AI CEO LabHow to Run a Marketing Agency From 3 to 30 Clients
A short primer if you're starting, then the operating system: departments with owners, onboarding and reporting SOPs you can copy, a weekly account review with dated decisions, churn watch, margin per client, and AI agents that draft and never send.
Getting clients is the part everyone teaches. Keeping them is where agencies break: a client whose onboarding stalled, an ad account nobody looked at for 3 weeks, a report that went out late, a retainer that loses money every month.
This guide is the operating file my agency, Dr. Lead Flow, runs on. You get 5 departments with an owner each, onboarding and reporting SOPs you can copy, one CRM and one project board, a weekly account review with dated decisions, a churn watch, a margin check per client, and where AI agents fit. They draft. People send.
I'm a doctor who runs a marketing agency, not an engineer. Starting an agency? Read the next section first. Already running one with 3 or more clients? Skip to step 1.
What you have by the end
Not concepts. Things that exist on your machine when you close the tab.
- A plain answer to what running an agency involves, and what to skip while you're still starting
- 5 departments, each with a named owner, including the seats you still cover yourself
- Onboarding and reporting SOPs you can copy and edit tonight
- One CRM and one project board as the only places client truth lives
- A weekly account review where every client gets a dated decision line
- A churn watch list and a margin worksheet per client
- Draft-only AI agents seated under people, starting with one
Start here: what running a marketing agency involves
Starting an agency and running one are different jobs. Most videos and articles on this topic cover the first. Here's a short version of that, then how the second works.
If you're starting: 3 decisions, then clients
Pick one kind of client you understand, one service you can deliver well, and one price. Then get clients from people who already know you before you spend on ads. My agency serves health businesses, and being a doctor helps me understand what their ads may and may not say.
Don't build departments, dashboards or agents at 1 or 2 clients. You don't have enough work yet to know what a good week looks like, and systems built on 2 clients encode guesses. Come back to this guide when you have 3.
What running one means
Running an agency is one loop, repeated for every client: onboard them, make a decision on their account every week, tell them what happened, watch for signs they'll leave, and check that they're profitable. Each box in the loop needs an owner, a written SOP (standard operating procedure) and a place the result gets recorded.
The difference from the starting phase is who does the work. At 2 clients, you do everything from memory. At 10, memory fails, and the loop has to run whether or not you're in the room.
What the numbers say
Promethean Research's 2026 State of Digital Services report (119 agency owners and managers, surveyed February 2026) found the average agency made a 13% after-tax net margin in 2025. Their profitability analysis puts studios under 10 full-time staff at 19%. In a separate survey of 165 agency leaders (published 3 July 2026), 24% said their typical client leaves within a year, and 62% said slow access to client systems during onboarding hurt client confidence (checked 15 September 2026).
So thin margins and early churn are normal. Both are operating problems, which is why this guide spends its steps on onboarding, weekly decisions, churn and margin.
Where AI fits
AI is now standard in marketing work. HubSpot's 2026 State of Marketing report (1,500+ marketers, updated 10 April 2026) found 86.4% of marketing teams use AI in at least a few areas. Among agencies, Promethean found 34% had implemented AI across the business and 28% were implementing it (checked 15 September 2026).
Owners are arguing about what that does to the business model in threads like Google is killing marketing agencies (34 comments) and anyone running an agency where AI changed the business model? (24 comments). On YouTube, Jordan Platten's 1-person AI agency walkthrough had about 39,600 views and Eric Siu posted how he runs an agency with 6 AI agents (checked 15 September 2026).
My position: agents read, pull numbers and draft under a named person. They don't send to clients, spend money or change accounts.
The catch
None of this replaces sales. An agency with great SOPs and too few clients is a well-organized agency that's running out of money. Install the weekly review first, because it costs 2 hours and no hires. Add the rest when quality starts depending on which day the work happened.
