. Most come through Shein and Temu. I'm very grateful and happy about the development and that the EU has actually taken responsibility in consumers' best interests.

I can recommend a series on Netflix called Broken, and Rotten too, for that matter. I'll look at them. I haven't seen them either. I'm not very keen on watching lots of TV, Netflix and such, but with things like this, you want to learn from it. Anyway, something else happening within the EU is new producer responsibility for packaging.

Small companies find it very difficult to ship parcels internationally, or within Europe, because we have to register in every country. If, for example, we at Såpa are going to ship soap products to different countries, we have to register and report to the country we ship to exactly which packaging and how much is being sent.

It's very difficult. We have so many products, different packaging and sizes, and ship to different countries. Each country has to have its own system and its own reporting. For a small company, it's almost impossible. The EU has taken a lot of shit, especially from southern European companies.

Now northern European companies and European companies overall are also starting to challenge this. What's being said is that only big companies can handle the new processes, which throws obstacles in the way of us small companies. It's a big issue on Västerbottens Såpa's table right now that we're trying to solve. Of course we'll solve it.

How hard can it be? I know the EU fundamentally wants to protect the consumer, so it has a good purpose. The thinking is right, but the implementation is wrong. I'm thinking of something similar I saw in a discussion within the EU: they want to reduce crop spraying.

That sounds great. But at the same time, they've got another thing through: they want to reduce the amount of land used for crops. Those two things don't fit together. It sounds good that we shouldn't use as much land for crops.

Why? What's behind it? No idea. Perhaps reducing the consumption of forests and such. I actually don't know; I haven't looked into it enough. It sounds good that we shouldn't cut down forests to grow crops, and that we can grow on less land, like vertical farming. I don't know, but those two things don't fit together. The long-term consequence could be too little food: more waste due to less spraying. Less spraying is fantastic in itself, because there's a great deal of PFAS and things like that in much of the spraying.

There's new technology allowing vertical farming, so you can grow upwards with fairly modest means and less water, around ninety per cent less. I bought lettuce from a company called [unclear in the Swedish source: Green and Growing].

But my question is: does it contain the same amount of nutrients? We humans often say we've invented an artificial solution to replace nature, and that very rarely comes without consequences.

Exactly. I don't know, because I haven't tested those crops. What I do know is that they're [unclear in the Swedish source: less], because they don't require spraying, since it's vertical and doesn't go down into the soil in the same way. It's still soil. But you're missing the soil, aren't you? It's still soil.

It's still water. It's just correctly proportioned per crop, precisely proportioned instead of a bit random. I don't know if it's less nutritious. I know it's very tasty, from the tests I've done. I don't know. Usually, taste tends to be connected with nutrition. When tomatoes are artificially produced, they basically taste like plastic here in Sweden.

Then you go to Italy and everything tastes much more. Yes, but it does in Sweden too if you've grown it the right way. I'd like to add a clause there: that applies to natural ingredients. I see it as the body often being drawn to what it needs, as long as it isn't processed food.

If something tastes extra good when it's processed food, it's usually for the wrong reasons. You can't say this pizza or Snickers was extra good, because then it's usually for the wrong reasons.

That's true. It's more engineering than nature that makes a Dorito taste the way it does and makes you want another, another and never be able to stop. We have a few other things, then. What's happening in AI right now that directly affects e-commerce businesses?

There are so many AI news releases that it's hard to keep up. Which is top tier? I saw a report that in 2023 and 2024 there were around 73 or 78 days between the major AI models from Anthropic, OpenAI, Gemini and so on. Seventy-plus days is quite a lot.

Now, in 2026, there are eighteen days between releases. Eighteen days, Fredrik. Almost every time we record, someone has released a new major, groundbreaking model. That says it all. In recent weeks, we've had OpenAI's Astra model, which is completely mind-blowing. I've had time to experiment with it a lot. It came at the end of August or beginning of September and set a completely new direction for what you can do with AI.

