Welcome, my name is Lena and I will talk today about how to become more productive using AI. My talk is called Managing AI Employees from Notion and it's not going to be a product that we sell, it's just something that we build for our company and we use extensively ourselves, so I hope everyone finds some inspiration here for the work that they do themselves.

So before we start just a a quick round of questions. Who here ever built CloudOS, second brain, you know, the folder structure for your company business?

Okay, maybe six people. Cool.

And who has built workflow automations? And then I take men, or make Cloud scheduled tasks, so like multi -step automations.

Okay, cool. More people.

And just out of curiosity, who can code? I

About the Speaker and the Project

slide about me. I think it will give you an explanation of why I'm talking about what I'm talking.

My name is Lena Shakurova. I'm an AI advisor and CEO of ParseLabs. We are a conversational AI and AI automation agency.

Worked with a lot of companies since 2018. I have a research degree in AI. That's good to know, meaning that I'm quite technical.

And I also build automations with my team for companies so we understand how the processes work, how to optimize them and then we develop the automations.

Plus I have a lot of students over the last couple of years because I give AI literacy workshops on a very similar topic. And all of this means that I spend a lot of time analyzing company processes, optimizing and automating them.

And we typically do this kind of work for our clients and for students but this summer with our team at Parse Labs we decided that hey let's try to do the same thing that we actually teach and do for others for ourselves as well so we sat down and we

analyzed our own processes and we built a bunch of automations and the AI operations layer for our company and that's the kind of project that I want to share with you today.

So what we're going to have today is I'm going to have a short intro about the thing that I'm going to be showing you, then we're going to have a practical demo where I'm just going to freestyle and show you how the thing works and then I'm going to explain a bit the architecture behind so if any of you is curious you can try to build something similar for yourself.

Then I'm going to share a couple of best practices that I learned about managing AI employees and building AI power team, and then we'll have questions. So let's start.

The Work AI Can Help With

I'll start with just giving you some overview of the kind of work that I need to do as a founder of an AI agency. A lot of research, and research can be looking for potential leads, looking for conferences conferences, and events where I need to present.

This is how I found MindStone event. By the way, this is why I came here as a speaker. Finding venues to host events because we also host our own events.

I get invited to podcasts, I give podcasts, so I research different gear that I need to buy to do this from home. I also teach online courses, so I need to make an overview of the market. So things like that. that.

Another recurring task that we do very often is lead generation. So here I was looking for a learning and development specialist in the Netherlands.

Another thing I do very often is writing LinkedIn posts. I need to write announcements of events where I'm speaking, anything we do at the company, tips, tricks, stuff like that. Then I also create Luma descriptions, I write SEO -optimized articles, so a lot of writing.

Another category of things we do frequently is creating carousels, in our case in Canva, so stuff that we post again on LinkedIn, and I have a YouTube channel as well, so creating YouTube covers is also something that we do frequently, or Luma covers, so anything related to images and stuff like that. So that's what the tool that I'll show you now helps me do.

Why Chat-Based AI Workflows Fall Short

And before how I used to do those kind of tasks. I'm an active user of cloud and cloud code and mostly you create a chat and then you do your task it creates I don't know an image maybe then you save it and then you start another chat and then you have a lot of chats in parallel and then you forget what you were working on and what finished and what didn't finish.

So these were the problems that I was trying to tackle.

No, that's not the problem. Okay, so we'll start from the second one. Open multiple chats in parallel. So that was quite annoying because I didn't know what was happening. I was constantly context switching. Then I would forget which task I was working on and if it finished at all and if I completed it or not.

And then all this chat is really hard to search through. Context switching pitching was hard for me, I would miss deliverables, like I knew that I at some point sat down and did this research of all the leads that I need to do, but I forgot to save it and now it's somewhere in the chat and I didn't know where it is.

Then I would lose track of which tasks were done and I was also losing track of which window was responsible for which task. So I found this all quite annoying.

Managing AI Employees in Notion

So what we did instead with my team this summer and we've been testing it for three months is basically connecting Notion to a bunch of AI agents so we can track tasks in Notion and delegate our recurring tasks to AI employees. And that's what I'm going to show you now.

So the way it works is and it works both from my laptop and from my phone which is quite convenient.

Creating and Delegating a Task

So I would go here and I I would create a task that I need to do now.

