How to build an AI OS for your business

Start with one job, not the whole company

An AI OS grows one job at a time. The first job decides whether the team trusts the rest, so pick one that is easy to check.

• It repeats. Weekly reports, inbox triage, client follow-ups, CRM updates after calls.
• Its inputs already live in your tools. Mail, calendar, CRM, documents, a finance system.
• A person checks the result today. That person becomes the approver, so nothing about who signs off changes.

Write the job down in two sentences: what comes in, and what a good result looks like. That is the brief for every step below.

The seven steps

1. 1

   Write down the context

   What the company does, who decides what, and the rules for this job: pricing limits, tone for clients, what counts as urgent. In Kylon this goes into Workspace knowledge, which every Agent reads, and into the Agent's own instructions.

2. 2

   Connect the data the job touches

   Add the Connections for the mailbox, calendar, CRM or documents the job reads. Grant each one only to the Agents and people who need it. Data with no home yet, such as a tracker kept in a spreadsheet, can become a Database App.

3. 3

   Create the Agent and give it the job

   Give the Agent a name, a role and the skills for the job. If the work already runs in Claude Code or Codex on someone's computer, connect that one instead with Invite my agent, so the setup you built keeps working. See Claude Code AI OS for teams.

4. 4

   Put it in the Room where the work is discussed

   A new Agent is private and in no Room. Add it to the Room where the team already talks about this job, so requests and results land where people read.

5. 5

   Keep a person in front of anything that leaves the Room

   Emails, record changes and anything a client will see come back as a draft with Approve and Edit. The person who checks this work today is the one who approves it.

6. 6

   Run it on a schedule or a trigger

   Once the job works when someone asks, make it an Automation: every Monday at 9, when a form comes in, when a message arrives. The result still lands in the same Room and still waits for approval.

7. 7

   Build the App when the same table keeps coming up

   When people keep asking for the same list or dashboard, ask an Agent to build it as an App. It opens next to the conversation and reads the same data the Agents write.

Review it after 30 days

Read the threads for the job and answer four questions. The thread already has the request, the draft, the edits and the approval, so this is reading, not a survey.

• How many drafts were approved as written, and how many were edited first.
• Which requests the Agent could not finish, and what context or access was missing.
• Who outside the original team started asking it for work.
• Which step a person still does by hand that the Agent could prepare.

Fix the context or the access behind each miss, then pick the next job. The second job reuses the context, Connections and Room from the first, which is the point of an operating system.

Watch one built live

Liam Ottley sets up his own business on Kylon in his full course livestream: onboarding, connecting Notion, Slack and Google Calendar, approving the first Agents, and later Agents handing work to each other. Kylon sponsors Liam's videos about Kylon.