The title was even ‘Welcome to the AGI era’. Since then, we've had the Opus 5.5 update, Sol and Luna. Now Sonnet 5.5 has come too.

Now Sonnet 5.5 and Sol 6.1 have come. Sol is OpenAI's model on [unclear in the Swedish source: Opus], really a cheaper model, but faster and above all more cost-effective. I tested Sol 6.1 yesterday. It has levels too. Max isn't the best, but the second-best Sol.

Then you get almost Astra Max-like results, much faster and with much lower energy requirements. Fewer tokens if you're on the API, or usage, as it's now become. Another update I saw Claude, sorry, OpenAI, make is throttling the 20X plans. First, we need to explain how AI usage works in 2026. You have a regular plan that's usually called the free plan.

The free plan lets you use chat, a little image handling and basic functionality. As soon as you start paying, usually two hundred kronor a month, you get access to Plus plans on different platforms allowing more usage. You still don't have the latest model or features. If you pay a hundred dollars a month, you now have the latest features and models, but weekly usage called 5X.

Two hundred dollars gives 20X. Until yesterday, that went a very long way, and I paid for it. Two hundred dollars for 20X usage. I almost didn't run out in a whole week. I burn it in one and a half or two days.

No, on 20X I built an app, a website and presentation material. I built all sorts of things and have my morning email. My record is less than half a day before I've burned it all. That's what happens if you don't know which model to use, how to use it, prompt and set end goals.

It's easy to burn through. Yes, but I'd say I'm doing very complicated architecture work. Exactly, that's what I was going to say: if you don't know how to use AI, or if you're doing extremely advanced reasoning, as it's called in AI. If you require a lot of logical thinking in [unclear in the Swedish source: station]…

Strategic, perhaps. Strategic and logical, reasoning or deep learning, will consume a lot because it reasons with itself. That's really the model. Now they've throttled [unclear in the Swedish source: the Quadex variant] on 20X down to 1X for the same money. You get 5X for a hundred dollars and 10X for two hundred dollars. There's really no reason if you truly need usage.

But it's still heavily subsidised, we should say. A hundred or two hundred dollars is absolutely not the actual value or actual cost for the companies behind it, which are bleeding money. I got 250 dollars in [unclear in the Swedish source: cloud credits] from Claude.

How long did it take me? Six hours to burn through them. Cloud credits or API tokens are their actual cost, one to one. Once you exceed what's included in the subscription, Jesus, it goes fast.

Incredibly fast. I think those tokens are their actual costs. I haven't looked into that in depth. Whereas these plans, Plus, Pro and so on, 20X and now 10X, are subsidised and have much more power and bang for the buck. But do those who had 20X get to keep it, or is it only for new people?

Is it new users who have this? No. First, they stopped taking more users on it because they quickly saw they were losing too much money.

Then they've told those who have 20X: ‘Now you have 10X.’ Including me, and me too, because I have both. The latest news OpenAI released: they had their dev launch with seven major new releases, a new dev API, bloody good, and new desktop features. Above all, they've released agents, agentic stuff.

Last week, we saw Meta release Muse, which has had two million downloads in a short time. Muse is an agent Meta created to help with everyday tasks, for example shopping, booking trips or following up an email. Little things and reminders in daily life: don't forget this, your schedule or if you have something in the calendar.

Muse reminds you proactively. You don't need to prompt, discuss or set anything up. Just download it and make sure you've connected your email, your life. And given all your data to Meta. No, nothing bad can happen there. But as soon as you do that, you get a lot back from the app, and lots of people have jumped on it.

‘Haven't dared’ is a strong expression, but I haven't tested Muse to the level of connecting my whole life. I already have other separate applications. I know Amazon has banned Muse. Amazon has also tried to sue certain AI agents that go in and shop, and tried to frame it as unauthorised computer access.

There will be lots of those. They've had several rounds now. So far, as far as I know, they failed in the last round and didn't get it through. It feels ridiculous to put a stop to new technology. Amazon's AI is a little behind, so it's not surprising they're trying to push back. But never mind.

