A Brooklyn founder spun up an AI marketing, sales, finance, and product department out of chat projects — and a chief-of-staff to summarize them. It works, right up to the point where a pile of chat windows stops being a team.
On August 17, Business Insider ran an as-told-to piece from Sara Russell, a 37-year-old founder of a Brooklyn bath-and-body business. She built what she calls an AI C-suite out of chat projects — a marketing department, a sales department, a finance department, and a product-development department, each one its own project, plus an AI chief of staff that compiles a monthly recap across all of them.
She told each project to collaborate, offer ideas, and defer to her as CEO. Twenty dollars a month. She reviews every recommendation and makes every final call.
Read that again, because it is the whole story of where small-business AI actually is right now. Not the vendor keynotes. Not the eleven-figure infrastructure announcements. A soap maker in Brooklyn independently reinvented the org chart — because that is the shape the work wants to take.
She did not ask for a bigger chatbot. She asked for a team. And she built the closest thing chat windows can give her.
Here is what we want to say plainly, because it is easy to miss under the "clever founder" framing: she is right about the destination and constrained by the vehicle. The instinct — separate roles, a chief of staff, one human at the top signing off — is exactly correct. The medium she is forced to use, a stack of disconnected chat projects, is exactly what will cap how far it goes.
Strip away the specific product and look at the design decisions. Every one of them is the decision a good operator makes.
She created roles, not tasks. A finance function, a marketing function — each with a standing purpose, not a one-off prompt. That is the difference between hiring and errand-running.
She defined authority and escalation. Every project was told: you advise, I decide. That is an intent boundary. It is the same thing a real Teammate's configuration encodes — where the coworker can act on its own, where it has to stop and ask.
She built a chief of staff to aggregate the other functions into one monthly briefing. That is coordination. It is the recognition that a team of coworkers who never talk to each other is not a team; it is a group of strangers.
None of that is prompt cleverness. That is organizational design. A solo founder, with no AI background and an explicit distrust of the technology touching her creative work, arrived at seams, roles, and human-in-the-loop review on her own — because those are the load-bearing walls of any operation, human or AI.
The demand signal could not be louder. People do not want an assistant. They want coworkers with jobs. The market keeps building teams out of whatever material is at hand.
Now the hard part. Here is where this architecture breaks, and it breaks in ways that get worse precisely as the business grows — which is the worst possible time.
Her marketing project does not know what her finance project decided last week. Each chat is its own island. When the sales function learns that a product line is underperforming, the finance function finds out only if she personally carries that fact across the gap and re-types it.
She is the integration layer. She is the message bus between her own departments. That works at one person and one product. It does not survive the second product, the first part-time hire, or the first month she is too busy to run the sync herself.
A real team shares context. When one coworker learns something, the others can see it — governed, scoped, and portable, not trapped in a transcript only the founder can read.
A chat project remembers what is in its window until the window fills up, the account changes, or the context gets summarized into mush. There is no durable, inspectable memory that survives a model upgrade or a plan change. When the vendor rotates the underlying model — which happens every quarter now — the "employee" she trained can quietly change behavior with no version history and no rollback.
This is the difference between a coworker and a session. A coworker picks up exactly where it left off. A session starts over, and you do not always notice it started over until the output drifts.
The entire operation lives inside a single provider's chat product on a single account. The memory is theirs. The model is theirs. The moment a better model ships somewhere else — or that vendor changes pricing, terms, or availability — the founder cannot move her "team" without rebuilding it from scratch. We wrote about exactly this failure mode after the June AI blackout pulled a major model offline globally: the businesses that had hedged kept running; the ones that had not went dark.
A team you rent inside someone else's subscription is not an asset you own. It is a dependency you hope stays online.
Her CFO project interprets sales data she pastes in. It does not connect to her accounting software, pull the numbers itself, and reconcile them. Her marketing project suggests Instagram angles. It does not publish, schedule, or check what performed. Every project is a very smart advisor that cannot touch a single real system.
That is the tell. This is still AI tools wearing name tags — advice on one side of the glass, the founder doing all the actual clicking on the other. It is not a coworker until it can act inside real systems, under real permissions, with a real audit trail.
The contrast got sharp this week, because the same day Business Insider published the soap-maker story, Anthropic published a case study on ABC Legal — a 1,100-person company that moved its people off exactly this pattern.
What the chat-window version looks like: Departments as browser tabs. Memory that lives in transcripts. A founder hand-carrying context between functions. Advisors that cannot touch a system. One vendor, one account, no portability, no version history, no rollback.
What a real AI team looks like (ABC Legal's own words): every agent defined as configuration in a repository — a prompt, a tool list, a schedule, credentials, and memory — so that "nothing about an agent changes except through a pull request someone approves." Every agent has a name, an owner, and a single job. Always-on, not on a laptop. Model choice is a one-line change. There is version history, code review, rollback, and an audit trail.
