As of 17 October 2026, rechecked on 18 October 2026, updated on 19 October 2026. Prices, features and availability change quickly in this area. This article will be reviewed again on 18 January 2027.

Many people who ask me about AI agents have the same problem at heart: there’s too much work for one person. Client projects come first, and outreach, newsletters or LinkedIn posts get pushed aside. Hiring someone is expensive, and as long as the work comes and goes, it feels too risky for many as well.

At the same time, a quick look at LinkedIn can leave you thinking that everyone else already has a whole AI team, with bots that win clients, write posts and sort inboxes. That leads to a hope I hear a lot: AI employees instead of staff.

One of these setups washed up in my feed in October: the “Grokbot GTM Team” by Tim Scheuer. In it, he describes 13 AI bots that handle lead generation and content for him. The app it all runs in starts at $30 a month, plus taxes. I took a closer look because it shows nicely what teams like this promise and where the catches are. But it’s just one of many.

So this article is about the question behind it: do you need a team like this, and how do you get started without getting yourself into trouble? Using the example, you’ll see how a bot team like this is built, what it costs and where it falls short. Then I’ll share what I’ve learned from my own AI team, and I’ll show you how to start small and safely, without 13 bots or any coding skills.

A quick word about me: I’m Saskia Teichmann. I’ve been working with WordPress since 2009 and with WooCommerce since 2012. Since December 2025, I’ve been working with my own AI team. One agent hands out the tasks to the others. That’s why, with setups like this, I mainly look at how the work is divided up and who decides in the end what goes out.

Video: What’s behind the Grokbot GTM Team

The video is loaded directly from the provider only after you click. This one-time choice is not stored.

Open at the provider instead

Spoken in German with English subtitles

What AI agent teams promise

Most people know AI as a chat window. You type in a question or an instruction, and the AI answers. That instruction is called a prompt.

An AI agent goes one step further. It gets a goal and works through the steps to get there on its own. It might open a website, create a file or save a draft in another program. To do that, it needs access to tools: a browser, files, or services such as your email inbox.

An agent team is several of these agents, each with its own role. Every agent gets a fixed job and its own instructions, and one of them hands out the work. Think of it as a small office where every desk is staffed by an AI program. What actually leaves the office should still be decided by a human.

The vendors promise a lot. xAI, for example, introduces its app “Grok Bot” as “your team of always-on agents” that keep working around the clock. In my experience, whether a team like this actually takes work off your plate depends mostly on who checks what the bots do and who maintains the whole thing. I’ll come back to that further down.

In Grok Bot, the agents are called “Bots”. The app launched as a beta on 11 August 2026, meaning a test version that isn’t finished yet. According to the documentation, each Bot works on a computer in Cursor’s cloud, with a browser, files and a terminal. The terminal is the text window where you type commands instead of clicking with the mouse.

An example from my feed: the Grokbot GTM Team

Tim Scheuer published the project on GitHub on 14 October 2026. GTM stands for “go-to-market”: everything that gets an offer to its customers, in other words sales and marketing.

The project lives in a repository, which is what a project folder on GitHub is called. There’s no program code in it. It contains 18 text files in Markdown format, three images and a license file. The texts describe roles, rules and workflows, and you copy them into Grok Bot, which turns them into Bots. The project doesn’t use an API, the kind of interface that lets programs talk to the AI directly. It’s published under the MIT License, a permissive license: you may use, change and share the texts as long as the license notice stays in place.

At the top sits a Bot called the Chief of Staff. It hands out the work and reports directly to the author. Below it are two departments, each with one lead and four specialists:

Outbound , with lead sourcing, copy for cold outreach messages, loading message sequences, and social media engagement. Outbound means you actively approach people who haven’t contacted you, for example with emails or LinkedIn messages to strangers. A lead is a potential customer.
Content , with one Bot each for LinkedIn, X and YouTube, plus one for graphics.

Two more Bots work on demand. One sorts the LinkedIn inbox and drafts replies; the other digs up older conversations and drafts follow-ups. On top of that there’s a group chat, the “Content Room”, where the content Bots and the Chief of Staff coordinate.

Other programs do the actual work

The Bots in Grok Bot plan and delegate; two other programs do the work itself. For copy, graphics and video, the quality in Grok Bot wasn’t good enough for the author, as he writes in the README, the project’s description page. So in his setup, the Chief of Staff passes every task on to Claude Code from Anthropic or Codex from OpenAI. Both are AI agents that run in the terminal, where they can write files and run programs.

Every handoff has four parts: the desired outcome, the material for it, the matching instructions (a so-called skill) and the limits. A limit might read: “Draft only, no publishing.” The result then comes back as a file or draft for review.

What the author cut again

I find one section of the README especially instructive: the one where the author describes what he cut again. He had set up ten routines , meaning tasks a Bot carries out on its own on a schedule. Today only one is left: a summary every weekday at 8:20 a.m. He calls his content Bot’s morning briefing “slop”, meaning useless mass output, and has paused it. His rule: a routine only earns its place if you read what it produces every time.