We're seeing more agents released that are easier to get started with, more for ordinary people than power users. Grokbot was the first, six or seven weeks ago. A superb platform; I think we mentioned it on the podcast. I've given the team lots of talks about it. I set up Grokbot for my business in Italy, and it works excellently behind the scenes.

It does everything I need: updates room prices and answers customers. For those who don't know, you rent out apartments. Yes, apartments I rent out. On paper, it's quite a simple, straightforward business. You need to update and work on it a lot, but AI can take over a large part of the work. Now it also does forecasting, logs, bookkeeping and much more.

Each bot has its own task, which I then coordinate through Oliver, who's head chief of staff. The news now is that OpenAI is releasing its agents. What do they call it? Dots. Have you tried them?

You can't, because they're not live in Europe yet. They've been released in the US first, like many things, but it'll happen any time, probably soon. Usually, it's about some kind of security approval. First from Trump and the administration in the US, that's what they care about most, then the rest of the world. One tip there is to actually use…

I use ChatGPT's agents. They're very good at controlling the mouse. I usually use it for bookkeeping, because they can drag in receipts and things without any problems. Historically, Claude hasn't managed that. It got stuck on how to take a receipt from Finder on a Mac and drag it onto the right area in the browser.

That was a problem it had. I've had the same problem, but Grok managed it. Grokbot has its own screens and computers. They don't take over your screen and do nothing on it; they have their own. You can see their screen, literally see an AI. Once you've got access and given it access to work, you can see it in real time.

It's quite amazing the first time you see it, and they communicate with each other. You see: ‘Send this to Oliver. Send this back. Oliver is reporting this.’ You see how they talk, the dialogue and communication.

Do they do real, useful things, then? Yes. So it doesn't just look cool but is a gimmick? No, it does exactly what you want. You can start it at night. We schedule it to start then and take over customer service at night too.

Then it learns from how you've chatted historically. Shit in, shit out. If we take three tips for e-commerce businesses: what could they use Grokbot for? Absolutely. Above all, look at simple tasks that can quickly be automated. One such thing is research. Every week before a creative meeting, you need to gather what competitors and others in the industry have done and what's trending right now.

That takes a lot of time. An easy way is crawling Amazon, which has blocked many AI agents: trending new products and reviews in your industry. Negative and positive reviews tend to stand out. I usually look at five-star and one-star reviews. What are the pain points? Let AI make a summary. Every Monday morning when I wake up, I have that for different brands.

When I go into growth discussions, I know what's trending in the industry, which ads are trending and which brands are doing well right now. I look mainly at American, Australian and southern European ones, but also a little at how they do it in India or China.

When you look at ads, do you let it go into Ads Library? What do you do to find the trending ads? I have an external tool that I let it browse through.

What's it called? [unclear in the Swedish source: Filter]. I look at AdSpy and BigSpy. I've also used AdSpy before. They basically download all the ads that are published, making it faster for AI to find them. You can also filter by how many days an ad has run or how much engagement it's received. You can't yet filter by how much money someone has spent; we don't know that.

But you can see how many it's reached, which is a proxy for how much money has been spent. If it's reached three people, they haven't spent anything. Exactly. If it's run for over two hundred days and reached over a million people, that's a proxy suggesting it may be a good ad. It breaks that down.

It can also see all the frames in it and make a short summary with a link: this is why the ad is doing well. We won't do exactly the same, but can take inspiration from an ad in a completely different industry or vertical and make it our own. When something trends, we quickly see that someone has often found a new format that works.

I usually download different videos myself using software. They might be ads or YouTube videos I want to understand better or change the context of. Then you just drag them into ChatGPT. For YouTube, I use 4K Video Downloader. Then you can choose 720p, download the video and ask ChatGPT to analyse the transcript if you're using Work mode.

You can also analyse frame by frame. It's wasteful if, for example, it's a slideshow, but it can read when it perceives a difference in the video. Smart. That was number one. Number two for an e-commerce business? There are so many use cases.

