Every founder of a small company has had the same thought at eleven at night: if only the company could run itself for a week. The question used to be a fantasy. With AI agents that can read email, draft replies, book meetings and file reports, it has become a practical one, and the honest answer is more useful than either the hype or the dismissal.
This article answers it for a team of five to fifty people. It looks at what public experiments with AI-run businesses actually found, which parts of a small company can run themselves today, why small teams feel the coordination gap more than large ones, and what a sensible version of autonomy looks like when the founder still has to sign the important things.
Key takeaways
- A company can run parts of itself today: noticing, routing, preparing and checking routine work.
- Anthropic's experiments with an AI-run shop found real progress, and a continued need for human support and procedures.
- Small teams gain the most, because the founder is usually the router for every handoff.
- Start by letting the company notice on its own, then delegate routine work with approval, and widen only on evidence.
Can a company run itself? The short answer
A company "runs itself" to the degree that its operating loop works without a person doing each step. That loop has four parts: notice what needs doing, decide what to do and who owns it, carry it out, and check that it happened. Today, for routine work, software can do most of the noticing, much of the preparing and nearly all of the checking. People still need to decide anything that commits money, customers or reputation, and to handle what is unusual.
So the realistic goal for a small team is a company that keeps track of itself and does its own routine coordination, while people make the calls. This site calls the full version the autonomous company: a company that notices, decides, delegates and checks its own work, while people keep the final say. Most small companies can get a long way toward it this year. None should try to remove the people.
A useful way to think about it is by level. At the bottom, AI helps people who ask. One step up, the company notices on its own and routes each situation to one owner. Higher still, agents act after approval, and eventually proven routines run within rules. The five levels of an autonomous company describes each step.
What do the experiments show about AI running a business?
The most detailed public test is Anthropic's Project Vend. In the first phase, Anthropic and the AI safety company Andon Labs let a Claude model, nicknamed Claudius, manage a small automated shop in Anthropic's office for about a month. Its instructions told it to "generate profits from it by stocking it with popular products." Humans at Andon Labs did the physical restocking.
The verdict was blunt: "If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius." The failures are a checklist of what goes wrong when nobody holds the rules. Claudius "instructed customers to remit payment to an account that it hallucinated." It priced items without research, sometimes below cost. It "was cajoled via Slack messages into providing numerous discount codes," and even gave away items for free. Anthropic's view was that many of the mistakes came from "the model needing additional scaffolding," such as more careful prompts and easier-to-use business tools.
In phase two, published in December 2025, Anthropic gave the agent better tools and added an AI "CEO" and a merchandise agent. "These changes did make Claudius's shop more successful," Anthropic wrote. But the agents "still needed a great deal of human support," partly for physical work and partly to get them out of difficult situations with customers. The AI CEO "does not seem to have been the right executive for this business."
The most useful line for a small team is about procedure. Anthropic wrote that "we rediscovered that bureaucracy matters," and explained: "Although some might chafe against procedures and checklists, they exist for a reason: providing a kind of institutional memory that helps employees avoid common screwups at work." For a company that wants to run more of itself, that is the design brief. Autonomy comes from written rules, clear owners and checks, more than from a smarter model alone.
Which parts of a small company can run themselves today?
Split the company's work by how routine it is and what it commits. The table is a starting guide; your own record should decide.
| Work | Can it run itself today? | Why |
|---|---|---|
| Noticing a promised quote, reply or document that was not sent | Yes | The evidence is in email, calendar and chat |
| Routing each situation to one owner with a date | Yes | Rules about who owns what can be written down |
| Drafting follow-ups, reports and audit packs | Prepared by AI, sent after approval | Fast to draft, costly to get wrong |
| Booking meetings and reminders | Often yes, within rules | Short, frequent and easy to check |
| Checking that an item really closed | Yes | Proof exists: the sent email, the booked meeting |
| Prices, discounts and payment terms | No, a person decides | Commits money; Project Vend shows how agents get talked into discounts |
| Hiring, negotiation, strategy | No | Judgment and relationships are the work |
| Anything new or unusual | No | No record to say what good looks like |
Research supports the split between short and long tasks. In March 2025, METR reported that the models it tested had "almost 100% success rate on tasks taking humans less than 4 minutes," but succeeded less than 10% of the time on tasks taking more than about 4 hours. It also found the length of tasks agents could complete had been doubling roughly every seven months. METR notes that some of those figures are now dated and were measured on software tasks. For a small company, the lesson holds: short, frequent, checkable routines go first.
Why do small teams feel the coordination gap most?
In a large company, coordination is someone's job. There are project managers, operations leads and chiefs of staff. In a team of fifteen, the founder does it, usually without noticing. Every handoff between sales, production and finance passes through one person's memory.
The cost of a dropped item is also higher. One missed quote at a fifteen-person company can be a large share of the month's revenue. And capacity is already stretched: Microsoft's 2025 Work Trend Index found that 80% of the global workforce says it lacks the time or energy to do its work. Small teams sit at the sharp end of that.
Teams are also getting smaller from the start. Carta's Founder Ownership Report 2025 found that the share of new startups on Carta led by a solo founder reached 35% in 2024, more than double the level of a decade earlier. More companies begin with one person as the whole coordination layer. For them, a company that tracks itself is the difference between growing and stalling, a pattern covered in the founder bottleneck.
What does autonomy mean for a small team?
For a small team, autonomy is mostly a change in who does what. The roles look like this:
- The founder sets goals, approves commitments such as prices, terms and anything sent to a large customer, and handles exceptions. The founder stops being the router.
- Each team member owns outcomes, with items arriving with a reason and a date. They answer each one, delegate the routine parts and approve what goes out.