| Where you are | What to install | What to skip |
|---|---|---|
| 1 to 2 clients | One offer, one price, a signed agreement, a shared folder per client | Departments, dashboards, agents |
| 3 to 10 clients | The weekly account review (step 5), onboarding and reporting SOPs, one CRM | Hiring ahead of a written SOP |
| 10 to 30 clients | Every step here: owners per department, churn watch, margin per client, one draft-only agent at a time | Letting an agent send anything client-facing |
| Term | What it means |
|---|---|
| SOP | Standard operating procedure. A one-page written checklist for a recurring job |
| Definition of done | One sentence a stranger could check to confirm the job is finished |
| Retainer | A fixed monthly fee for ongoing work |
| CRM | Customer relationship management software. Where leads, clients and their history live |
| Project board | A shared board of tasks with owners and due dates |
| Account review | The weekly check of every client's ad account and results |
| Decision line | A dated one-line record of what you decided on an account and why |
| CPL | Cost per lead. Ad spend divided by leads |
| CTR | Click-through rate. The share of people who saw an ad and clicked |
| Frequency | The average number of times one person saw your ad |
| Learning phase | Meta's early testing period for an ad set. Big edits can restart it |
| Churn | A client leaving |
| Gross margin per client | What a client pays minus the direct cost of serving them, as a percentage |
| Net margin | What's left of revenue after every cost and tax |
| Draft-only agent | An AI agent that prepares work for a person to approve and never sends it |
| Reserved list | The decisions only the owner makes |
Should I be building this at all yet?
I have 3 clients and I'm drowning. Does installing departments and agents fix that, or stop me getting to 10?
At 3 clients it mostly stops you. The research on this guide pushed back on my own first draft hard enough that I changed it.
- Small numbers are noise. At 3 clients one bad month looks like a trend, and a system built on it encodes the mistake.
- Install the rest when quality starts varying by which day or which person did the work. Until then, the constraint is volume, and volume comes from sales calls.
- At 3 clients, install step 5 only. A fixed weekly day with a dated decision per account costs 2 hours and no hires.
The steps below come from documents my agency runs on. The ad bands in step 5 are mine and change monthly, so copy the shape, not the numbers. The one rule I'll defend without qualification is draft-never-send. The first time an agent emails a client something wrong, you lose trust in the whole setup.
The operating system, in the order to install it
1. Name 5 departments and put a person on each
CEO, Marketing, Tech, Client Success, Media. At Dr. Lead Flow that's 6 people plus me: a media buyer, a CTO, a client success manager, a social media manager and a video editor in Media, and an operations and compliance person across all of it. Write one name per department. Where a seat is empty, write "CEO covers until hired."
My hiring rule: is there an SOP, can an agent run it, and if not, hire against the SOP.
Why: a department without a name is still you. Writing "CEO covers" makes the load visible, so you can see which seat to fill first.
2. Pick one CRM and one project board
Every client's truth lives in 2 places. The CRM holds leads, pipeline stages and client contact history. Mine is GoHighLevel. The project board holds every task with an owner and a due date. We use monday.com as the source of truth for clients and work. SOPs, client notes and decision lines go in one shared wiki. Mine is an Obsidian vault.
If client numbers live in 2 spreadsheets, a CRM and someone's inbox, fix that before anything else in this guide. Any tools work if everyone agrees there's one place to look.
Why: the weekly review, reports, churn watch and agents all read from these places. If there are 2 versions of the truth, every later step inherits the argument about which is right.
3. Write the onboarding SOP first
Onboarding touches every department, and a bad one stays invisible for weeks, then shows up as churn. Use the same one-page shape for every SOP, and write the definition of done before the steps. If a stranger couldn't verify it, the SOP is a wish.
Mine has 8 checks. Two matter most: collect the client's compliance constraints and who approves creative at kickoff, and verify lead routing with a real test lead (here's how I build the GoHighLevel workflows). Ask their budget on the kickoff call.
If you do something twice and it isn't written down, write it the second time.
Why: Promethean's July 2026 survey found 62% of agency leaders said onboarding access delays hurt client confidence (source, checked 15 September 2026). Access and routing are exactly what this checklist forces.
SOP TEMPLATE (one page, every recurring job)
# SOP: <job name>
Owner: <department> Cadence: <when it runs>
Systems: <the exact tools it touches>
Definition of done: <one sentence a stranger could verify>
Steps:
1. ... (each one a thing you could watch someone do)
Guardrails:
- <what goes wrong if a step is skipped>
- <the threshold that needs an ask instead of an action>
ONBOARDING: DEFINITION OF DONE (all 8, or the client isn't onboarded)
[ ] CRM sub-account live; lead form -> pipeline stage -> notification
[ ] Verified with a REAL test lead that a person watched arrive
[ ] Partner access to the ad account; account ID in the client file
[ ] Client file created: status, account ID, offer, budget,
empty performance section
[ ] Client on the project board with onboarding tasks and owners
[ ] Kickoff assets in hand: offer, pricing, booking link, brand assets,
COMPLIANCE CONSTRAINTS, and the named person who approves creative
[ ] First campaign live: conservative budget, one proven angle + one test
[ ] Kickoff call logged in the client file; first report date on the calendar4. Write the reporting SOP second
Every client hears from you every week, logged. Client Success sends a short update on the client's preferred channel: the numbers, one insight, one next action. Report leads and booked appointments, not impressions. Once a month, walk through a scorecard on a call, because a call is a retention moment and an emailed PDF isn't.
