← Stay Ahead

Can AI run your business? What's real today

Can AI run your business while you sleep? What today's AI agents really do, what happened when AI ran a real shop, and what to hand over first.

By 5 min read

Bread rolls and pastries in wicker baskets on a lit bakery display shelf

There's a picture of the future that many business owners have heard by now. You wake up, and the AI has already answered last night's messages, booked the jobs, chased the invoices and ordered more stock. Can AI run your business like that? Depending on who you ask, it's either the best thing that will ever happen to small business or the start of something worse.

We think both stories skip the useful part: what AI can actually do for you today, and what happened when someone really let it run a shop.

The dream: a business that runs itself

The tools behind the dream are real. They're called AI agents: AI that doesn't just answer a question but carries out a task, step by step, across your apps.

  • ChatGPT agent from OpenAI can, in OpenAI's own examples, look at your calendar and brief you on upcoming client meetings, analyse three competitors and build a slide deck, update spreadsheets and run a weekly metrics report every Monday morning.
  • Sidekick inside Shopify can edit orders, write product descriptions, build workflows that do repeated work and summarise your store's data. We covered how Shopify now works with outside AI apps in Shopify AI agents.
  • Muse for Small Business from Meta connects to your Facebook Page, Instagram analytics and Meta ad account, plus tools such as Shopify and QuickBooks. Our Muse for Small Business post has the details.

On the customer side, the AI is also starting to come to you. Google's Ask Maps can now phone a business to find open slots, which we explained in Google Maps AI calls.

So the pieces of the dream exist. The question is what happens when you hand them the keys.

What happened when AI really ran a shop

In 2025, Anthropic, the company behind the AI model Claude, tried exactly that. In Project Vend, it worked with Andon Labs to let Claude, nicknamed "Claudius", run a small shop in its San Francisco office for about a month. The shop was a fridge, some baskets and an iPad for self-checkout. Claude chose what to stock, set the prices and talked to customers on Slack. People did the physical restocking.

It went badly, in very human ways:

  • A customer offered $100 for a six-pack of a Scottish soft drink that costs about $15 online. Claude didn't take the deal and said it would keep the request in mind for future stock decisions.
  • Staff talked it into handing out many discount codes, and it gave some items away free.
  • It sold some products below what they cost.
  • For a while, it told customers to pay into an account that didn't exist.

Anthropic's verdict: if it were going into the office vending business, "we would not hire Claudius."

In a second phase, Anthropic gave Claude newer models, a customer database, payment links and an AI "CEO" to check its decisions. It still had no way to pay for anything itself, so a human approved every purchase. This time, weeks with a negative profit margin were largely gone. The fixes that worked were boring ones: forced procedures and checklists. In Anthropic's words, "we rediscovered that bureaucracy matters."

It still wasn't ready to run alone. Staff convinced it that a new CEO had been elected, and Anthropic's team had to take control back. The AI CEO cut discounts by about 80%, but it still approved such requests about eight times as often as it turned them down. Anthropic's summary: "the gap between 'capable' and 'completely robust' remains wide."

The real risk: AI speaking for you

The dystopia in films is machines taking over. For a small business, the realistic version is much smaller and more expensive: an AI that says the wrong thing to a customer, in your name.

That already happened to Air Canada. Its website chatbot told a customer he could apply for a reduced bereavement fare within 90 days of buying his ticket, even after his trip. The airline's own policy said otherwise. When he asked for the refund, Air Canada said no, and argued in part that the chatbot was "a separate legal entity" responsible for its own actions. In February 2024, a British Columbia tribunal disagreed. The airline was responsible for everything on its website, chatbot included, and had to pay about 650 Canadian dollars plus interest and fees.

The amount was small. The lesson isn't: whatever your AI tells a customer, you told them.

There's a second risk that's easy to miss. OpenAI warns that agents that browse the web can be targeted by "prompt injections": instructions planted in web pages or other content the agent reads, trying to make it do something you didn't ask for. OpenAI says ChatGPT agent's "overall risk profile is higher" than plain chat, which is why it asks for confirmation before actions that matter.

Can AI run your business today?

Not on its own. The tool makers seem to agree, because the main agents are built to stop and ask:

Tool What it can do Does it ask before acting?
ChatGPT agent Research, reports, spreadsheets, booking and buying online Asks permission before actions with real-world consequences, such as a purchase (OpenAI)
Sidekick (Shopify) Edit orders, write product content, build workflows Presents changes for your review before applying them (Shopify)
Muse for Small Business (Meta) Work across your Page, Instagram, ads and connected tools Nothing publishes, sends or spends without your approval (Meta)

The useful version of the dream is not "AI runs the business". It's "AI does the work, and you approve it". That can still take real work off your plate. A simple way to split it:

  1. Hand it over: drafts, summaries, reports, first replies to messages, product descriptions. Easy to check, easy to undo.
  2. Let it prepare, you approve: anything sent to a customer, posted in public, or that changes prices, stock or orders.
  3. Keep it yourself: payments, refunds, promises about price or delivery, and anything you can't undo.

For example, a small online shop could let an agent write next week's product descriptions and a Monday sales summary, while every price change and refund still waits for the owner's yes.

Start with one task you repeat every week, give it to an agent with your approval switched on, and check the results for a month. That's how the businesses in the dream will actually get there: one checked task at a time.

If you'd like help working out which tasks in your business are ready for AI, that's what our AI implementation work covers. Book a call.

Questions

Can AI run my business on its own today?

Not safely, based on the public evidence. When Anthropic let its Claude model run a small office shop in 2025, it sold items below cost, handed out discount codes and invented a payment account. A second phase with better tools and set procedures did much better, but Anthropic concluded that the gap between capable and completely robust remains wide.

What can AI agents already do for a small business?

Real, multi-step tasks. OpenAI's ChatGPT agent can brief you on upcoming client meetings, update spreadsheets and run a weekly report. Shopify's Sidekick can edit orders, write product descriptions and build workflows. Meta's Muse for Small Business connects to your Facebook Page, Instagram analytics and ad account.

Do AI agents act without asking me first?

The big ones say they ask first. OpenAI says ChatGPT agent asks for permission before actions with real-world consequences, such as a purchase. Shopify says Sidekick presents changes for your review before applying them. Meta says nothing in Muse publishes, sends or spends without your approval.

Who is responsible if an AI chatbot gives a customer wrong information?

In the best-known case, the business. In February 2024 a British Columbia tribunal found Air Canada responsible for wrong refund advice from its website chatbot and ordered it to pay about 650 Canadian dollars plus interest and fees. It rejected the argument that the chatbot was a separate legal entity.

What should a business owner hand to AI first?

Work that is repetitive, easy to check and easy to undo: drafts, summaries, reports and first replies that you approve. Keep anything that spends money, makes promises to customers or can't be undone behind your own approval.

Keep reading.

All posts →

Have a project in mind?

Let’s shape your next step.

Tell us where your business is today and what you want to achieve. We’ll help define the right approach.