A simple AI setup for a 10-person agency

Most 10-person agencies already pay for AI. Some people use ChatGPT. The developer uses Cursor. The founder lives in Claude. Everyone has a few good prompts saved somewhere.

And yet the results depend on who is asking. Two people get great work out of it. The rest paste the brand voice into a new chat. They get something generic and fix it by hand.

This post is a worked example of a simple setup that changes that. It is an example, not a case study. The agency is made up. The roles are the ones you probably have.

The agency in this example

Picture a small marketing agency with 10 people:

• The founder, who runs sales and the big clients.
• Two account managers, who own client relationships.
• Three content and design people.
• One developer, who builds landing pages and fixes sites.
• One operations person, who does invoices and scheduling.
• Two freelancers who come and go, project by project.

They work across Outlook, Teams, Linear, Figma and GitHub. Finance runs on an accounting tool. Each client has a brand guide, a tone of voice, and a few rules nobody writes down.

What goes wrong without a shared setup

You probably recognise most of this.

Everyone explains the client again. One agency founder described the routine on Reddit: "They open ChatGPT, copy-paste their brand voice, paste the email they're replying to, paste the SOP for how they handle this situation, then ask it to help." Multiply that by ten people and every client.

Everyone has their own version. An SEO agency owner wrote: "The problem with everyone having their own is that the knowledge bases and skills start contradicting each other, and nobody benefits from what anyone else figured out." The same owner added: "At one point we had the same rule copied into eight different skills, which is exactly how they drift apart."

Account managers build the same thing twice. One commenter described "six account managers each building a slightly different deck for the same client set."

The good stuff leaves with people. A freelancer builds a great reporting prompt. The project ends. The prompt goes with them.

Logins get passed around. The freelancer needs the client's Figma, so someone shares a password. Nobody remembers to change it.

Your people know how to do the work. What's missing is one shared place for it.

The setup, in four parts

The goal is simple. Everyone keeps the AI they like. Everyone's AI starts from the same playbooks, the same client facts and the right access. Here is how this agency would set it up in Homla. Each part maps to one room.

1. Playbooks for the jobs you repeat

Start with the five jobs the agency does every week. For this example:

1. The weekly client report.
2. A first draft of a blog post in the client's voice.
3. A reply to a client asking for changes.
4. A proposal for a new project.
5. A handover note when a freelancer finishes.

For each one, ask the person who does it best. Have them do it once in their AI, then type "save this as a skill". It goes into the team library, with folders and search.

Now anyone can ask their AI for "the weekly client report" by name. Fix it once and everyone gets the fix. Old versions are kept, so nothing is lost.

Put finance playbooks in a folder only operations and the founder can see. Client playbooks stay with the client team. More on this on the Skills page.

2. One memory for client details

Next, the facts. Each account manager saves what the AI keeps getting wrong. For example:

• "This client wants invoices on the 1st."
• "This client never uses the word cheap."
• "We agreed to two rounds of edits per page."

They just say "remember this" in their chat. It is saved for the whole team. When a fact changes, the new one replaces the old one. Every change is logged, so you can see where a fact came from.

Next time anyone asks about that client, their AI checks this memory first. The new hire gets the same answers as the account manager. See the Memory page for how it works.

3. Access that follows each person

Then the tools. The founder connects each one once, in the browser. A guide shows each step. Logins are stored encrypted and never shown again.

Then they set access by role. In this example:

• Account managers: Outlook and Teams in full. Linear to read and change. Accounting tool hidden.
• Content and design: Figma and Linear. Outlook look only.
• Developer: GitHub and Linear in full. Figma look only.
• Operations: the accounting tool and Outlook. Nothing else.
• Freelancers: only the Linear project and Figma file they need, with an end date.

When a freelancer leaves, you remove them. Their access ends within 30 seconds. No password to change, because they never had one. The Access page shows each control.

4. One agent for the boring job

Last, pick one job nobody wants. For an agency, it is often the inbox. Client emails arrive, someone sorts them, finds the context, and drafts a reply.

An agent can do the first pass. It reads new mail, looks up the client in memory, and drafts a reply. Before it sends anything, it waits for a yes. You can approve, edit, or reject each one.

Start with one agent and one job. Agents never take a seat. It can run on a teammate's computer, or on an always-on cloud box from $25 a month. See the Agents page.

What changes for the team

Nobody had to switch AI. The developer is still in Cursor. Marketing is still in ChatGPT. The founder is still in Claude.

What changes is where they start from:

• The new hire uses the agency's playbooks on day one.
• Nobody pastes the brand guide into a chat again.
• Freelancers get exactly what they need, until the job ends.
• When someone leaves, their playbooks stay with the business.
• The founder can see who can open which tool, on one page.

How to roll it out without a big project

Don't try to do all of it at once. A small agency can do this in stages.

1. Week one: connect two tools. Save three playbooks. Pick the best person for each.
2. Week two: ask the account managers to save five client facts each.
3. Week three: set access by role. Move freelancers off shared logins.
4. Week four: turn on one agent. Approve every change yourself for a while.

Keep a short list of what people still paste by hand. That list is your next set of playbooks and memories.

What it costs for 10 people

Homla's Team plan is $79 a month for up to 10 people. Extra seats are $8 a month each. It includes a setup call to connect your first tools. Your Claude or ChatGPT plans stay as they are. Full details on the pricing section.

If you have one person who already gets great results from AI, start there. Ask them to save one playbook this week. Then see who uses it.