Escalate to the owner immediately when cost per lead passes twice your benchmark for 2 weeks or a client is unhappy, and bring a proposed fix. Collect case studies when a result lands, not at renewal.
Why: clients who always know what happened and what's next don't get surprised by an invoice. A logged touch also gives the churn watch in step 6 something to check.
SOP: CLIENT REPORTING Owner: Client Success Cadence: weekly update + monthly call Systems: CRM (source numbers), wiki (client note), client's channel Definition of done: every active client has a logged update dated within the last 7 days, and a scorecard call this month. WEEKLY UPDATE (paste, fill, send) Hi <name>, here's your week (<date range>): - Leads: <n> (last week <n>) Booked: <n> Cost per lead: $<x> - What we saw: <one insight in plain words> - What we're doing next: <one action, with a day> Anything you need from us, reply here. LOG LINE (client note) <date> | weekly update | <channel> | sent by <name> MONTHLY SCORECARD CALL (15 minutes) 1. Leads, booked, cost per lead vs last month and vs target 2. What worked, what we stopped 3. Next month's one priority 4. Anything bothering you? (write the answer down word for word)
5. Run a weekly account review with dated decisions
Pick a day and never move it. 6 things, in order: 1. Scoreboard against your targets. 2. A decision for every client ad account, with a dated line in the client's note. 3. Comms audit: every active client logged within 7 days. 4. One test lead through your own intake funnel. 5. Content cadence check. 6. Project triage: what moved, what's blocked, what gets build focus.
The bands below are from my own client health standards, recalibrated monthly. Yours will be different numbers in the same shape, so don't use mine as benchmarks. Log the decision even when it's hold.
The 20 to 30 percent scale step is my own rule, not Meta's. Meta doesn't give a percentage. It says budget and bid changes may count as a significant edit depending on their size, which can put an ad set back into learning, so I never double a budget overnight. For the account check itself, run the Meta ads audit.
Done when: every account has a dated line this week, and nothing flagged kill last week is still spending.
Why: without a fixed day, a quiet account coasts for weeks. The dated line proves someone looked.
AD DECISION BANDS (my book: health lead gen on Meta.
Recalibrated monthly. Yours will differ.)
CPL <= $10 -> SCALE raise budget 20-30% (my rule), never
double overnight
CPL $10-14 -> HOLD note the trend direction
CPL $14-20 -> WARNING diagnose creative / audience / offer,
refresh within a week
CPL > ~2x book avg -> KILL pause, move budget to the winner
sustained
CTR < ~2.5% -> creative fatigue: refresh, don't raise budget
Frequency > ~2.5x -> audience saturation: broaden or rotate
(7-day, cold; my rule of thumb, not a Meta limit)
One campaign -> fragility flag regardless of CPL
carrying a client
THE DECISION LINE, pasted into the client note:
2026-09-08 | <campaign> | HOLD | CPL $11.40 (was $9.80) | watching
frequency at 2.3, refresh queued for next week6. Keep a churn watch list
Churn rarely arrives as a surprise. It shows up first as a missed touch, a rising cost per lead, a short reply or a client who stops showing up to calls. Once a week, during the review, run every active client through the signals below and put any client with 2 or more on the watch list with an owner and a fix date.
Then do something the same week: a call from the owner, a creative refresh, an offer change or an honest conversation about fit.
Why: 24% of agency leaders in Promethean's July 2026 survey said their typical client leaves within a year (checked 15 September 2026). A written list turns a feeling that a client is drifting into a dated action someone owns.
