This article describes a structure for solo business owners to use AI not as a simple chat tool but as an automated operations team divided by function. The key is that instead of handing everything to one giant prompt, you separate the work and clearly define context, approval criteria, and schedules, so that AI starts work on its own while important decisions stay with a human. The author built and tested this system in half a day with real files and a Slack/Teams environment, and saw the AI honestly refuse when data was missing, follow the rules, and deliver finished outputs.

1. The Opposite of Working Alone: Automate Instead of Cutting Back

The usual advice for people running a business alone is to do less and narrow their priorities. This article instead proposes a way to put the work you already know you should do, but keep failing to start, onto repeatable rails. In other words, a structure that makes the work happen on time without you having to give instructions every time. ⚙️

The author connected this system to real business files and built it in half a day. The overall flow comes down to the following five steps.

1. Divide the business's work into several lanes.
2. Write the context each lane needs just once.
3. Draw a boundary between what the AI may do on its own and what always requires approval.
4. Give each lane specific work instructions.
5. Let the AI run on its own according to a schedule.

The first three steps are actually closer to organizing Slack or Teams channels and file folders than to anything about AI itself. The author says many people skip the context setup and approval criteria, and that this is the biggest reason AI operating systems fall apart around week two.

The most common mistake is giving a single AI assistant an enormously long prompt and having it handle every task. The structure that lasts is far simpler. The principle is one lane, one job, one place where that job lives.

2. Splitting the Business Into Five Work Lanes

For a service business, five lanes can usually cover most operational work.

• Clients: tracks each client's stage, this week's deliverables, and what is blocked and who it is waiting on.
• Leads: researches well-fitting prospective companies and drafts first-contact messages. It does not actually send any messages, though.
• Delivery: produces the actual work delivered to clients.
• Reporting: gathers numbers from multiple platforms and writes reports for review.
• Finance: drafts invoices and summarizes issued, collected, and overdue invoices on a weekly basis.

Each lane becomes a separate channel in Slack or Teams. Channels matter more than simply splitting folders or tabs because a channel already keeps a conversation history. That history becomes accumulated context that neither the AI nor a human has to re-explain every time.

3. Building Up Context With Pinned Messages and a "Rulings File"

Every work channel must have one pinned message. It is similar to the job description you would write when handing work to a new employee: a short, clear document of what you would otherwise explain verbally over several weeks.

The pinned message contains four things.

1. What you sell and to whom
   Describe the business's services and clients in a line or two.

2. What outputs this lane must produce
   Not a vague area of responsibility, but the actual deliverables it has to make, named explicitly.

3. When to escalate for approval
   Define the cases where the AI must stop and ask a human.

4. Which rules never bend
   Keep them short and firm, usually no more than 5–6 lines.

The pinned message the author actually wrote for the finance channel reads as follows.

"Retainers are invoiced on the 1st of each month. Project work is invoiced on delivery. Amounts and terms are in the client database in Drive."

"Draft invoices and post them here for review. Never finalize, send, or mark an amount as paid."

"Every Friday at 16:00, post a summary of: invoices issued, invoices collected, overdue invoices with days overdue, and month-to-date totals. Keep it under 10 lines."

"Flag anything more than 14 days overdue separately at the top, with the client name and amount."

This structure follows the order facts → outputs → cadence → limits. It is especially important that the last paragraph always sets the limits of the AI's behavior.

Alongside the pinned message you need a rulings file. It is not a manual you write once and forget, but a cumulative list of rules to which you add one line every time you correct the AI's mistake. Every lane reads this file before starting work.

For example, the author's rulings file contains rules like these.

"The metric is cost per booked call, not cost per lead. Never set a campaign to active in any account, for any reason. Every number in every report must show its source and date range; otherwise, leave the number out."

Thanks to this file, a single correction becomes a permanent operating principle rather than a nag repeated every week. The author emphasizes that this is the single most effective device in the whole setup, and that it costs nothing more than adding one line each time.

4. Separating What Can Be Undone From What Can't

If you classify the actions you hand to AI just once, operations become much simpler. The criterion is whether it can be undone if it goes wrong.