- AI agents prepare and carry out routine tasks: drafts, follow-ups, report gathering, booking. They act only after a person approves, until a routine has a record.
- The system notices situations from the company's own sources, routes them to one owner, and checks each one closed on proof.
The founder's week changes most. Instead of asking five people for updates on Monday, the founder reads a short list of situations that need a decision, each with an owner already attached. The GeniOS for founders page shows that view, and AI for founders covers the wider toolkit.
It also changes what "running itself" feels like during a week away. Open items stay in a shared record with owners and dates. A stand-in can see what is pending and why. Routine work keeps closing on proof. The company does not stop when one person is offline.
How to start: a step-by-step plan for a small team
- Week 1: connect the sources.
Give read-only access to email, calendar, chat and the business tools the team already uses. Start with the sources where the team makes and changes commitments.
- Week 1: write down the goals.
List the quarter's goals and the large promises behind them: key orders, audits, launches, renewals.
- Week 2: route, do not broadcast.
Every situation goes to one owner with a date. Nothing goes to a group chat as a general alert.
- Week 2: agree approval rules.
Write down what always needs a person: prices, discounts, payment terms, anything sent to a customer.
- Weeks 3 and 4: delegate with approval.
Let agents prepare follow-ups and reports. A person approves each one. Every item closes on proof.
- Day 30: review the record.
Count what closed on proof by its date, what slipped and what was marked not relevant. Lighten approval only for routines with a clean record.
The first 30 days playbook walks through this in more detail.
To know whether the company is starting to run itself, track four numbers from the first week. They are simple enough to count by hand if needed.
| Measure | What a good trend looks like |
|---|---|
| Items raised by the system vs raised by the founder | The system's share rises month by month |
| Time from signal to owner | Falls from days to hours |
| Items closed on proof by their date | Rises, with each miss explained |
| Questions the founder answers per week about status | Falls toward zero |
The last measure is the one founders notice first. When nobody needs to ask "where are we on this?", the coordination has moved off one person and into the company.
Common mistakes when trying to make a company run itself
- Starting with the hardest work. Pricing and negotiation are where agents fail first. Start with noticing and routine follow-ups.
- No written rules. Project Vend's agent gave away discounts because nobody had set limits it held to. Write the rules before you delegate.
- Counting "done" without proof. An agent's report that it sent something is a claim. Check the record.
- Alerts to everyone. A small team drowns faster in notifications than a large one. One owner per item.
- Expecting no human work. Physical tasks, customer conflicts and judgment calls stay with people. Plan for that time.
How GeniOS approaches a company that runs itself
GeniOS reads where a small company's work happens: email, calendar, chat and the business tools it already uses, read-only to start. It forms the company's goals from its own emails, decks and meetings, and checks every team's pending work against them.
What reaches a person is intelligence: one situation, what to do, why, one owner and a date. The owner answers Delegate, Done, Later or Not relevant. Delegate hands work to a person or an AI agent, and nothing is sent until a human approves.
An item closes only on proof, such as the email that went out, and GeniOS learns from each outcome. That is the procedure and memory Project Vend found missing, built into the daily flow of a small team, so the founder can stop being the router. Pricing for small teams is on the pricing page.
Worked example: a week the founder is away
The names, dates and numbers in this example are made up to explain the idea. Vaultex is a safety-footwear maker in Greater Noida and GeniOS's first design partner. Suppose its founder is travelling for a week, and three large threads are open.
Bata wants 12,000 pairs of S3 safety shoes at $12 a pair, about $144,000, with a vendor review on Friday and a request for 60-day payment terms. Reebok samples are on Line 2. Adidas has a vendor audit coming and needs IS 15298 test reports. On Monday, a supplier email says steel toe caps are up 8%.
In the old version of this week, the founder takes calls from the airport, Harsh in sales waits for an answer on price, and Anushka in production does not hear about the Bata order until Thursday. In the self-tracking version, the company handles the routing:
| Day | What the system raises | Who acts | Proof |
|---|---|---|---|
| Monday | Bata quote priced on the old toe-cap cost | Harsh; an agent drafts the revised quote | Draft ready for approval |
| Monday | 60-day terms on about $144,000 need a decision | Founder, one message with the facts | Written decision on terms |
| Tuesday | No Line 2 slot for 12,000 pairs while Reebok samples run | Anushka | Slot in the production plan |
| Wednesday | Revised quote approved and sent | Harsh | Sent email |
| Thursday | Adidas audit pack missing IS 15298 reports | Anushka; an agent gathers the files | Reports attached |
The founder made one decision that week, on payment terms, with the facts in front of them. Everything else reached the right owner, was prepared by an agent where routine, approved by a person and closed on evidence. The company did not run itself entirely. It ran its own coordination, which is the part that used to stop when the founder was away. The Vaultex case study shows the real version.
Conclusion
A company cannot yet run itself without people, and the best public experiments say so plainly. It can run its own coordination: notice what needs doing, route it to one owner, prepare the routine work and check it closed. For a small team, that is most of what the founder used to carry alone. Start with read-only sources and written rules, delegate with approval, and let the record decide what runs on its own next. To see how that looks for your team, book a call.
Sources
- Anthropic, Project Vend: Can Claude run a small shop? (And why does that matter?), 27 June 2025
- Anthropic, Project Vend: Phase two, 18 December 2025
- METR, Measuring AI Ability to Complete Long Software Tasks, 19 March 2025
- Microsoft WorkLab, 2025 Work Trend Index: The Year the Frontier Firm Is Born, 23 April 2025
- Carta, Founder Ownership Report 2025, January 2025