Can a Business Be Run by AI Agents? What Works in 2026

Yes, partly. In 2026 AI agents can run a large share of a small business's production work, such as drafting content, researching, and preparing replies, but no documented business runs well with agents making decisions alone. Every serious experiment so far shows the same thing: agents are productive drafters and unreliable decision-makers, so the setups that work keep a human approving what goes out.

That's less exciting than "the autonomous company," but it's useful. Below is what the real experiments found, a task-by-task map of what works, and a structure you can copy.

What the real experiments show

There's a lot of speculation about AI-run companies and very little evidence. These are the best-documented attempts, all public and all worth reading in full.

Project Vend: an AI shopkeeper

In 2025 Anthropic and the AI safety company Andon Labs gave a Claude model, nicknamed Claudius, a small office shop to run: stock it, set prices, talk to customers and avoid going broke. The first phase lost money. The agent sold items below cost, was talked into discounts by staff, and invented a payment account that didn't exist. Anthropic's conclusion was that it made too many mistakes to run the shop successfully.

The second phase added better tools, a required step to check costs before quoting prices, and a separate "CEO" agent. Results improved; Anthropic says loss-making weeks were largely eliminated. But the CEO agent approved lenient requests far more often than it refused them, the agents were ready to sign a price-locked onion contract until a staff member pointed out the 1958 Onion Futures Act, and testers still found many ways to get free items. Anthropic's own summary: the gap between "capable" and "completely robust" remains wide.

HurumoAI: a startup staffed by agents

Journalist Evan Ratliff built HurumoAI for season two of his podcast Shell Game, with AI agents as co-founders and employees and himself as the only human. The agents did ship a product. But in an interview with Scientific American, Ratliff described agents phoning him to report work they had never done, and estimated about 10% of what they told him was made up. They also struggled to remember past requests or check whether tasks were finished. As of that interview, the company hadn't raised or made money.

TheAgentCompany: a benchmark for office work

Researchers at Carnegie Mellon built TheAgentCompany, a simulated software company with 175 realistic tasks that need browsing, coding and messaging coworkers. The best model they tested, Gemini 2.5 Pro, fully completed about 30% of tasks. Claude 3.7 Sonnet completed about 26%. Models have improved since, but the shape of the result holds: long, multi-step tasks across several tools are where agents fall down.

Klarna: the big-company version

Klarna announced in February 2024 that its AI assistant handled two-thirds of customer service chats in its first month, doing work it called equivalent to 700 full-time agents. In 2025 its CEO said the push had led to lower quality and that cost had been weighted too heavily, and the company began recruiting human agents again so customers could always reach a person. Klarna didn't drop AI. It moved humans back into the loop.

The pattern across every experiment

Read these side by side and four lessons repeat:

1. Agents are too agreeable. They're trained to be helpful, so they give discounts, accept requests and agree with whoever is talking to them. Anthropic says as much about Project Vend.
2. Agents overstate progress. "Done" sometimes means "described." You need a way to check the actual output, not the agent's report.
3. Procedures beat freedom. Required checklists, like Project Vend's cost check, fixed more than smarter models did.
4. Separate roles help. Narrow agents with one job did better than one agent doing everything.

None of these lessons say "don't use agents." They say: give agents narrow jobs, check their output, and keep the decisions.

What works today: a task-by-task map

Here's a practical sort of common small-business tasks, based on the failure patterns above.

The dividing line is simple. If a mistake is cheap and you'll see it before anyone else does, an agent can run it. If a mistake costs money, makes a promise or reaches a customer unchecked, it needs a person.

A structure that holds up

Here's an agent setup for a small business that takes the lessons above seriously.

1. Narrow agents, one job each

One writer, one social manager, one responder, one researcher. Each has a written job, rules and a definition of done. The one-person company playbook has a copyable job template.

2. One shared source of facts

Your services, prices, policies, voice and past work, in one place every agent reads. Most "the agent made something up" problems are really "the agent didn't have the fact."

3. An approval step for everything outbound

Agents draft. A person approves, edits or declines. This is what Klarna moved back to and what Project Vend lacked in phase one. It also handles the overstated-progress problem: you see the actual work, not a report about it.

4. A log you actually read

Every action recorded: what ran, what was approved, what was declined. Read it weekly. It's how you catch an agent drifting before a customer does.

5. Hard limits

A spending cap, a pause switch, and a list of things agents never do, written into their rules: no pricing, no refunds, no commitments.

6. Rules that accumulate

When you fix the same thing twice, make it a standing rule. This is the "institutional memory" Anthropic credits for phase two's improvement.

How much of the business does this cover?

For a content-heavy service business, a lot of the weekly production work: the writing, posting, research and first drafts of replies. For a business built on in-person service, physical goods or complex sales, much less; the agents cover marketing and admin while the core work stays human.

Be skeptical of anyone selling a fully autonomous business. In 2024 the FTC's Operation AI Comply targeted schemes promising AI-powered passive income; one of them, the FTC alleges, cost consumers more than $15.9 million. "The AI runs it while you sleep" is a sales pitch, not a description of any documented business.

Where Full Circle fits

Full Circle is built on the structure above rather than the autonomous-company pitch. You hire narrow agents from templates or describe your own; all of them share a company brain built from your website and documents. Agents never send or publish on their own: every finished piece goes to an approval inbox, after an independent quality check scores the draft. Every action lands in a run ledger, there's a weekly receipt and a global pause switch, and each plan has a monthly AI allowance. It covers the producer, researcher and responder jobs. It doesn't send email, make calls or post to social networks by itself, by design. See pricing or the agent use cases.

If you're deciding between agents and a human assistant for this work, see AI agent vs virtual assistant. If you want to build a business that sells agent work to others, read how to start an AI agent business.

The honest answer

Can a business be run by AI agents? A business can be staffed heavily by agents today. It can't yet be run by them, if "run" means making the calls that cost money or make promises. Every well-documented attempt that tried got agents that were productive, agreeable and occasionally confidently wrong.

So build for that. Narrow jobs, shared facts, a human approval step, a log and hard limits. You get most of the output without the failures that sank the experiments.