CHURN WATCH (weekly, every active client) Signal Yes/No 1. No logged touch in the last 7 days [ ] 2. CPL above 2x benchmark for 2+ weeks [ ] 3. One campaign carrying the whole account [ ] 4. Missed or rescheduled the monthly call [ ] 5. Complaint, short replies or silence [ ] 6. Onboarding item still open after 30 days [ ] 7. Late on the invoice [ ] 2 or more yes -> watch list: <date> | <client> | signals: <numbers> | owner: <name> | fix: <one action> | review by: <date>
7. Check margin per client every month
An agency can grow revenue and lose money at the same time, one underpriced client at a time. Once a month, fill in the worksheet below for every client. You need 3 numbers: what they pay, the hours your team spent on them, and the tools or contractors that exist because of them.
Use a loaded hourly cost for each person: pay plus taxes and benefits, divided by the hours they work. Rough is fine. You're looking for the client at the bottom of the list, not a perfect figure.
Why: the average agency made a 13% after-tax net margin in 2025, per Promethean Research (checked 15 September 2026). With that little room, one client who takes twice the hours of the others can erase the profit on 2 or 3 good ones. The fix is a price change, a scope change or a planned exit, and you can't choose without the number.
MARGIN PER CLIENT (monthly)
Client: <name> Month: <month>
A. Monthly fee $______
B. Team hours this month x loaded hourly cost $______
(media buyer __h x $__ + client success __h x $__
+ media __h x $__ + tech __h x $__)
C. Direct costs for this client $______
(contractors, client-specific software, stock, etc.)
Gross margin $ = A - B - C $______
Gross margin % = (A - B - C) / A x 100 _____%
Sort all clients by gross margin %. For the bottom one, write:
<date> | <client> | margin __% | decision: reprice / rescope /
reduce hours / exit at renewal | owner: <name>8. Write the reserved list before any agent runs
Write 2 columns: what people and agents may do without asking, and what comes to you. Do this before step 9, so every agent starts inside the rules.
Mine, without asking: ad budget shifts within the standing rules and kill-zone pauses, sending scheduled client reports, publishing inside the content strategy, fixing broken pipeline steps and restarting services, writing and updating SOPs.
Comes to me: any new spend swing above $50 a day outside my standing rules, any new tool subscription, pricing changes, discounts and refunds, anything near a medical claim in copy, deleting data or rotating keys, changing a shared schema, and signing or ending a client or contractor.
Plus 3 things only I do: sales calls, content on camera, and the final scale-or-kill beyond the standing rules.
For agents, one extra line: drafts anything client facing, never sends. Here's the permission model I use.
Why: if people still ask you about everything, your "without asking" column is too short. If an agent can act without a rule, the first mistake lands on a client.
9. Seat one draft-only agent, then add the next
An agent sits under a person and works from that person's SOP. It reads and drafts. The person above it decides, so a name is attached to everything the agent produced.
Start small: one agent, one job, for one week. The best first job is drafting the weekly client updates from step 4, because you can check every draft against the CRM. Fill in the job card below.
My roster in Slack today runs on Viktor, all draft-only: Marcus reviews every Meta account Monday at 7am and recommends scale, hold, refresh or kill, under the media buyer. Devon audits CRM workflows Tuesday, under the CTO. Jade grades every DM conversation daily at 10am, under Marketing. Sophia scans the week's recorded calls and flags at-risk accounts daily at 9am, under Client Success. Ciara pulls Instagram and YouTube metrics and competitor ideas Wednesday, under Media. Atlas runs the CEO pulse Friday, for me. Every run ends with DID, DECIDED, BLOCKED, NEXT in the day's log.
For the full setup, with one channel, rules and schedule per client, follow AI for marketing agencies. Here's the Viktor roster in detail.
Why: 6 agents added in one week produce 6 reports nobody reads. One agent read closely for a week teaches you what to fix in its instructions before the next one starts.
AGENT JOB CARD (fill in before the agent runs)
Agent name: <name>
Under: <person and department>
Job: <one recurring job from an SOP>
Cadence: <day, time>
Reads: <exact sources, read only>
Posts to: <one channel>, as a draft for <person>
Never: sends to clients, spends, edits records, changes budgets
Log: every run ends with DID / DECIDED / BLOCKED / NEXT
Review: <person> reads every output for 7 days before the next
agent is added10. Protect the daily 15 minutes and the owner's week
Daily, about 15 minutes: read the morning brief and act on flags, check overnight leads landed in the CRM complete, clear or reschedule anything overdue, and review drafted replies. If it regularly takes longer, a system is broken, not a person. The fix goes on the build queue.
