Introduction

All right. Good afternoon. Good evening.

My name is Sjoerd de Grij. My name is... Oh, sorry.

I'm the co-founder and CEO of Typetone.

We are an AI startup in Amsterdam, and we started out building a tool to automate content creation. Well, I saw one of the use cases of content creation and marketing automation there on the board. We were there before ChatGPT building this tool.

We take some pride in it. Doesn't mean a lot, but still, we were there before the AI hype.

And now we build AI digital workers or AI agents. We built Sarah, maybe you've heard of her, maybe not, but I will tell you more about it.

The Future of AI and Work

And today I will tell you more about AI agents and the future of work and how LLMs and basically prompting is already an outdated term. And I want to start with a small number, 0.5%. Who has any clue what this means? Who has an idea what this number can represent?

This is actually the predicted productivity increase that MIT, the MIT economics department, last year predicted of AI. Super, super low.

We, of course, everybody in this room probably thinks this is way higher, but there's a reason they think this. And funnily enough, we still see quite low productivity increases now because of AI.

And I want to let you guys leave after this talk with this question, how much is your business truly growing or saving because of AI? Because this is what it's all about.

And what is currently the single biggest use case of AI? probably this one. What should I eat tonight? No, I'm just kidding.

Current Limitations of AI Productivity

This is probably not the biggest use case, but this is part of the problem. This is why productivity increases are not that big, because what we get right now is tools, and tools have a big problem, and that is switching costs, and I will tell you a little bit more what that actually means.

The moment that you give a worker or an employee a tool, that means that every time they have to switch and have to Basically, invest a little bit of cognitive time to switch from cost. Basically, you prompt, you get an output. It's not exactly what you want. You have to switch again, go to the next task.

Versus doing the task all by yourself isn't giving us the productivity increase that we actually predicted. 1This cognitive efficiency diminishes as the brain needs time to refocus each time. So switching costs right now prevent us through AI productivity gains.

And this is an example.

For example, I'm a worker and I have a task. What you can see here, for example, I'm writing an SEO article. The task is to get this article in as little time as possible.

You start with researching, then you write it, you get feedback, you have probably a meeting, you get more feedback, then you rewrite, and then finally you have your end product.

And this is what AI is doing. You still have to research, or maybe AI can do that nowadays. Then you generate the article, you fine-tune it, you finalize the article, and in the end you save some time, but not 95%.

It's probably more like 30% because you had to switch all the time.

Case Study: Samantha's Agency

And to give you a little bit of an example, this is Samantha. She has an agency. She has eight full-time employees. She does half a million in revenue.

And she needs to grow her business, but she's not findable online, so she has to do content marketing. Her options are writing or prompting SEO articles with a tool like ChatGPT or Cloud or whatever she wants. She can hire a freelancer or hire a full-time employee where she doesn't have time for.

So is Samantha Health with this master ChatGPT prompting cheat sheet? Well, probably not because that's still a tool. This is not solving her complete problem and that's what I want to focus on.

Moving from Co-pilot to Autopilot

The true productivity game is an autopilot, not co-pilot. And that means stitching all of these steps together. And that's what we're working towards. So not a co-pilot, but actually an AI employee who can talk to you, who can automate full workflows.

So where we're going from now is a team without AI with output of one unit of work. And then now we have a team with a co-pilot outputting one and a half or two units of work. But what if you have a team plus an AI agent who can do infinitive amounts of work limited only by compute?

So we are moving from AI tooling and prompting towards agents that can handle entire processes. And these agents don't perform isolated tasks, no, they do complete workflows. So this is where we're going.

Not putting extra headcounts to your business to grow your business, no, actually putting more compute to one agent performing every task of your business. And funnily enough, the biggest advantage and biggest gain here is actually to be gotten by SMEs, which is still 50% of GDP because headcount matters the most here. And this is finally the time where SMEs have a fighting chance against enterprises because they don't have to increase headcount anymore. They can just hire Sarah or whatever AI agents can do the work.

Maybe you guys know the big venture capital firm NFX. They predict that if you are in SaaS, probably quite a big share of you guys work in software, that the big shift right now is not towards AI software and that's it. No, this is actually the shift.