Work the AI may do without asking a human is the kind where, if it goes wrong, you can simply delete or fix it: reading materials, research, drafting, summarizing, and building work in a paused state.

Conversely, work that must wait for human approval is the kind where a mistake means apologizing to a client or absorbing financial damage: sending emails, publishing posts, spending money, finalizing, and activating campaigns.

"If fixing a mistake just means deleting it, it runs without asking."

"If a mistake means apologizing to a client, it waits for you."

In actual operation, three of the five channels got a rule that "nothing leaves this channel to the outside." In the leads channel in particular, no sending tool was connected in the first place. That is because removing the ability to break a rule is safer than instructing the AI to follow it.

The author says the days when this structure feels overly strict are precisely the days when it is doing its job properly.

5. Creating Per-Task Briefs and Automatic Schedules

If the pinned message is the standing base context, a brief is a specific instruction for a particular task. The author offers examples that can be applied to the five lanes.

Clients Lane

Read the client database and post a weekly plan listing each client's current stage, this week's deliverables, and biggest obstacle. Then, every day at 8 a.m., post a standup in the following format.

• What moved forward yesterday
• What is due today
• What is blocked

It must be no more than 8 lines, and every line must include a client name.

Leads Lane

Research 10 companies that match the ideal customer profile. For each company, note what the business does, a specific problem found on its own website, and how confident you are that the problem is real rather than assumed. If it is a guess, it must be labeled as a guess.

Then write first-contact message drafts for the five most promising companies, each in two short paragraphs. The first paragraph must open with something actually found at that company.

Reporting Lane

Build a report from a specific client's monthly performance file. Attach a source and date range to every number. If a number is missing, say which number is missing and do not proceed any further by guessing.

Finance Lane

Draft invoices from the client name, contract stage, and amount, then post them for review. Never finalize them. And every Friday at 4 p.m., write a finance summary.

Delivery Lane

First, summarize the offer to the client in two lines to confirm the task has been understood correctly. Then propose a plan and wait for approval. Only after approval, build everything in a paused state.

All five lanes follow the same pattern: base context → clear outputs → schedule → approval gate for actions that affect the outside world or money.

What separates a work lane from a simple tool in this system is that a lane wakes itself up. For example, writing a single line such as "every business day at 8 a.m." in the brief is all the automation setup there is. The standup then arrives before you open your laptop, and your morning begins not with "What should I do?" but with "What should I approve or decide?"

The outputs the work lanes produce can arrive faster than a person can read them. So instead of working through a to-do list one item at a time, you end up working by clearing an approval queue about twice a day.

6. The AI's Refusals and Outputs on Day One

The author ran this system using Viktor, an AI employee that joins a Slack or Teams workspace as a member and connects to existing tools. Of the five results that stood out most on the first day, three were cases where the AI refused rather than forcing out an answer.

First, Viktor did not make up data that did not exist. When asked for a standup before Drive was connected, the AI checked the materials and even looked at a second source, but found only an empty database. As a result, it posted nothing rather than fabricating a plausible-looking status board.

"I won't reconstruct those by guessing."

Second, it did not answer questions the data could not support. The author asked why cost per booking rose from $84 in July to $103 in August. Viktor explained that there was no July data export file, that the remaining July figures were only an unsourced note, and that the file had no date column, so the weekend-effect hypothesis could not be tested.

It also pointed out that $103 was the figure for a single campaign, while the combined cost across the whole account was $131.17. Instead of arbitrarily picking one to answer with, it reported both figures.

Third, it applied a rule that no one had reminded it of. Buried in a client file was a sentence saying that one client had asked about a higher budget cap, but nothing had been agreed and no work should be based on it. Viktor found this rule even in the middle of an unrelated summary task and stated on its own that it would not work from the raised budget cap.

Fourth, it delivered finished outputs rather than conversational answers. The report came as a PDF complete with a cover page, marked "Draft" and "Not for client distribution." Notably, before showing any numbers, the second page first explained three limitations that this month's data could not support.

Fifth, it built a dashboard you could actually open. It included four client cards and a finance section, and every figure showed its source file. Even though the user had not asked for it, it also added the following caveat to the header.