The delegation rule: when I propose a build, my chief of staff first asks whether it's owner work (revenue, strategy, content) or delegate work (build and integration). Delegate work goes to the CTO as a ticket with an outcome, inputs, definition of done and deadline. Being able to build it isn't a reason to. My own operating notes name this as my bottleneck, which is why an agent asks me the question instead of me remembering to. The solo operator version of this rule works the same way.
Why: the owner's week is the agency's scarcest resource. Every hour spent building an automation is an hour not spent on sales calls, which is where the next 10 clients come from.
Starter prompts
Paste these as written. They are short on purpose, because the long ones drift.
It is ad day. For every client account, using the connected read only ads data for the last 7 days and the 7 before: recommend scale, hold, refresh or kill with the number behind it. Use MY bands, which I am giving you now and which you must not substitute with your own: cost per lead at or under [X] is a scale candidate, raise 20 to 30 percent and never double overnight; [X] to [Y] is hold, note the trend; [Y] to [Z] is a warning, diagnose creative, audience or offer and refresh within a week; sustained above twice my book average with meaningful spend is a kill, reallocate to the winner. Click through under [C] percent or frequency over [F] is fatigue, queue a creative refresh and do not raise the budget. Flag any account where a single campaign carries more than half the spend. Write the dated one line decision for each client note in this shape: date | campaign | decision | the number | what changes and why. Flag for Client Success anything the client should hear about this week. Do not change anything. The media buyer acts.
Draft this week's update for [client] using only the numbers in [CRM or report source] for [date range] and last week. Use this shape: leads, booked appointments and cost per lead with last week's figure beside each; one insight in plain words; one next action with a day. No impressions, no jargon, under 90 words. If a number is missing, write MISSING instead of estimating. Post it as a draft for [person] to review. Do not send it.
I am about to do this: [task]. Classify it as owner work (sales, content, offers, reserved decisions) or delegate work (build, integration, automation, reporting). If it is delegate work, write the ticket for [name of tech owner] with the outcome, the inputs, the definition of done and the deadline, and tell me plainly not to build it myself. If it is owner work, tell me the single next action and stop. Do not offer to help me build it.
If this happens, do this
Every one of these has happened to me or to someone I set this up for.
| Failure | Fix |
|---|---|
| Installed 5 departments at 2 clients | Wrong problem. Get more clients first, and run only the weekly review |
| Client numbers disagree between the CRM and a spreadsheet | Pick one source of truth (step 2) and retire the other |
| A new client waited 2 weeks for ad access | Put access on the onboarding checklist with an owner and a due date |
| An account coasted for 3 weeks | Fixed weekly day. Done means a dated line per account |
| A client said they never hear from you | Weekly update, logged, from the reporting SOP |
| A client left and nobody saw it coming | Run the churn watch weekly and act on 2 or more signals the same week |
| Revenue grew and profit didn't | Fill in margin per client and deal with the bottom client first |
| An agent sent a client update on its own | Draft only, under a named person. Set the permissions |
| 6 agents added in one week and nobody reads them | One agent, one week of reading its output, then the next |
| Owner built the automation over a weekend | Run the delegation check first. Ticket it to Tech |
The tools I use for this
| Tool | What it is for here | |
|---|---|---|
| GoHighLevel † | The CRM backbone for every client: lead routing, pipelines and what each client's leads did. | Get it |
| monday.com | The project board and source of truth for clients and who owns which task. | Open |
| Viktor † | The Slack coworker that carries the draft-only agent roster under each department. | Get it |
| Claude | Drafts SOPs, weekly updates and the ad day read from my bands. | Open |
| Meta Ads Manager | Where the media buyer acts on the weekly decision. Agents read it and never touch it. | Open |
| Obsidian | The shared vault for SOPs, client notes, decision lines and the daily log. A wiki works too. | Open |
What I'd tell you over coffee
- Write the definition of done before the steps. Most bad SOPs are a list of actions with no way to tell when the job is finished.
- Log the hold decisions too. A month of holds on one account is a signal you'd miss if you only logged changes.
- Ask a new client their budget on the kickoff call. People spend more with someone who's visibly comfortable with the number.
- Put the margin worksheet next to the churn watch. A low-margin client on the watch list is a different conversation from a profitable one.
Questions people ask
How do you run a marketing agency?