Same instinct — roles, owners, coordination, human sign-off. Radically different foundation. One is a person heroically holding a system together by memory and willpower. The other is a system that holds itself together and lets the humans steer.
The gap between those two pictures is not talent. Sara Russell is clearly a sharp operator. The gap is the operating layer underneath — the thing that turns a pile of capable pieces into a team that remembers, persists, and stays in your control.
If you have caught yourself doing the chat-window version — separate projects, a manual monthly sync, you as the human glue — you already have the right design. Here is how to put it on a foundation that survives growth.
Do not throw out the roles, the authority boundaries, or the chief-of-staff idea. That thinking is correct and hard-won. The problem was never your design. It was that chat projects cannot carry it. Move the same structure onto real coworkers that persist.
Your departments should not depend on you to relay facts between them. The finance function should be able to see what sales learned. Insist on memory that is shared where it should be, scoped where it must be, and inspectable so you always know what your team knows.
Before you build anything deeper, ask: if I wanted to switch models next quarter, could I keep my configuration, my memory, and my track record? If the answer is no, you are building on rented ground. Model-agnostic is not a luxury feature — it is the difference between an asset and a lease.
A coworker that can only give advice is still a tool. The step change happens when a Teammate can read your accounting data itself, publish the post itself, and follow up on the lead itself — inside clear permission boundaries, with an audit trail of what it did and why.
Sara Russell's best decision was staying the final authority. Keep it. But encode it, do not enforce it by hand. A good AI coworker continues working, pauses when it hits ambiguity, asks for your judgment, and resumes without starting over. That is steerability built into the system, not a founder personally reviewing every chat before anything moves.
Q: Isn't building "departments" out of chat projects good enough for a solo business? A: For interpreting data and drafting ideas, it is a real upgrade over nothing. The ceiling shows up the moment you need those departments to share context, act on real systems, or survive a model change. At one person it feels like a team. At two products or one hire, it reveals itself as a pile of disconnected windows held together by you.
Q: What is the actual difference between an AI tool and an AI coworker? A: An AI tool responds to what you paste in and forgets it when the window closes. An AI coworker holds a standing role, remembers across sessions, coordinates with your other coworkers, acts inside real systems under real permissions, and keeps its track record even when the underlying model changes. Roles and memory and action, not just answers. We break this down further in what makes an AI Teammate more than an app.
Q: Why does it matter that everything lives in one vendor's app? A: Because your memory, your configuration, and your model are all controlled by one company. If they change pricing, terms, or availability — or if a better model ships elsewhere — you cannot move your "team" without rebuilding it. Portability of memory and model is what makes an AI team an asset you own rather than a subscription you are trapped inside. Here is why that lives-in-someone-else's-app problem is bigger than it looks.
Q: Do I have to be technical to run a real AI team? A: No. The ABC Legal case study is instructive here: a 15-person steering committee of non-developers — finance, marketing, operations — built production agents in a week because the hard runtime work was handled for them. You configure roles, boundaries, and workflows. The operating layer handles persistence, memory, and infrastructure.
Q: What should I keep from the DIY chat-project approach? A: All of the organizational thinking. The separate roles, the advise-don't-decide authority boundary, the chief-of-staff coordination, and you as the final sign-off. That design is correct. Only the foundation needs to change — from chat windows that cannot remember, persist, or act, to real coworkers that can.
A solo founder in Brooklyn independently rebuilt the org chart out of chat windows because that is the shape the work wants — roles, coordination, and a human steering the whole thing. She got the design right and hit the ceiling of the medium. If you have built the same thing and felt the same walls closing in, the fix is not more clever prompting; it is putting your org chart on coworkers that share memory, persist through model changes, act inside your real systems, and stay yours to move. That is exactly what Associates AI Teammates is built to run. If you are ready to stop stitching departments together out of browser tabs and start running a real team of AI coworkers, choose a plan and get started at associatesai.team.
Written by
Founder, Associates AI
Mike is a self-taught technologist who has spent his career proving that unconventional thinking produces the most powerful solutions. He built Associates AI on the belief that every business — regardless of size — deserves AI that actually works for them: custom-built, fully managed, and getting smarter over time. When he's not building agent systems, he's finding the outside-of-the-box answer to problems that have existed for generations.
More from the blog
In one week of August 2026, HoneyBook shipped a Claude connector to your client pipeline, Anthropic...
In one week, a supply-store platform raised $30M for an 'AI-native operating system,' a startup laun...
A St. Petersburg business owner cut $41,000 a year by handing his marketing and bookkeeping to a cha...
Want to go deeper?
Get started today. Hire your first Teammate in minutes and put it to work on what you're reading about.
Get Started