I'd easily say Claude, at least as I've tested it, is a good-enough media buyer today. It beats nine out of ten Swedish media buyers on the media accounts I've looked at, both in structure and optimisation. But isn't that based on you giving it the right instructions from the start? No, I've seen use cases where I wasn't even involved.

Above all, I've seen a complete beginner, a woman with a digital product, which isn't the same thing. She's sold a simple, cheap digital product on Etsy and got Claude to generate a lot of money for her. I dug into how it set up the account and ads, and how day by day it optimised, lowered and increased the budget. I saw: perfect, well done, applause.

I'd have made roughly the same adjustments. It even found a breakdown for certain audiences: this audience works better, but if we change to that breakdown, we'll perform worse across the account overall, so we won't make the adjustment. The reasoning behind it is mind-blowing. Most media buyers see something and just: ‘Bam, now we'll change it so we have more of this,’ without the underlying logic in mind.

The AI model Claude, and this was Opus too, handled it much better. It wasn't even the new 5.5, but the old 5.4. Today, I'm very positive about all media buying being done through AI.

So you're saying ninety per cent of all media buyers can be replaced with Claude? I think a hundred per cent can be replaced, but still that there are ten per cent specialists who beat Claude. That shift is happening now, over the coming quarter. Then no percentage will remain. What's valuable then?

Take a growth marketer, for example. A growth marketer can easily manage an AI with instructions. Just as you've had a specialist buying media, you now have a specialist who's a Grokbot, a Claude bot or now an OpenAI agent. That's your executor.

You still set strategy, goals and methodology and filter which creatives we should run, where and how. A lot of work remains. You can outsource everything, but it won't be as good as if you actually do certain things yourself. AI handles the tactical media buying itself, I'd say, to a large extent better than people today.

That's number two. Number three is manual reporting. It's so tedious and laborious to go into every system, export data and put it into Excel. Excel is still very clear, functional and easy to compare and send around. It's laborious and takes a lot of time, but you still want it done.

Now you can connect up. You have a built-in Shopify connector to these platforms and lots of others: Meta, Google, Analytics. It's all built in. You can extract the data simply and cost-effectively, get a clear overview and chat with it. You can ask follow-up questions: ‘What's your thinking here? What are you thinking here?’

Hallucinations don't happen when it's actual data. It doesn't write it wrong, because it's literally retrieving your data from a connector. You don't have to sit double-checking, although it can be positive to continue doing that. You can trust that if it's laid out your cohorts, you have the right data in front of you.

Then you can ask it deep, tough questions. Do you have an example? Absolutely. A person would have struggled to produce the data quickly; it would have taken time. An example is: what are my best products for first-time buyers? Surprise, often they're not the ones you're advertising, but a different product from the one you think is your hero product.

By ‘best’, I may mean highest lifetime value. Which entry products make people more likely to return for a second or third purchase? There can be big differences between products. At Såpa, a starter kit, limescale remover, window cleaner or deck soap as the entry product can give completely different outcomes for how customers make their second purchase.

We see that people who buy deck soap might not make a second purchase within twelve months. Those who buy our all-purpose laundry detergent, however, return within forty-five days and make a second purchase, both of the detergent and something else. We have great stickiness in that audience, which means we can pay more to acquire that customer.

Or reactivate the customer. Do you understand how much the marketing strategy changes when you know the underlying factors, how valuable the customers actually are? We call them cohorts. How important they are, when they return, from which product, category or ad. That takes time. The basic data from Shopify isn't that valuable in itself without context and a correlation with other data.

What does spend look like for the category or product? What does repeat purchase frequency look like for those categories and products? One problem for some marketers is that without the right data, they may accidentally switch off an actually profitable ad and let an unprofitable one continue. In the direct data, it looks like the best result or decision.

With more overarching data, it may turn out to be a worse choice. It's very common for a manual media buyer, now we'll call it an agent, to switch off the ad that doesn't seem to perform, but then the whole account performs worse. Why? Perhaps the ‘worse’ ad reached top-of-funnel customers and moved them down so they saw other ads, then the third and fourth they converted on.