Give every department a named owner, write an SOP for each recurring job, and keep client truth in one CRM and one project board. Then run a fixed weekly review where every account gets a dated decision, send every client a logged weekly update, and check churn signals and margin per client. The 10 steps above cover each part.
What SOPs does a marketing agency need?
Start with client onboarding and client reporting, since they touch every client. Then add weekly ad management, content production and publishing, and an incident process for broken tech. Each should fit on one page with an owner, a cadence, steps and a definition of done. Steps 3 and 4 have templates you can copy.
How do I start a marketing agency with no experience?
Pick one kind of client and one service you can deliver, practise on your own accounts or a friend's business, and get first clients from people who already know you. Keep systems minimal until you have 3 clients. The Start here section covers what to install at each stage.
Is running a marketing agency worth it?
It can be, but margins are thinner than most people expect. Promethean Research's 2026 report found the average agency made a 13% after-tax net margin in 2025, with studios under 10 staff averaging 19% (checked 15 September 2026). Whether it's worth it depends on pricing, retention and how much delivery depends on you.
Can one person run a marketing agency with AI?
A solo operator can serve a small number of clients with AI doing the drafting, reporting and number pulling, and videos like Jordan Platten's 1-person AI agency walkthrough show that approach. Sales calls, client relationships and final decisions still need a person. Keep every agent draft-only so nothing reaches a client without your yes.
How do marketing agencies use AI?
Mostly for content, media and admin work. HubSpot's 2026 State of Marketing report found 86.4% of marketing teams use AI in at least a few areas (checked 15 September 2026). In agency operations, AI works well drafting weekly client updates, reading ad accounts against your rules and flagging at-risk clients. See AI for marketing agencies for a full setup.
The free skill
It checks whether you're ready for structure, maps your agency onto 5 departments with owners, writes the onboarding and reporting SOPs, installs the weekly account review with dated decisions, runs the churn watch and margin worksheet, writes the reserved list, and seats one draft-only agent at a time. Related guides: AI for marketing agencies and AI agent guardrails.
--- name: how-an-agency-runs description: Sets up the operating system for a marketing agency with 3 to 30 clients. Checks readiness, maps 5 departments with owners, picks one CRM and one project board, writes the onboarding and reporting SOPs, installs the weekly account review with dated decisions, runs a churn watch and margin per client, writes the reserved list, and seats one draft-only agent at a time. Trigger on "structure my agency", "how do I run my agency", "agency SOPs", "org chart for my agency", "which agent under which person", "get me out of the day to day". --- # How An Agency Runs Everything either has an owner and an SOP, a standing rule, or reaches the owner. ## Step 0: Check it's time Ask how many active clients the agency has. - 1 to 2 clients: say the constraint is sales, not structure. Help with one offer, one price, a signed agreement and a folder per client. Stop. - 3 to 10: install Step 6 (weekly review) first, then onboarding and reporting SOPs and one CRM. - 10 to 30: run every step. ## Step 1: Interview Ask one at a time: 1. Who is on the team, and what does each person do this week? 2. Which recurring jobs exist? (ads, onboarding, reporting, content, automations, intake, billing) 3. Which of those does the owner still do personally, and why? 4. Where do client numbers and tasks live? Is it one CRM and one board? ## Step 2: Departments and seats | Department | Owns | Typical SOPs | |---|---|---| | CEO | Strategy, offers, pricing, sales calls, final scale or kill | Scoreboard, monthly review | | Marketing | Paid ads and lead flow, for clients and the agency | Weekly ad management, intake funnel | | Tech | Builds, integrations, infrastructure, the agent fleet | Infrastructure map, incidents | | Client Success | Onboarding, reporting, retention, comms logging | Onboarding, reporting, churn watch | | Media | Content production and publishing | Production, publishing | Assign a person to each. Mark an empty seat "CEO covers until hired". To fill a seat: is there an SOP? Can an agent run it? If not, hire against the SOP. ## Step 3: One source of truth One CRM for leads, pipeline and client history. One project board for tasks with owners and due dates. One wiki for SOPs, client notes and decision lines. If numbers live in two places, flag it and resolve it before anything else. ## Step 4: SOPs, onboarding first, reporting second ``` SOP: [name] OWNER: [department] CADENCE: [when] SYSTEMS: [tools] DEFINITION OF DONE: a sentence a stranger can verify (write this first) STEPS: numbered, each something you could watch someone do GUARDRAILS: what