The original outbound Bot, which was supposed to do everything at once, is gone too. A lead and four specialists have taken its place.

Transparency note: According to the README, Tim Scheuer is a co-founder of OXYGEN, a sales and lead generation platform. The outbound Bots are tailored to that product, and the link to it in the README contains tracking parameters. That’s why I only link to the repository itself.

What it costs and what you need (as of 17 October 2026)

According to the pricing page , Grok Bot is included from the “Premium” plan at $30 a month. The higher plans, “Super” ($100) and “Ultra” ($200), include more usage. The “Free” and “Plus” plans don’t include Grok Bot. The listed prices don’t include applicable taxes or fees, such as VAT. If you pay annually , Premium costs $300 a year.

There are two things to keep in mind. Usage comes out of a shared allowance that refreshes weekly, and I couldn’t find the exact size of that allowance anywhere. And how the spending limit works depends on your plan. For the older Cursor plans, the help pagesays a limit is not a hard stop: a Bot that’s already working can finish past the monthly limit. For the current subscription, “The X Subscription”, which includes the Premium plan, the xAI documentation says something different: if you set a monthly limit for paid additional usage, it acts as a hard stop on those extra costs.

And the $30 only gets you the app. For the team as described in the repository, the author lists further requirements:

Claude Code or Codex with their own subscription, if work is to be handed off to them. Without these programs, the Bots do the tasks themselves.
A so-called second brain: a collection of your own texts, proof points and information about your audience for the AI to draw on.
A sales platform for contact lists, message sequences and an inbox.
For the handoff, a Mac that allows local commands, or a server that runs all the time.

What all of this adds up to depends on your setup. The Claude Code route alone requires a paid Claude plan (claude.com/pricing). I haven’t checked prices for Codex or sales platforms. Either way, $30 a month won’t cover rebuilding the team as it stands in the repository.

Where the catches are

Your data sits in the US. According to the security page , Grok Bot computers currently run in the United States. The page itself doesn’t mention the GDPR, the EU’s data protection regulation. It does, however, link to Cursor’s Data Processing Agreement, which explicitly refers to the GDPR. If you want to process customer data, read that agreement yourself first and make sure you’re allowed to. This is a pointer, not legal advice.

All Bots share one computer. They all work on the same computer in the cloud. Files, browser sessions and credentials on that computer are available to every Bot. The documentation explicitly says not to use separate Bots as a security boundary. So if one Bot can get into your inbox, technically they all can.

Rules in a prompt aren’t a lock. So in the project, the end of each Bot’s instructions lists which services it may use. These services are called connectors: links between the Bot and programs such as Gmail, Notion or Google Drive. The author himself writes that Grok Bot has no per-Bot switch for connectors. Once you connect Gmail, any Bot could use it. A line in the prompt is a request to the AI, not a lock.

Grok Bot does have real approvals, though. Before an action runs, you can allow it once, always allow it, or deny it. With “Ask first” rules, which you set up yourself, Grok Bot always stops matching actions and asks you. The documentation recommends keeping sending, publishing, purchasing and deleting behind an approval. It also points out that an approval only controls the proposed action and doesn’t undo work that’s already been done. One more detail matters for routines: if an approval request comes from work that started without you, such as a routine, it expires after about ten minutes, according to the documentation. The action then doesn’t run. So let your routines end with a draft that you read when you have time.

Here’s what the documentation lists as defaults: for commands on your own computer, the setting is “Ask every time”. Hardware security keys, small devices for secure sign-in, are switched on, but every use asks for your approval first. And according to the FAQ , sensitive actions can stop and wait for your approval. That assessment comes from a review model called “Car Review“.

The app is available in Germany, with one restriction. Grok Bot is listed in the German App Store, and the maker doesn’t name any country restriction for the app. The only thing blocked in the EU is tagging the Bot on X (docs.x.ai). I haven’t tested myself whether signing up and using it from Germany works smoothly.

Legal (a note, not legal advice): Section 8.2 of LinkedIn’s User Agreement prohibits bots and other unauthorized automated methods, including for sending messages. In Germany, under Section 7(2) No. 2 of the Act against Unfair Competition (UWG), advertising by email without the recipient’s prior express consent counts as an unreasonable nuisance. Subsection 3 provides an exception for existing customers, tied to four conditions. That is German law; other countries have their own rules. This is why you deliberately won’t find instructions for automated cold outreach here. If you’re planning something like that, get legal advice.

What I’ve learned from my own AI team

You don’t have to rebuild the GTM Team to learn something from it. Some of its ideas match what I’ve experienced with my own agent team, which I’ve written about in a behind-the-scenes series . Here are the points that also work for a small business without its own tech department.

Write the role down once instead of explaining it every time

In the GTM Team, every Bot has exactly one job and a fixed description to go with it. That’s how I work, too. You write the role down carefully once and don’t have to explain from scratch every time what you want. That works with a single assistant as well.