And all of those tasks, we already have written SOPs for them, because we've done them many times. So SOP is standard operation procedure for those who doesn't know.

So we know that for LinkedIn task, we always need to read through my past posts. This is the tone of voice. We need to maybe research the doing search engine optimization research to see which keywords rank. Then we write an article, then we review it. reviewed.

So I would create a task and I would say write linkedin post for mindstone and typically I would also write a comment here but I just want to show you the quick setup. So I would describe what the task is about so I would put the voice recording and I would say write that I'm coming here it will be cool please come join here is the link to join and then all I need to do is assign it to manager.

So here you see manager, planner, researcher, writer, reviewer, and then the team. I only need to assign the tasks to manager. Manager is AI manager.

And then you'll see now it's in the backlog, but as we will be talking and looking at other tasks, it will move in progress. Every 60 seconds the AI manager is checking which tasks do we have. So we'll have a look at the one that I already created just to save us time.

so I wrote here this is the link to the event and then the AI manager picked it up and it said hey I picked it up and this was important for me because I needed to know if AI stuff is actually working and it's not broken and I don't need to restart the server so I needed them to tell me I'm working on it everything is fine so what happens then is that manager picks it up and then it transfer it to planner then planner transfers it to OUC now it's in progress progress, and it said it's working on it.

So planner transferred it to researcher, writer, then reviewer, and then back to manager. And then when the manager picks it up again, then the manager needs to write the whole description of the ticket, which I personally am really bad at, I'm not as organised as my team members or AI in this case. So the AI would do their best to come up with a goal, context, tools, acceptance criteria, and the plan with a description of which of the sub AI agents will be picking up what so planner researcher then the writer drafts a LinkedIn post and Lena needs to approve so we also have the approved gates for different processes

so we have for example for the presentations we would say in our SOPs that you always need to approve the research plan the outline and then you you send the final Canva link, for example.

And then, because I am bad at providing all the context that it needs, it knows how to interview me. So it would ask me which angle is your talk about? Is there a CTA? Do you need to mention two other tickets and stuff like that?

So I would give it an answer, and then I would assign it again to manager, and then it would pick it up and go again through the SOPs. And that's the whole idea behind it actually. So if

Reviewing Deliverables and Collaborating

we look at the tasks that we had here before, let's look at some interesting ones.

Well I was writing a post about, oh that was a long post, about the hackathon. So maybe what's interesting here is that it creates the final deliverable. So I I always know where my documents are. So here is the post about the hackathon that I attended. And I can share this post with the team.

I can myself also leave a comment here on any of the text that I can edit that I want to give feedback on, although I don't do this for LinkedIn posts usually. It's just easier to edit it myself. But I can also leave comments here later and say, go and edit this. Go create another version.

Now I want an image for this post. and it will keep creating deliverables which because it's a human readable format it's just an ocean document that people can also very easily write which is not the same in cloud and cloud you cannot edit the text directly so very easily especially if you are having cloud code user so that makes it easy also this board is shared with my team members so I'm no longer the bottleneck

and what I can do and what I often do is I go for a walk and then while I'm walking I transcribe my voice and I create like five research tasks I know that they have an event coming up I need to prepare a client proposal so I just do this then it interviews me while I'm walking to the supermarket with this kind of questions then I answer the questions then I go for a meeting with a friend or a business meeting and then when the meeting is done I basically already have the initial versions of the stuff that I need which I can forward to my human team and they can pick it up and they can either continue working working with the AI employees, or they can pick it up and do their own work in real life.

And as you can see, we've been using it quite heavily. We started, I think, in May. And the green is the stuff that we've done, so there are 44 tasks done. A lot of them are blocked, blocked meaning that it's waiting for my input. A lot of them are in review, meaning that I'm really happy with the output of AI, but But unfortunately, I cannot make calls or send connections on LinkedIn, so I still have to do my part. So there is that.

And yes, I think that would be it for the demo. Let me see if there was anything else interesting.

Tracking Usage and Costs

Well, maybe I want to also highlight that we are also measuring the cost. We are using the cloud subscription for this. So actually, the cost is not the real cost. it's heavily subsidized but it gives us an impression of how research -heavy the task itself was.

Now we can go back to the presentation and if you'll have questions about this later we can go back to the demo.

