Claude Cowork Setup: Replacing a $16K Agency Quote
A coach had been quoted $16,000 a year for four jobs she was already doing herself, badly. Here’s what she handed to Claude Cowork instead, and what she kept.
Table of Contents
- The four jobs one person was doing badly
- What went over first, and why it wasn’t the interesting one
- The four functions, and what each one actually got handed
- What stayed with her, and why that list matters more
- The first fortnight, including what didn’t work
- Doing this yourself, one function at a time
- Frequently Asked Questions
The Short Version
A solo coach we worked with (details changed to protect her privacy) had been quoted $16,000 a year by an agency to cover marketing content, lead outreach, client operations, and reporting. She built a Claude Cowork setup instead, splitting the work across four agents modeled loosely on a marketing, sales, operations, and analytics function. What actually took the longest wasn’t the flashy part. And a good chunk of the work, client judgment, the actual coaching, what to say in a hard conversation, never left her hands.
The four jobs a solo coach is already doing badly, because there is one of her
Sunday night, and she’s still writing captions. Not because she wants to. Because Monday’s post has to go up, the CRM hasn’t been touched since Wednesday, a client proposal is half-finished in a Google Doc from two weeks ago, and nobody else is going to do it. She runs a coaching practice on her own: no agency, no assistant, no marketing hire. She is, functionally, the marketing department, the sales team, operations, and the analyst, and all four jobs get done at whatever quality level is left over after the actual coaching, the part she’s good at and the part clients pay for, is finished.
This isn’t a story about a business that was failing. It was working. It just had a ceiling built into it, and the ceiling was her calendar. A while back, she got a quote from an agency to take some of this off her plate: roughly $16,000 a year to cover content, outreach, and reporting support. That’s a real number, a quote she actually received, not money that left her account. She didn’t take it. What she built instead is a Claude Cowork setup: four agents, each doing one of those four jobs, that now run most weeks without her writing a single caption from scratch.
A quick note before the rest of this: some identifying detail here has been changed at her request, since she hasn’t given permission to be named publicly. The numbers, the $16,000 quote and what the setup replaced, are real and confirmed directly with her. And to be precise about something that’s easy to blur: Claude Cowork is Anthropic’s Claude Cowork, a different product from Microsoft’s separately named Copilot Cowork. If you’ve read about the other one, this isn’t it.
What went over first, and why it wasn’t the interesting one
If you asked her which function she expected to hand over first, she’d probably have guessed content, since that’s the most visible time sink. What actually went first was reporting: pulling numbers from three places into something she could glance at once a week without opening four tabs.
That’s a boring answer, and it’s the right one. Reporting is low-risk to hand over. Nobody gets hurt if a first-draft summary is a little off, because she’s the only one reading it before anything goes anywhere. It also gave her a fast, safe way to learn how to work with an agent, how specific an instruction needed to be, what “good” looked like, before she handed over anything client-facing where a mistake would actually cost her something.
Hand over the lowest-risk function first, not the most visible one. Learn what a good instruction looks like before you hand over anything a client will see.
The four functions, and what each one actually got handed
Once the reporting agent was working, she built out the rest, one at a time, loosely modeled on the four functions a small agency or a first hire would normally cover.
The four-function split: what each agent got, and what came back
| Function | What it was given | What it returned | What got rejected |
|---|---|---|---|
| Marketing (AI CMO) | Her past posts, her tone, the themes she coaches on | First-draft captions and carousel outlines, several options per week | A batch of overly generic “5 tips” posts that read like every other coach’s feed; she rewrote her own voice back in |
| Sales / outreach (AI CRO) | Her ideal client profile and a list of past discovery call notes | Draft outreach messages and a lead magnet outline | An early outreach draft that oversold her results with a specific number she’d never actually claimed; she caught it before sending |
| Operations (AI COO) | Her proposal template and her client onboarding checklist | Drafted proposals from a short client brief, and onboarding emails | Nothing structural, but every proposal number still gets checked against her actual calendar before it goes out |
| Analytics (AI CAO) | Access to her scheduling and basic revenue tracking | A weekly one-page summary: bookings, response rate, where leads are stalling | An early version buried the one number that mattered under six that didn’t; she asked for it to lead with bookings, and it does now |
Based on the setup described directly by the coach, with identifying details changed. Illustrative of the four-function pattern, not a verified case study with published figures.
None of the four run unattended forever. She reviews outreach messages before they send, reads every proposal before it goes to a client, and edits nearly every marketing draft for voice before it posts. The work didn’t vanish, she’s just editing instead of originating, which is a completely different amount of Sunday night.
What stayed with her, and why that list matters more than the other one
This is the part worth spending more time on than the setup itself, because an article about handing four functions to AI that lists no failures and no limits reads like a sales page, and this isn’t one. Three things never moved off her plate, on purpose:
- The actual coaching: what to say in a session, and when to let a point sit instead of pushing it
- Every client relationship, including any message a client would read as coming from a real conversation
- The final check on anything client-facing, proposals, outreach, and invoices, before it goes out
The actual coaching never left her hands, and it was never going to. What a client needs to hear in a specific session, the judgment call about whether to push someone or let a point sit, the read on a client’s tone shifting mid-call, none of that is delegable, and she never asked it to be. Client relationships stayed entirely hers: no agent has ever sent a message a client would read as coming from a real conversation that didn’t happen.