needs the CEO, what never happens ``` If something is done twice and not written, write it the second time. Onboarding definition of done: CRM sub-account live with lead routing verified by a real test lead; partner access to the ad account with the ID recorded; a client file (status, account ID, offer, budget); client on the project board with onboarding tasks; kickoff assets in hand, including compliance constraints and who approves creative; first campaign live with one proven angle plus one test; kickoff call logged and first report dated. Reporting definition of done: every active client has a logged update within 7 days and a scorecard call this month. Weekly update shape: leads, booked, cost per lead with last week beside each; one insight; one next action with a day. Log line: date | weekly update | channel | sent by. Escalate to the CEO when cost per lead passes 2x benchmark for 2 weeks or a client is unhappy, with a proposed fix. ## Step 5: An agent under a person, one at a time Only after Step 8's reserved list exists. First job: drafting the weekly client updates. ``` AGENT: [name] UNDER: [person] CADENCE: [day, time] READS: [exact inputs, read only] POSTS: [report shape] to [one channel], as a draft NEVER: sends to clients, spends, edits a record, changes a budget LOG: every run ends with DID / DECIDED / BLOCKED / NEXT REVIEW: the person reads every output for 7 days before the next agent ``` Reference roster: ads review Monday morning (under the media buyer, recommends scale / hold / refresh / kill), CRM workflow audit Tuesday (under Tech), DM grading daily (under Marketing), client health from the week's calls daily (under Client Success), channel metrics Wednesday (under Media), CEO pulse Friday. ## Step 6: The rhythm **Daily, about 15 minutes:** read the brief and act on flags; confirm overnight leads landed in the CRM; clear or reschedule overdue tasks; review drafted replies. Longer than that means a broken system. **Weekly, about 2 hours, fixed day:** scoreboard against targets; a dated scale / hold / refresh / kill line for every client account, including holds; every client logged within 7 days; one test lead through the intake funnel; content cadence check; project triage; churn watch (Step 7). Done when every account has this week's line and nothing flagged kill last week is still spending. Decision bands use the owner's own numbers, recalibrated monthly, never benchmarks: scale (raise 20 to 30 percent, never double; this step is the owner's rule, not Meta guidance, Meta gives no percentage), hold, warning (diagnose creative, audience or offer), kill (sustained, well above book average). Fatigue means refresh creative, not budget. Flag any client carried by one campaign. Decision line: date | campaign | decision | the number | what changes and why. **Monthly, half a day:** close the numbers; margin per client (Step 7); scorecard calls held; thresholds recalibrated; the wiki brought back to true. ## Step 7: Churn watch and margin per client Churn signals, weekly per client: no logged touch in 7 days; CPL above 2x benchmark for 2+ weeks; one campaign carrying the account; missed monthly call; complaint, short replies or silence; onboarding item open after 30 days; late invoice. Two or more: watch list line with owner, one fix and a review date. Act the same week. Margin, monthly per client: fee minus (team hours x loaded hourly cost) minus client-specific direct costs, as a percentage of fee. Sort clients. For the bottom one, record a decision: reprice, rescope, reduce hours or exit at renewal. Ask the owner for hours and costs; never estimate them. ## Step 8: Reserved list and standing rules **Reserved for the owner:** sales calls, content on camera, offers and pricing, final scale or kill beyond standing rules, signing or ending clients and contractors, anything touching a medical or regulated claim. **Act without asking:** budget shifts inside an agreed percentage and kill-zone pauses; sending scheduled reports; publishing inside the content strategy; fixing broken pipeline steps; updating the wiki. **Ask first:** spend swings above an agreed daily amount outside the rules; new tool subscriptions; pricing changes, discounts, refunds; deleting data or changing shared schemas. Thresholds are numbers the owner gives you. ## Step 9: The delegation rule Whenever the owner proposes a build, classify it: owner work (revenue, strategy, content) or delegate work (build, integration). Delegate work becomes a ticket for Tech with outcome, inputs, definition of done and deadline. Say plainly that being able to build it isn't a reason to. ## Rules - Roles, rhythms and templates only. Never invent clients, counts, revenue, hours, margins or thresholds. - Every agent is draft-only under a person. Nothing client-facing ships without a human yes. - One source of truth for client numbers. If there are two, flag it before anything else.
Want your agency's operating system built on a call?
Inside the AI CEO Lab we map your departments, write your onboarding and reporting SOPs, set up the weekly review and seat your first draft-only agent in one session.