Architecture of the AI Operations Layer

Just to show you the architecture once again so it's clear. We have five sub -agents and 24 skills and skills also represent the SOPs and SOPs are based on how we were doing things manually then we tested it a lot with this agency RM that's how the project is called and then when we were happy with this then we packaged it into skills so

the workflow is like this I delegated to manager manager makes a plan and then it delegates it to planner planner makes a further plan with a research and the MCP tools we have fire crawl connected heavily connected and other web search MCPs. Then the researcher actually does all of this research.

We also connect with VidIQ which is the MCP that can give you statistics about YouTube that's useful in case we are doing YouTube videos or data for SEO so this is when we are writing SEO articles. So that's all the stuff that researcher is doing.

Maybe researcher gets stuck and then it can go back to me and ask me for extra context or if it doesn't get stuck then it goes to the writer and then the reviewer and reviewer checks with the notion dog if all the acceptance criteria were met and if they weren't then they keep talking to themselves and if it was all fine then they can hand it over to me and then it goes on until we

Automation Is Delegation

are happy with the task and one thing that I've been learning as I've been working a lot on optimising processes and automating processes, is that automation is essentially delegation. And the same principles that apply to how you delegate applies to automating and delegating stuff to AI agents.

And when I was building this project together with my team, I was asking myself a lot, how does the new team constellation look like when you no longer work just with humans but with AI agents? and what is needed from me if I want to become as little of a bottleneck as possible and if you are a founder like me you know how painful it is and everything just depends on you and you're just droning in the tasks and you do it to yourself yourself so I defined three things that I need

to give to my AI employees and everything else they can handle autonomously so either I can can provide context because it was missing, or I give feedback on the stuff that they did, or I make decision. Everything else they can figure out themselves.

They can figure out the plan. They can figure out the design. I don't care. They do it. I only need to do these three things. So we design it in such a way that it only comes back to me when it needs one of those three things.

Principles for Building an AI Power Team

And here is just a summary of things that I learned when building this project, which I think work well for building your AI power team.

Connect AI to Your Company Operating System

First is connecting it to company OS. So this project is essentially a Python code that works on cloud SDK and it lives inside a folder with a Python script inside our company OS. So it already has a lot of context about our company, all the references, images and stuff like that. So we are not starting from scratch and the company OS is

also growing the more tasks we do because when I give feedback when I give the tasks I also give a lot of context about our company and then once in a while we can run the cloud on top of all of this to tell us okay what can we learn from how Lena or any of our team members were giving feedback can we make it part can we make it stable part of our company OS another principle that we build in inside of this is continuous learning and process improvement so this

I already just mentioned basically we are following the SOPs but we also create the new SOPs as I start noticing that I start delegating new forms of tasks that we didn't build the SOP yet for but I've been doing it repeatedly to this agency CRM It means that we need to package it into a standard operating procedure

Make Work Visible and Human-Readable

Another thing that I found helpful is that deliverables are human readable So, this is why we selected Notion instead of just, you know, you could also potentially share cloud chats with each other, I guess, with a share link, but we found this less comfortable for teamwork if you want multiple people to work inside your company OS and not just you. Yeah, and agency RM, so this project is also accessed by human employees, so it makes it not just me and AI, but me, AI, and my team. so I finally stopped being the bottleneck.

All the tasks are tracked so that's another thing that I really like it's just a standard Kanban board we are all familiar with this done not done reviewed and I know when my AI finished and when I also finished my part of the work plus all deliverables are also in one place and are easily searchable I can delete the ones that I don't need and we built AI loop in such a way that it only stores as deliverables things that actually need to be deliverables. Any intermediate research that it's doing is never stored on Notion because I don't need to see it. So I only see what I need to see. And it works from mobile, which I also really love.

Key Takeaways

Yeah, so that's kind of it, just to sum up. Automation is delegation, so similar principles apply.

If you are thinking of building something like that, think about building separate specialized agents rather than having one agent to rule them all and design for human in the loop because human to AI collaboration works better than just AI with AI and you want to keep your original input your feedback to amplify your work if

you want to learn more about this this is my LinkedIn I give trainings at companies on this kind of topics we run an AI automation agency where we do stuff like that and I also share in general a lot of knowledge for free on my LinkedIn through posts, YouTube videos or talks, so feel free to connect and happy to answer any questions that you have. Thank you.