There was at least one clear failure. Early on, the marketing agent produced a run of generic advice posts, the kind that could have come from any coach’s account, technically fine and completely unusable. She’d given it her past posts but not enough of the actual thinking behind them, and it defaulted to safe, generic phrasing instead of her specific point of view. Fixing it meant feeding it more of her actual reasoning, not just her finished posts, and it took a few rounds before the drafts sounded like her again.
The first fortnight, including the parts that didn’t work
A realistic first two weeks with a new agent function
Set up the reporting agent first. Output is rough: right structure, wrong emphasis, needs specific correction.
Reporting stabilizes. Marketing agent goes live and produces the generic-sounding batch of posts. It genuinely feels like a step backward this week.
Feed the marketing agent more of her actual voice and reasoning, not just past output. Drafts start sounding like her.
Outreach and operations agents come online. The oversold outreach draft gets caught and corrected. By day 14, three of four functions are producing usable first drafts most days.
A worked illustration based on the sequence described directly by the coach. Setup duration overall isn’t stated in this article; see the note below.
Worth being straight about something most stories like this leave out: how long the whole build actually took isn’t something we’re stating here, because it isn’t something we’ve confirmed with her precisely, and guessing at it would be inventing a detail about a real person’s case. What we can say is that it took real, sustained work across those four functions, not an afternoon. Once it was built, running the monthly reporting cycle that used to eat a chunk of her Sunday now takes about two minutes. Those are two different clocks: the one-time build, and the weekly run once it exists. Don’t confuse them, because the second number is only true because of the first one.
Doing this yourself, at the scale of one function rather than four
You don’t need all four functions to get something real out of this. If you’re a solo consultant, coach, or service business owner reading this and thinking about where to start, pick the lowest-risk, most repetitive piece of your week, the thing closest to what reporting was for her, and build one agent for that first.
Starter worksheet: your first agent, not all four at once
| Field | Fill this in for yourself |
|---|---|
| The recurring task that eats the most low-value time | Name it specifically: not “marketing,” but “writing this week’s three social captions” |
| What you’d give it to start | Ten to twenty examples of your own past work in that exact format, not a style guide, actual examples |
| What it must never do without you | Send anything client-facing, quote a number or result you haven’t confirmed, publish anything live |
| How you’ll know it’s working | You’re editing a draft in minutes instead of starting from nothing, most weeks, for four weeks straight |
A blank version of the pattern this coach used for her first function. Fill it in before you touch any tool.
If you want the guided version of this rather than working it out alone, Future Factors runs a Claude Cowork course, Your AI C-Suite: Build AI Agents that Run Your Business with Claude Cowork, where all four functions get built live over the course of the program with the prompts and workflows included. As of publishing, the course is sold out with a new cohort planned; you can join the waitlist on the course page. You can absolutely start with one function on your own using the worksheet above. The course mostly saves the weeks of trial and error this coach went through building the other three.
One thing worth stating plainly: none of this is free. Claude Cowork needs a paid Claude plan (Pro, Max, Team, or Enterprise), and the course itself is a paid program. If budget is the constraint, start with the single-function approach above using whatever Claude plan you already have.
For more on what Claude Cowork actually is before you build anything, see our plain-English guide to what Claude Cowork is, and for the operating mechanics, context files, delegating outcomes instead of steps, connectors, safety defaults, see how to build AI agents in Claude Cowork. This article is about the what and why of splitting one person’s work into four functions. Those two cover the how.
Frequently Asked Questions
What is Claude Cowork and how is it different from Claude chat?
Claude Cowork is Anthropic’s environment for building and running standing AI agents that handle recurring work over time, rather than a single back-and-forth conversation. Where a chat answers one question at a time, a Claude Cowork agent can be set up once with context about your business and then run a repeatable task, like drafting a weekly report, on an ongoing basis.
Can a solo business owner set up AI agents without any code?
Yes. The setup described in this article was built by a solo coach with no coding background, using Claude Cowork’s own interface. The skill that actually matters is giving clear context and examples of your own past work, not writing code.
What should a coach or consultant hand to an AI agent first?
Start with the lowest-risk, most repetitive task, something like internal reporting that only you review, rather than anything client-facing. It’s the safest way to learn what a good instruction looks like before handing over something a client will actually see.
Do you need a paid Claude plan to use Claude Cowork?
Yes. Claude Cowork requires a paid Claude plan, Pro, Max, Team, or Enterprise. It isn’t a free tool, and any setup built on top of it inherits that cost.
How long does it take before an AI agent setup actually saves time?
The build itself takes real, sustained work, not an afternoon, and the first outputs usually need correction before they’re usable. Once a function is genuinely working, the ongoing time savings can be dramatic: in the case described here, a weekly reporting run that used to take a meaningful chunk of a Sunday now takes about two minutes, but that’s only true because of the setup work that came before it.
About This Article
Based on a real Claude Cowork setup described directly to Future Factors by the business owner, with identifying details changed since she hasn’t given permission to be named. The $16,000 agency quote and the roughly two-minute weekly run time are confirmed figures. Setup duration is not stated because it hasn’t been confirmed precisely, and estimating it would mean inventing a detail about a real person’s case. Course details and pricing confirmed live against the Maven course page on the day of publishing.
Sources
- Direct account from the business owner described in this article (anonymised at her request)
- Maven, Your AI C-Suite: Build AI Agents that Run Your Business with Claude Cowork (course page, checked live September 9, 2026)
