If you run a business and employ people, you've heard this one: AI agents are going to take our jobs.
I don't think that's something that just happens to you. I think it's a choice. You can keep the people you have and use AI to get a lot more out of the business, or you can cut people and use AI to try and keep the output where it is.
I started circle.cloud in 2015. It has over 100 people now, and I'm not sacking any of them to make room for AI. I'm keeping them and building AI employees to work alongside them.
The harder question is how you put AI agents into a company that size without making a mess of it. That's what the video is about.
In 1980, Steve Jobs called the computer a bicycle for the mind. I see AI the same way, as a tool. Used properly, it takes a business a lot further. Used badly, it produces slop, and we've had a few examples of that at circle.cloud.
If you're not careful, it gets into the social posts, the contracts, the privacy policy and even the software. I'm sure plenty of companies are being built right now on buggy AI-written code and AI-generated contracts. There'll be legal cases. I wouldn't want to be a lawyer right now.
It's coming from customers too. Someone I was talking to recently told me long-winded, clearly AI-written complaints have shot up, and each one takes more than twice as long to deal with. At circle.cloud, complaints now arrive about double the length they used to be.
What gets me is what happens next. You go back to the customer with a solution based on what they wrote, and they don't know what they complained about. They know the outcome they want, but all they typed into Claude or ChatGPT was "get me a discount", and they never read what came out. They don't mean any harm. But somebody has to sift through it, and that cost ends up in everyone's prices.
Sloppy output comes from sloppy input: give AI a vague, generic instruction and you get something vague and generic back.
So if you're putting AI employees into a business, the first question is what you're feeding them.
Think about a new person joining your company. You don't hand them a Post-it note with four lines on it. You give them the handbook, introduce them to the team and get them into the intranet or Notion. Even then, it's a couple of months before they really understand how the business works.
An AI agent doesn't get a couple of months. It has to know the business from day one.
There are two parts to that. One is the base knowledge of the business, the handbook and the intranet: call it the company brain. The other is the instructions for what this particular agent is there to do. Keep them separate. A marketing agent, a sales agent and an accounts agent get different instructions but share the same brain, so you're not explaining the business again for every new one.
There's a catch with a handbook. The moment a new client comes on board or someone posts an announcement in Slack, it's out of date. If your company brain is a PDF you uploaded to Claude, you'd have to redo it every time something changed.
So the brain has to be live. One way is to have it call every system you use each time someone asks a question. That's the wrong approach: you run into rate limits, and the answers take ages, if they come back at all.
The better way is one initial run that reads everything from your systems (Google Drive, Slack, Teams, Gmail, whatever you use) and indexes it into a database so it can be searched quickly. After that, it only looks for what's new and copies that across. Ours is connected to Slack and Google today, with the CRM next, and a new Slack message is in the brain within a couple of minutes.
We didn't build that from scratch. For the brain we use GBrain, the open-source system by Garry Tan, who runs Y Combinator. We picked it because it's well built and it's been leak tested. For the agents we use Mastra, an open-source framework for building AI agents in TypeScript. We forked both and changed them to suit us, and that's what we're putting into circle.cloud.
When I ask ours something, it knows it's me from my Slack account, and it only answers with information I can already see. If the answer is in a private Slack channel I'm in, I get it. Someone who isn't in that channel can ask the exact same question and they won't.
This is the part I think a lot of businesses are getting wrong. Once a brain has read everything in your business, private channels included, the answer has to depend on who's asking. A sales rep asking about finance shouldn't get anything they couldn't already open themselves.
Getting information you don't already have access to is a leak. Before we connected ours to any real data, we tried every way we could think of to pull out something that shouldn't come out, and found zero leaks. It gets more important the bigger you are and the more your departments keep to themselves, so it's worth setting up properly from the start.
Next, where does your team talk to these agents? I don't think the answer is another app. They should live wherever your team already talks to each other. If that's Slack, put them in Slack. If it's Microsoft Teams, put them in Teams. If it's Olatti, the business phone system we've spent the last five years building, put them in Olatti.
One agent that does everything, or several? Some products take the single-agent route. Viktor is one: you add it to Slack as an app and mention it whenever you need something. Claude Tag, Anthropic's Claude in Slack, works the same way. It's simple, and I can see the appeal.
My bet goes the other way, because that isn't how a company works. When you want to talk about marketing, you talk to the marketing exec or the CMO, not the finance director as well. So I'd split agents by function, and which functions depends on the business: a plumbing company won't need the same ones as a software company.
Each agent gets its own account on your messaging app, and you talk to it like a colleague. Say Sam is your cold email outreach agent. You start a conversation with Sam and set up a campaign. Then you want new creatives for your Instagram ads, so you start a separate conversation with Olivia. Put everything through one agent and that one conversation fills up with every topic at once.
For a personal assistant, one agent is probably fine. For a business, I think separate agents win on control and simplicity. I could be wrong, and we'll see.
One thing I didn't get to in the video: who approves what the agents do. Right now, a person has to sign off anything that goes outside the business, and any code an agent writes before it gets merged, and for now I'm happy with that.
The more confidence you have in an agent, the more it should do by itself, and I'm still working out how we hand that over. The simplest way might be to let the agent ask: do you want me to check with you every time? You say yes or no, the same way a new starter gets more room once they've proven themselves.
Keep your people and give them AI agents. Give the agents the business once, in a brain that keeps itself up to date. Make sure it knows who's asking, so nothing leaks. Put the agents where your team already talks, give every job its own agent, and keep a person signing off until each one has earned the trust.
None of this replaces anyone at circle.cloud. It's the same team, getting a lot more done.
Join 480+ founders and operators getting my learnings from building and running my companies, straight to your inbox.