B2B software as a service is becoming part of the labor force where basically AI will be a new workforce. So, current roles that we see being automated are SDRs, for example, AI marketeers, customer service, and the list will grow and grow the more tasks will be automated.

Funnily enough, these are tasks that are now done by ChatGPT. After this, you can probably see finance as well on this list.

So I was actually announced as the technical talk. I will leave the technical talk for the questions.

Understanding AI Agents

But how does this actually work? What is an agent?

Well, in a nutshell, it all comes down to this. An AI agent is basically a chain of LLMs or self-prompting LLMs with a short-term memory, basically to remember what kind of task you gave it.

Then there is the knowledge. You could say long-term memory. You can basically say this is the knowledge of your business, your tone of voice, your USPs, your competitors, your target audience, maybe even your company culture, while the list goes on.

So you have the short-term memory for the task and the long-term memory to actually letting this AI agent feel like an employee. And then there's tools and skills, because the output is not just plain text. The output should be actions, because what do you want the content marketer to do?

Not only to write the text, but probably also to publish this. So the tooling would be the ability to call APIs or talk to software to actually perform actions.

So that means that an AI employee like Sarah doesn't only know about your company and create text, but she can also schedule something in your calendar, publish the text, and also comment underneath it. For example, talking to the LinkedIn API.

Then there's task orientation. There's a sequence of tasks to be done in work. And this is where the magic happens in the AI agent, is that the LLMs actually talk to each other and give each other feedback in terms of what's the next step in the sequence of these actions should be.

And to make this very plain, probably research, then write, then integrate with LinkedIn to create a post, publish it, and then learn from that in terms of engagement for the next post, that is this task orientation.

And then, of course, there's accessibility, collaboration, and planning, and this is that these AI agents work best if they actually feel like humans, funnily enough. I won't go into depth there, but it shouldn't feel like a tool. It should be an employee.

So, I hope you guys are not scared yet. because I will actually show you how this works in practice.

I have a demo link here, so I'll quickly open that one. I just heard from our developers, we have a small outage, so I will just show you this.

So, how does it work?

Onboarding AI Agents

1Sarah is onboarding your company, and that means that she starts with scraping your website, like you would do with an employee. This person checks your website, checks your info base, checks your notion, checks your confluence. Then she integrates with all the tools of your company, so that would mean not only Confluence and Notion, but also in this regard, because Sarah is a content marketeer, Facebook, LinkedIn, Instagram, and Slack.

You give her your goals, so that means she is going to upload her own images because she scraped them from her website, and she integrates with your Instagram, Facebook, and LinkedIn user account, and she already learned your tone of voice by scraping your website. Well, this is the moment where she can do this 100% autonomously, but we still make the decision of showing this to our customer because our customer still wants to see what Sarah comes up with.

This is the same like collaborating with an employee. She's probably not doing everything autonomously. She will actually collaborate with you and talk to you, send you a Slack message. That's what Sarah does as well.

She scraped the web and found industry reports, found everything about your competitors. She repurposed your own content. She made a whole list.

And if you want to give feedback, that's of course possible. So if you want to say, well, Sarah, there's this one competitor and I really want to repurpose their content or I want the edge to be AI this time, please make some content on that. Well, you can give that feedback. You can also do this in Slack and WhatsApp.

It's an employee in the end. And that's the moment where she starts creating.

And we now chose to give it a user interface. So that means that you can visually see what she's generating at the moment.

So as you can see, she's now generating 35 posts in a couple of minutes. And we also give it a visual visualization, a content calendar to actually show what she's planning. And she's also autonomously publishing this.

How AI Agents Work in the Background

And actually the part where you probably came for is what happens on the background here. And what happens in the background here is that about 36 different LLMs, it's not all different LLMs, but you can think of clods, fine-tuned models, fine-tuned LAMA, think of Sora, not Sora, what is it? flux as well.

And all of these models talk together by letting Sarah make a complete content strategy, start executing on the content strategy, and then making all the content. And then, of course, she integrates with the APIs as well. So that is what Serra does, and that is what we believe the future of work, especially for SMEs, will go to.

Again, do you have to worry? No, I don't think so, because these are companies who couldn't pay for a human content marketeer, and now they actually have the hands available instead of a tool. And that is, I think, super cool.

Conclusion

So that was it, and now it's time for questions. Thank you.