You don’t need an AI agent for that.

A customer asks whether you deliver to her area. You answer. She asks for the price. You answer. She sends a payment screenshot. You scroll through yesterday’s messages to find her address. Somewhere between that chat and the next one, an order stops existing everywhere except in your head.

Then someone tells you the solution is an AI agent.

Hold on. What, exactly, are we asking it to do?

There is useful work that automation can take off a small business owner’s hands. There is also a very expensive way to avoid deciding how an order becomes an order. Before you buy a clever assistant, give the ordinary work somewhere to live.

Start with the part you keep forgetting

Consider a fictional cake business. Enquiries arrive through messages. Customers send dates, flavours, reference pictures and delivery details in whatever order occurs to them. The owner asks for missing information, checks availability, sends a quote and waits for a deposit.

The problem isn’t that the owner needs a machine capable of having a conversation. The immediate problem is that each conversation is also acting as an order book, a calendar and a payment record.

A useful first improvement would be one enquiry form that collects the event date, serving size, delivery area and contact details. Each submission creates a record. The owner gets a notification. The customer gets an acknowledgement that says when to expect a reply.

That acknowledgement should not promise a booking. A request is a request. Availability, price and payment still need checking.

Already, there is less information to chase. Nothing in that first step requires a system that invents answers.

The form is part of the same question I asked in Your website is not a flyer: what useful job does this piece of your business actually do?

Write down what happens next

Try describing your process without naming an app. “When an enquiry arrives, save it, tell me and mark it as new.” Good. “When I approve the quote, prepare the payment instructions.” Also good. You can point to the event, the action and the person responsible.

“Handle my customers” is a wish. It leaves too much undecided to build or judge anything useful.

For the cake business, the record might move through five stages: new enquiry, details needed, quote sent, deposit verified, confirmed. A cancelled order needs its own status too. Otherwise a perfectly enthusiastic reminder might go to somebody who has already cancelled.

Decide who moves each record between stages. Decide what counts as proof. A customer uploading a screenshot should not automatically count as a verified payment. The system can flag it for review while the owner checks the actual payment record.

The boring automation might be the best one

Once the stages are clear, you can choose a small piece to automate. Perhaps every new enquiry needs an acknowledgement. Perhaps tomorrow’s confirmed orders need to appear in a morning list. Perhaps an unanswered enquiry needs a reminder for the owner.

These are rules: when this happens, do that. They can be useful without generating a single sentence.

AI becomes more interesting when the input needs interpretation. It might turn a long customer message into a draft brief, suggest a category or prepare a reply for someone to check. That is a different job from copying a known delivery date into a calendar.

Keep the distinction visible. A suggested date extracted from a message is not necessarily a confirmed date. A draft reply is not an approved promise. If a wrong answer would cost you money, embarrass a customer or commit your business to something, put a review step before the action.

Show me what happens when it goes wrong

A demonstration where every field is complete tells you very little. Ask what happens if the customer submits twice. If the email doesn’t send. If someone enters next year instead of next week. If an order changes after the reminder has been prepared.

For duplicate submissions, the system needs a way to recognise the same request or flag it for review. For failed messages, somebody needs a visible warning. For a changed date, the old reminder needs to be cancelled or updated. You should be able to see what happened without opening six different apps.

And a retry should not create another order every time it runs. That is an unglamorous requirement right up until the morning it saves you from three delivery bookings for one cake.

Ask who owns the accounts, where the records live and how you can export them. Use only the customer information the task needs. A person setting up your delivery reminders does not need unrestricted access to every customer conversation you have ever had.

Give one workflow a fair trial

Start with a task you repeat often enough to observe. Write down what happens today: how long it takes, what gets missed and how much checking it requires. Those are your starting points, even if the first record is just a page of notes.

Test the new process with sample enquiries before real customers depend on it. Include missing details, duplicates, cancellations and a deliberately failed notification. Check both the customer’s message and the record you receive.

Then run a small live trial. Count the time spent reviewing and correcting the automation as well as the time it saves. If you spend the afternoon supervising something that used to take ten minutes, you haven’t gained much.

You may discover that a clearer form and a reliable reminder are enough. You may find a genuine need for help interpreting messages. Let the work tell you which one it is.

Buy a better Tuesday

When someone offers to automate your business, ask them to describe an ordinary Tuesday after the work is finished. Which task disappears? Which decision stays with you? What will you look at when something fails?

“You’ll get a list of tomorrow’s confirmed orders without collecting them from separate chats” is an answer you can test. “Your business will be powered by AI” leaves you with a slogan and another subscription.

I build automations. I would still rather help someone remove one persistent headache than give them a complicated system they are afraid to touch.

Pick the task you keep doing twice. Give it a clear starting point, a reliable record and an owner. That is enough work to begin with.