For me, every role has four parts: who the agent is, how it works, what it’s allowed to do and what it should remember long term. You can see what that looks like in “For me, an AI agent consists of four files”. There’s a practical reason for the split: in one long prompt, every correction risks changing something else you didn’t mean to touch. To start with, four headings in your assistant’s instructions are enough.

Being able to doesn’t mean being allowed to

The first rule in the repository says, roughly, that nothing gets sent, posted, bought or deleted without the author’s explicit yes to that exact item. Drafts are the default output. That attitude is worth adopting.

For each of my roles, it’s written down what it may do on its own and where I decide. My writing role, for example, may create drafts in WordPress, but it may not publish them. That still applies when the task reaches it through a different tool. That doesn’t change when another agent passes the task on to it. Anything that costs money, deletes something or goes live stays with me. As described above, though, this attitude needs real technical safeguards; a rule in the prompt isn’t enough on its own.

AI only remembers what’s written down

An assistant doesn’t reliably remember what mattered last week just because the app has a memory feature. That’s why every session with my agents ends with a short entry: what was the task, what got done, what was decided and what’s still open? When I correct something, it also becomes a fixed rule that gets read the next time. One example: a German headline sounded as if it had been translated from English. Since then, the rule is that a headline has to sound the way I’d actually say it to a colleague.

Anything nobody wrote down stays marked as a gap. I don’t let the AI piece it together after the fact. There’s more on this in “How My AI Agents Remember”. You can do the same on a small scale with a chat assistant: at the end, ask it for a short note covering those four points and save it in a document you give it again next time.

Start small and factor in the upkeep

The author of the GTM Team himself went from ten routines down to one. The Grok Bot documentation also advises starting with tasks that only read and produce drafts. And even a test run isn’t a dry run there: according to the documentation, it does real work, so it can open websites, change files and use connected services. Only test with examples where nothing can go wrong.

From my own experience, there’s one more thing: an AI team also creates work of its own. In February 2026, I set up a team of fifteen specialized agents with an orchestrator, an agent that hands out the work. After updates to the tool, it sometimes took me half a workday to get everything running smoothly again. I was busier fixing things than doing the work the team was actually there for. If you already have too much on your plate, factor that upkeep in. So start with one task you have every week anyway, and look closely at the results for a few weeks.

Your instructions belong to you, not the app

At the end of March 2026, after just four weeks, I parted ways with the tool my team had been running in since February. The only things I could take with me were the ones stored as plain text files: the role descriptions, the collected knowledge and the daily notes. Everything that only existed inside the tool, I had to rebuild. The full story is in “OpenClaw Discontinued: What Remains of the AI Team”.

So my advice is: write the role, the rules and what you’ve learned into a document you own, and copy it into the tool from there. That way you can switch later, say from ChatGPT to Claude, without starting from zero. How I now do this across several tools is described in “My AI agents do not belong to their tool”.

How to start small and safely

To get started, you need neither Grok Bot nor a terminal. A single assistant in Claude or ChatGPT is enough. Both offer what they call projects. In a project, you store instructions once, and they then apply to every chat inside it.

That’s where you define the role: what the assistant writes for you, in what tone, which facts it may use, and that it only delivers drafts. You copy the finished text yourself and publish or send it yourself. Don’t connect any services the assistant could use to send or post. It’s safest when the assistant technically can’t send anything at all. If your app has agent or browser features, though, those might be able to act even without a connected service. So check in the settings what the app is allowed to do on its own.

If that goes well for a few weeks, you can think about whether a second role makes sense. The Grok Bot documentation offers a good rule of thumb: a separate role is only worth it when the work has its own goal, needs different tools and sources, or has its own approval rules. If it’s just a new task in the same area, a written set of instructions for the assistant you already have is enough. Grok Bot calls this a skill. And if you want to go further: I’m currently building the Agents Brain Starter Kit, a template that gives your assistant a fixed role, a memory and approval rules. It’s coming soon.

The checklist to take with you

So you don’t have to keep all of this in your head, I’ve put together a free checklist: “Before you put an AI agent team to work”. It has 16 points on costs, data, approvals, legal questions and starting size. It also comes with a template for a content assistant with approval that you can copy straight into a Claude or ChatGPT project.

To get it, sign up for my emails about new isla Studio posts. Once you confirm your address, I’ll send you the checklist. You can unsubscribe at any time using the link in every email.

Free checklist from isla Studio

Request the AI agent team checklist

16 checkpoints on costs, data, approvals, legal issues and starting size, plus a template for your first content assistant with approval. To get access, sign up for a short email with a link whenever I publish a new isla Studio post. Once you confirm your address, I'll send you the link and password.

I use your details to manage your subscription and deliver the checklist. Your subscription starts only after you click the confirmation link; you can end it using the link in any email. I delete unconfirmed requests after 48 hours and the delivery record after 30 days. I keep your confirmed subscription until you unsubscribe or ask me to delete it. Learn more in my Privacy Policy.

This article is not legal advice.

Sources

Accessed on 17 October 2026, rechecked on 18 October 2026; further sources added on 19 October 2026.

AI employees instead of staff? What an AI agent team can do and where it falls short

What AI agent teams promise