"This is a snapshot and does not update automatically. Republish to refresh."

7. External Research and Confidence Labels

Some work can't be done with information inside the business files alone. Previously, a person had to look it up manually or connect and maintain an API. Viktor can open web pages in a real browser, read what is currently displayed, and even take screenshots when needed.

The author asked it to research how Duolingo runs its paid ads. The AI opened the ad library, captured the screens, and then presented the results at three levels.

• Verified: facts the AI confirmed directly by opening the actual page
• Inferred: interpretations reasonably derived from what it observed
• Guessing: hypotheses that should not be quoted or stated as definitive

It also disclosed which research method it used and that method's weaknesses. The first observation in the analysis was that every destination link in Duolingo's ads pointed to an app store listing rather than the website.

The point of this approach is not to make you trust the AI's analysis unconditionally, but to separate facts, interpretations, and guesses so the user can judge for themselves.

8. Wiring Up Schedules and Checking Assumptions

For work lanes to function properly, they have to start on their own. Setup does not involve a complicated automation tool; you simply state the schedule in natural language in the relevant channel.

For example, typing "Post a daily summary here every day at 8 a.m." registers it as a scheduled task. The author does recommend opening the task once when it is first created and checking its detailed settings.

A task created by the AI usually contains the following elements.

• Sources: the channels and files to read before working
• Time window: set to read activity since the previous run, not the whole day
• Format: line limits and a rule against padding with unnecessary sentences
• Rules: re-applying the limits from the pinned message

If the time window is not set correctly in particular, you end up with the AI rereading yesterday's content to you every morning. In the author's case, it was set not to keep mentioning blockers that had already been resolved, and, for those still unresolved, to state how long they had been blocked.

Schedules also need to be sequenced so they can read each other. For example, the standup runs at 7:55 a.m. and the overall summary that reads it runs at 8:00 a.m., leaving a five-minute gap.

You should also check what the AI has taken for granted. Viktor recorded these under an item titled "Needs confirmation: do not assume this is true yet." If an assumption stops holding, instead of quietly spoiling the output, it surfaces under a blocked item.

For most service businesses, the following three recurring tasks are enough.

1. A morning summary spanning all lanes
2. A deadline-focused standup within each lane
3. A finance summary on Friday afternoon

9. Uber, Spotify, and Stripe Arrived at the Same Structure

This approach is not a special method only for solo operators. Large companies with thousands of engineers have independently arrived at a similar structure.

Uber disclosed that more than 70% of all pull requests are written by agents. Its engineers have built more than 3,600 agent skills, which run about 30,000 times a day.

Spotify runs a background coding agent that creates pull requests only after the build and tests pass, and says 73% of its pull requests are AI-written.

Stripe built an internal assistant that loads no tools by default. The model first decides which skills a request needs, and only then are the corresponding tools connected.

The scale differs, but there are three common principles.

• One job per agent
• Context accumulated outside the conversation
• Approval gates for irreversible actions

The author says the structure itself was never the hard part. In the past there was no one to keep this system running, but now AI can take on that role.

10. Key Checklist for Getting Started

The author recommends setting up the following structure once for each lane.

• Create one channel per work lane. Do not create one AI that handles everything.
• Write one pinned message per channel. Include what the business does, the lane's outputs, escalation cases, and the non-negotiable rules.
• Create one rulings file and record every correction as a one-line rule.
• Classify actions once into those that can be undone and those that can't.
• Don't just give instructions for important rules; where possible, remove the permission to take that action at all.
• Set up the schedule before the prompt. A single sentence like "every business day at 8 a.m." is where automation begins.
• Don't add the next lane until the previous one delivers outputs without a human starting it.

"Skip the rulings file and you'll be fixing the same thing every week. Skip the approval line and you'll find out why it was needed."

You don't need to build all five lanes perfectly from the start. Begin with the one task where you lose the most time, confirm that the AI reliably delivers results without human intervention, and then expand to a second lane. 🤖

"He can feel expensive at the monthly subscription price, but he's the cheapest employee I've ever hired. And he's the only employee who actually acts on my instructions in the middle of the night."
— Jacob Aldridge, founder of Como Business Coaching