In the data, it looks as if the third, fourth or fifth ad performs, while the first ones that got people interested in the brand don't convert that well. We switch them off, bam, and the other ads suddenly perform much worse. You see that relationship by looking deeply at how much people watch the ads.

Those who've seen this ad have also seen this one. The data is there, but people don't have time to go through it. The agents do, because they export the data. It's all code, so in the code you can see data and the relationship between them. We've said for many years that a media buyer will be replaced by AI, but now we're at the point where it can actually do it.

OpenAI's platform can do it with Astra underneath. With the new agents, you'll be able to do it more repetitively. Before, you needed to schedule a piece of work as a job for Astra. Now the agent handles it and becomes much more proactive. It'll even ask: ‘I see this happening. Do you want to act on it?’

If you don't answer, it'll act on its own. A little risky if it does it at night and asks a question at four in the morning. But you can instruct it to wait on those things.

Yes, you barely need to instruct it. After a while, it learns: ‘You haven't answered at four in the morning. I understand you're asleep.’ A tip when using agents: when I use Claude, we can see… Before it was Sonnet 5, now Sonnet 5.5.

You can get up to thirty per cent more done, with reduced cost, or rather a thirty per cent increase in efficiency. That's why I usually use Opus 5.5 when I'm working in my systems. It's the overarching brain for planning and strategies.

Then I've instructed it to have [unclear in the Swedish source: fable advisors]. If things are really complex and major errors could occur, or it gets stuck and can't find results itself, it has a strategist behind it.

It has a strategist in its instructions that it can call on without asking me. I even have several [unclear in the Swedish source: fable advisors] depending on the type of problem, so it doesn't become biased. If you ask an AI to criticise itself, it'll be kinder than if an external agent does it.

True. Then, when you ask it to do research, you use Sonnet 5.5. You don't want to use the strategic power to collect data. That way, you get much more data collection in a more cost-effective way.

We ought to rename the podcast the AI podcast, with all the AI tips we have, Fredrik. My God, that's what I do. Noll till hundra with AI, Fedja and Fredrik. What do you think? Should we start a new podcast that's only about AI?

Yes, then the old Noll till hundra can be called the Västerbotten podcast or something. Why not? We talk so much about soap. One last thing in AI, a life hack: I usually leave the computer on overnight and give it a goal to work towards. It doesn't end the task before reaching the goal.

To do that, in OpenAI, ChatGPT, Codex or Work, which I use, I write ‘/goal’. That starts the prompt. For example: ‘/goal. I want you to create this and this presentation, research or DD,’ due diligence. Something you want to do.

Then it works, keeps grinding away and can go on for hours. It doesn't stop before reaching the goal. The danger is that if you have tokens, it burns through them unless you have limits in between. For example: ‘If you can't complete the task after three repeated attempts and have tried different parts, stop.’

But I love that life hack right now. Otherwise, you can give it a task that it doesn't really solve, but still ends. You come back and wonder what the fuck: that wasn't exactly what I wanted, not the task's end goal.

One of the most important things when you have a goal is not just saying: ‘Set a goal and work towards the task.’ I've done that, and then it burns through everything without almost getting anything done. Then it's been too abstract. I recently spoke to a developer who said: ‘Among the most dangerous things we can ask is whether there's anything more around the goal we should work on.’

‘Then the agent will invent it.’ It's better when setting the goal to also define a plan and a checklist. Then you can check how far it's got through the checklist.

If you wake up in the morning and wonder whether it's finished, you can't say: ‘Are you done with the goal?’ There's a risk it starts making things up. Instead, you can say: ‘Have you completed these items on the list?’

Then it's more yes or no. An AI is non-deterministic. A bloody complicated word, but it means if you ask the same question five times, you get different answers each time. If you ask overly broad questions without defined items, there's a risk it starts inventing things to do just to please you.

