Before you buy an AI agent, check what it can actually read

Anthropic has released an open-source blueprint and a merchant guide for building Claude agents that serve customers on a business’s own channels and do the unglamorous back office work behind them, covered by SmartCompany. The reference build handles product recommendations, payments and inventory lookups, and you can deploy it through the Claude API, Amazon Bedrock or Shopify.

Most of the coverage this week is a launch summary. Here is the part nobody writes up, and it is the part that decides whether any of this works for you.

The agent is the easy part

An AI agent is a program that can hold a conversation with your customer and then go and do something about it. Look up whether the part is in stock. Check what a job like theirs costs. Take the booking. Flag the order to you.

Every one of those actions is a question the agent asks your business. And it can only ask questions of something that exists and can answer.

That is the whole catch. When a build like this goes badly, it almost never goes badly at the AI end. It goes badly because the agent asked “do we have that in stock” and there was nothing to ask. The stock lives in a spreadsheet on one laptop, and the real answer lives in the head of whoever does the ordering.

What an agent actually needs to read

Three things, and none of them are AI.

Clean product or service data. One list of what you sell, with current prices, written in a way a machine can parse. Not a PDF price list from three years ago, not a page of prose, not three versions that disagree. If a new staff member could not work out your pricing from what is written down, an agent has no chance either.

A real booking or ordering system. Something the agent can write into, not just read from. An enquiry form that lands in an inbox is a message, not a system. If taking a booking still means someone opens a calendar and types it in, the agent can only ever hand work to a human, which is roughly where you started.

One source of truth. This is the one that catches people. Most small businesses have the same information in four places: the website, the quoting spreadsheet, the accounting software, and the owner’s memory. That is survivable when a person is in the middle, because a person notices when two numbers disagree and rings someone. An agent does not notice. It confidently quotes whichever number it was pointed at.

Why this is good news, mostly

If you read that list and recognised your business in it, that is not a reason to sit this out. It is a reason to do the boring work first, because that work pays whether or not you ever switch an agent on.

A clean service list means your quotes go out faster and match each other. A real booking system means the enquiry that comes in at 9pm on a Sunday is still there on Monday, with the details intact. One source of truth means you stop discovering that the website has been advertising last year’s price for five months.

Those are ordinary business improvements. They just happen to also be the exact prerequisites for the AI thing everyone is excited about. That is a rare piece of luck, and worth taking.

The honest sequence

If you want to end up with an agent that is genuinely useful, the order goes like this:

1. Pick one thing you sell and get its data straight. Not the whole catalogue. One line of work, priced properly, written down once.
2. Give it a system that can be written into. A booking tool, an order flow, a quoting form that produces a record rather than an email.
3. Kill the duplicates. Decide which place is the truth and make everything else read from it.
4. Then add the agent and point it at that one clean thing.

Most businesses want to start at step four, because step four is the one that is fun. Steps one to three are the ones that make step four work, and they are the reason the AI project you heard about at a barbecue quietly stopped being mentioned.

Where this leaves a Melbourne small business

You do not need Bedrock, Shopify Plus or a developer on retainer to act on this week’s news. You need to be able to answer one question honestly:

If a customer asked us something about price, availability or timing right now, is there one place that holds the correct answer, and can a computer read it?

If yes, you are closer to an agent than you think, and the build is genuinely a few weeks of work.

If no, you have found something better than an AI project. You have found the reason your enquiries take two days, your quotes disagree with each other, and your team keeps checking with each other before answering a customer. Fixing that returns hours immediately, and it leaves you standing on the exact foundation the next few years of this technology are going to need.

That is the work we do with Melbourne businesses: looking at how your site, your enquiries and your day-to-day systems actually fit together, then getting the data and the plumbing into a shape that can carry automation instead of collapsing under it.

If you want a plain read on where your business sits on that list, book a free digital systems audit. No jargon, no hard sell, just an honest look at what an agent could read today and what it could not.