The Solo Agency Ops Playbook: Let AI Agents Run Your Back Office

Seven clients at $1,500 a month is $10,500. What sinks a one-person agency is not the marketing. It is the back office: logins, invoices, status updates and follow-ups. We mapped six months of ours, found an agent job in every kind of work, and turned it into 12 jobs you can run from a laptop anywhere.

The $10K math, and where it breaks

Here is the math behind a $10K month:

Illustrative math, not a promise. We use a conservative $1,500 retainer; your results depend on your effort, skill and market. Try your own numbers in the AI Builder earnings calculator.

Getting to seven clients is a sales problem. Keeping seven clients while you travel is an operations problem. Every client brings accounts to get into, payments to match, reports to write and questions to answer. As clients add up, that work fills the week, and one login request missed while you are on a plane can cost you a client’s trust.

Agents change that. They do not sell for you and they do not replace your judgment. They take the repeatable back-office work, so your hours go to clients, content and the next sale, from wherever you happen to be.

What we found when we mapped our own back office

On October 7, 2026 we pulled six months of activity out of our Operations function’s project system: 5,016 events across 611 threads. We sorted every hands-on action into nine kinds of work and asked one question about each: can an agent do this, and with how much supervision?

That same day our Access Desk ran its first job: five project memberships, each read back from the system, with a receipt for the team. This playbook is built from that plan, with our names and clients taken out, for one person with a laptop.

Three levels: what an agent may do on its own

Every job in the kit carries one of three levels. The level, not how sure the agent sounds, decides what happens next.

Graduation: a level B job can move to level A after 10 clean runs in a row, with receipts posted and zero corrections, and only if it can be undone. Removals, money, domains, client messages and anything you cannot undo never graduate.

The nine kinds of back-office work

Here is the full map, with what each kind looks like when you run a one-person agency. The shares come from our own six months of data, so treat them as a starting estimate for yours.

Share of 2,994 hands-on actions. Level A means the agent runs it end to end; level B means the agent prepares it and you say yes. A kind of work can hold jobs at both levels; the level shown is the most common one. The client access checklist we use is public: the Access Checklist.

The 12 agent job cards

Each card is one bounded job: a trigger, the steps, a never-list, a definition of done, a receipt and an instruction you paste into your AI assistant. Eight run end to end at level A. Four prepare the work and wait for your yes.

The full cards, with every step, an example receipt and the paste-ready instruction, are in the free kit below.

Six rules that keep agents safe with client accounts

These are the lines in the one-page agent policy you sign on day one.

1. Agents draft. You send.Anything that reaches a client, the public or a bank goes out from you, after you read it. Agents prepare it; they never send it.
2. No passwords in chat. Ever.Use each platform's invites and partner access. Keep the rest in a password manager. The scanner reports where secrets are, never what they are, and agents only ever see that list.
3. Agents get their own identity.Create an ops@ address and an agent user in each tool. Agents never work as you, so you can see what they did and switch them off without locking yourself out.
4. Every action ends with a receipt.What changed, where, and a Verified line: what was opened, what it showed, and when. No receipt means it did not happen.
5. Add first. Remove with a yes.An agent may add people or create things on your approved list. Removals, payments, DNS and mail changes need your yes every time. Contracts and anything you cannot undo stay with you.
6. Silence never stops the work.Every ask carries a rescue date and a fallback. When the date passes, the agent runs the fallback or asks you one yes-or-no question.

We wrote up the longer versions: how a request nobody answers still gets done, how we check that done means done, and how agents run our company day to day.

Your first 30 days

Twenty tasks over four weeks. Agents own twelve. You own eight, and each of yours is a credential, a decision or a review. Shadow mode means the agent posts what it would do and changes nothing, so you can compare it with what you would have done. Every task in the kit has a done-when line, so you always know whether it is finished.

What stays yours

Agents take the volume. These five never leave your hands.

Take it free: what is in the kit

Everything on this page, as files you can use today. It is free to use and adapt for your own agency.

More free material: our skill packs, the Task Library of published SOPs, and the Content Factory, where everything Content Intake files goes next: Produce, Process, Post and Promote.

Or build it with us

The AI Builder Program is where young adults put a team of AI agents to work on a real Lighthouse business, under the same rules as this kit: bounded jobs, approval boundaries and receipts. It is selective: we choose for character before technical skill.

• A real Lighthouse client to prove the system on, often a parent's business, or one we supply.
• An installed team of AI agents working from published SOPs, each with a bounded job, an approval boundary and a definition of done.
• A live group call every Thursday at 2 PM Pacific, and a Metrics, Analysis, Action report you submit every Friday.
• Office Hours and a community of more than 400 people when you are stuck.
• In-person meetups, and a chance to speak at events as you advance.

Still deciding? Read why people pay for a working AI system when they could build it themselves, or see what an AI builder does all day.

Questions