If you ask whether it should have something to do, it wants to invent it. I get it. Good tip with lists. I haven't done that before, but I'll try it too.

I've had it because it… But my prompts are long too. When I write the end goal, I usually write: ‘This is the end goal.’ Then I don't want to say exactly how it should get there, what list or order. This is my end goal, what I'm actually trying to achieve with the report, analysis or presentation.

This is the approach I like. I'm inspired by these books, psychology, marketing tactics or something else. I usually put that in, and then it learns a lot. Here, I think in the same way as when delegating to a person. Last week, I had the privilege of meeting experts from France and the US who research decision-making, recruitment and optimising results. I think that's… Okay, that's amazing.

Bloody fun. It could be an episode in itself. The researchers talked about delegation. If we take a recruiter as an example, often their goal is to recruit new staff, but there's no follow-up. They don't find out how it went after two years.

That's true. But do they care about that? Yes, how else can they get better? At getting people to stay? You have no idea. You can say this was the best person.

Yes. Compared with tossing a coin, those who have an idea and plan for how they recruit were sixty-two per cent better than tossing a coin, as an example. Similarly, there's another book called Peak, by Anders Ericsson.

There, they have an example of a doctor doing an initial examination. Perhaps the doctor looks at an X-ray, sees that you probably have cancer and refers you to the cancer department. If the doctor never finds out whether the diagnosis was right or wrong, how will he learn which cancer patients he missed and which he was right about?

If you don't get any feedback. I think the same way here with the agent. I've put into my AI agents that they should go in and verify visually if needed. It can't visually verify code. But if the code is connected to a button, it should verify visually.

Take Claude, for example. It sends Sonnet into the browser to click around and on the buttons. Then it reads logs I've connected to different systems to see whether bugs occur there. For the visual side, I also have an instruction: it should assume the person looking is a first-time user who's never seen this before.

Then it has that perspective, and that way you can improve the experience. Shit, then you get much more out of it. Yes, it came from how, when I'd built complex things, it could say: ‘It's finished.’ Then I'd look and think: ‘But I don't understand anything as a human.’

I asked: ‘Why have you done it this way?’ ‘I build for agents because I'm an agent.’ It starts from an agent's perspective, doesn't look visually but at the code. Then perhaps it's a box inside a box inside a box inside a box inside a box.

Very logical to it. Yes, because there's a hierarchy in the code. Visually, there's no clarity: what's important here? Is it information or an action? I could get absorbed in any number of details there.

We don't need to, but there's a lot around AI. Another development I've seen is in content for e-commerce businesses. If we rewind one or two years, we could see AI getting fingers wrong: six fingers or short fingers. Then we saw mistakes with strands of hair. If an AI moved too quickly, you could see a little flickering on the screen.

Then we could detect that it was too polished. It absolutely didn't look… We could hear from the voice that it felt very AI, straight away whether a voiceover was made by AI or not. It also couldn't handle the details in watches. For [unclear in the Swedish source: mockups], we had huge challenges. This got solved.

Everything I'm mentioning has been solved. Then it was an enormous challenge to have two AI UGC creators or personalities on the same screen interacting, handing something to each other or passing a bottle naturally. It couldn't do it. Another problem was when they collided or bumped into each other. You could immediately see it didn't look real.

That's been solved. Everything I'm mentioning, they've solved over time. Now AI handles most content tasks in all the languages we need, around seventy-eight languages we advertise in, without problems. You need a few different platforms. Claude and OpenAI handle most things as the brain behind it.

It's often used by Higgsfield for motion graphics and by ElevenLabs for the voice. You should connect them through MCPs, an easier way to connect through connectors available in both pieces of software. We're going to watch a clip. You have to see this. I'll show a video first, and I've chosen an AI ad someone else generated.

They've put in a photo of the person and a photo of the product. The person in the source photo was in the same setting as in the generated ad. Then there was a description of roughly what the person should say, how long the ad should be and roughly how she should move the product.

