You’ve probably seen the demo by now. One sentence of instruction, and an agent has a working online store live before dinner, with products listed and checkout working. Owners get told that’s the hard part done. My view of AI agents for business is that the demo proves assembly is cheap now. Running the store, and keeping its promises, still sits with you.
Canadian law has already tested this. In Moffatt v. Air Canada, a BC tribunal held the airline responsible for what its website chatbot told a customer and rejected the idea that the bot answered for itself.
This is the real lesson in AI agents for business right now. Assembly got cheap, but the standardised part of a store was always the easy part. The shelves, the till, the sign over the door, these come in kits, and a kit is exactly what an agent is good at arranging. What a store actually sells is a set of promises. The price is right. The item is in stock. It ships when the page says it ships. When it arrives broken, someone makes it right. An agent can print those promises on a page in an afternoon. Keeping them is the business, and keeping them still has your name on it.
What does the demo leave out?
It leaves out every moment where the store meets real customers. The supplier who shorts the shipment the week your promotion lands. The customer whose order crossed a border and picked up a fee nobody mentioned at checkout.
The refund that policy says no to and good judgment says yes to, because that customer buys from you four times a year. The review that gets one fact wrong and costs you the next ten shoppers who read it.
None of those moments arrive as tidy tasks. They arrive as judgment calls with your money and your name attached, and they are the part of running a store that was never for sale in a kit. The demo skips them because a demo has to end. Your store does not get to end at the tidy part.
Where does the owner’s share start?
It starts at the first promise the store makes to a stranger. Prices, stock counts, delivery dates, refund terms, every line of product copy that claims a thing the product has to live up to. An agent can draft all of it, and should, because drafting is assembly too.
A person from your firm reads it before it goes live, the same way a person reads a contract before signing it. The reading is not a lack of trust in the tool. The reading is the job. I wouldn’t let any agent publish a price without that read.
Money movement sits on the same side of the line. An agent can prepare the refund, draft the reply, and flag the order that looks wrong. The release of money and the words that go to an angry customer carry the firm’s name, and a person signs both. Draw that line while things are calm, because it is much harder to draw at midnight with a queue of complaints building.
What should an owner hand over first?
The inside work, the jobs that never touch a customer until a person has seen them. Building listings from your real stock records. Drafting descriptions you then correct. Watching stock levels and flagging what runs low. Pulling the week’s numbers into a report you read with coffee.
These jobs are bounded, their mistakes are catchable, and they give the agent a track record you can inspect before it earns anything wider.
That track record is the whole game. I lay out the checks before output counts as done in the action log post, because an agent with real work needs two things: proof its output holds up, and a designed path for the moments it should stop and hand the job back to you.
Which store jobs can an AI agent run this week?
Write two lists. On the first, the jobs in your store an agent may run because a person checks the output before it matters. On the second, the promises and the money, the jobs where a person signs, every time, no matter how good the demos get.
Put one job from the first list into service this week, with the check in place, and let the second list sit where your team can see it. Drawing that line for a specific firm is the first thing I do in AI Agent Architecture and Orchestration work.
AI agents for business, quick answers
What can an AI agent safely run in a small online store?
The inside work a person checks before it reaches a customer. Building listings from real stock records, drafting product descriptions, watching stock levels and pulling the week’s numbers into a report. These jobs are bounded, their mistakes are easy to catch, and they give the agent a track record you can inspect.
Who is responsible when an AI agent makes a promise to a customer?
The business. In Moffatt v. Air Canada, a BC tribunal held the airline responsible for wrong information its chatbot gave a customer. Prices, delivery dates, refund terms and replies to unhappy customers carry your name, so a person reads and signs them before they go out.
Which promise in your store would you want a person to sign before it goes out? Start your second list there. I cover more of these risk calls in the AI Newsletter with ZAK, out Fridays.
4 Comments