The funny thing is, it didn't follow the instructions exactly, but it still turned out very well. Look at this. [Ad clip in English:] ‘I've been using this tumbler every day lately, and I didn't expect to like it this much. My drinks stay cold literally all day, which is kind of crazy. It doesn't leak, it fits in my car, and the handle is actually super comfortable.’

[The ad clip continues:] ‘I just end up taking it with me everywhere now.’ That's cool. That's unedited, raw output from the AI. I don't like the ad. It's generic, the hook is bad and everything about it feels like something that would have been great in 2023. Now that kind of content doesn't work, because it's become so easy to mass-produce.

But it's almost perfectly made. You don't see any specific mistakes: hands, strands of hair, the clicks on the bottle, lip sync and pronunciation. A lot has become good. There's still a slightly too-perfect appearance, or whatever you'd call it. But I wonder whether it's the source photo and person chosen? The person, yes. It depends on which photo you've chosen as the basis.

The interesting thing is that you have a window behind her and a wardrobe full of clothes. When she moves her hair, do you see? None of them become distorted or anything like that.

No, everything stays consistent. It looks good, even between scenes. I saw a cut in the middle, after five or six seconds. Yes, the AI did that editing itself. No human edited it afterwards.

It feels very realistic today. Perhaps eight out of ten for realism. But in quality as an ad, would I run it in my ad account? No, perhaps four out of ten. Okay, show the video I was talking about now, Fredrik, to show how far we've come in AI for content.

[Demonstration clip in English:] ‘Babe, I want to do a demonstration for the viewers on how realistic AI video is now.’ ‘Using us?’ ‘Exactly. Okay, look at my face. Skin, hair. I'm not real.’ ‘Do the hands. AI always messes up hands.’ ‘Five fingers, five fingers. Groundbreaking. Okay, hair.’ ‘Every single strand.’ ‘And you're still fake.’ ‘Rude. Jewelry, fingers, hair. And nothing's melting into anything.’

[The demonstration clip continues:] ‘Exactly. Okay, reflections?’ ‘Reflections are ridiculous now. Movement used to expose AI instantly.’ ‘Okay, watch this then.’ ‘Still fake.’ [inaudible] ‘Okay, catching things too?’ ‘Apparently.’ ‘You are just teasing them now.’ ‘Show off. Walking, touching things, clothes moving properly, bumping into your annoying friend. What about this?’ ‘AI. Even drinking.’ ‘Babe, you're not even thirsty. You don't exist.’ ‘That's actually quite rude.’

[The demonstration clip continues:] ‘I don't think they believe us. Let's prove it.’ ‘How the hell do you even learn this?’ ‘AI video bootcamp. AI is getting ridiculous.’ This is a good transition to our soon recording the next episode, which will be about creating authentic content.

Exactly. In this content era, AI has become so good at things it used to be bad at: maintaining consistency between scenes and keeping a character. Now we even have [unclear in the Swedish source: the proverb], the flow between the characters. It's flawless. I didn't find anything in this minute-long video I could point to…

[inaudible] …and say: ‘That's AI’, or that it feels very fake. I don't think you saw anything either? Nothing I noticed directly, no. It still points out most of the weaknesses we've seen in AI over the past few years and solved them so well. I don't know: are we now at the point where AI actually makes better content than people? I would say so.

We'll talk more about that in the next episode. Then we'll also discuss community: what it means and how to build one. Brilliant, I'm looking forward to it. Thank you so much for listening to Noll till hundra. We sit here every week, every Monday, and try to produce the best content we can that's relevant to you listening.

So we have something to talk about every Thursday. It means a lot to us if you follow the podcast and us on social media, like what we post when you see it and write a comment now and then on LinkedIn or down here on Spotify.

Or on YouTube. The whole episode is also there in full-length video. Just do it. If you like us, share it with a colleague too, so we know we can produce even more fun content.

[unclear in the Swedish source: gensant]. Keep doing good things completely without sponsors. The only way we can manage it is by having you with us. Thank you so much for listening. See you again next Thursday. Bye!

Bye!