How BuyFactory.direct uses AI with human guardrails

I started out selling advertising space for local newspapers around Sydney. In 1991 I left and started a party and equipment hire business in Blacktown, out in Sydney’s west. It grew into a multi-million dollar company with two branches, and I sold it in 2010 when I was forty.

Retirement lasted about twelve months. With nothing to push against I could feel myself going soft, so I went back to work. For the best part of a decade after that I spoke on business and marketing, touring Australia, New Zealand and the UK. The travel was exciting at the start. By the end I was sick of living out of a suitcase.

In 2019 my wife Renoo and I started again, this time in ecommerce. I was forty-nine. BuyFactory.direct is now 90-plus brands across more than 800 specialist stores, shipping to 32 countries out of our own warehouse.

What makes it different sits right there in the name. Buy factory direct. You pay for the product, not for the layers of markup between you and the factory that made it, and it’s often the same factory the big brands use. It arrives tracked and insured, and when something goes wrong there’s a person to talk to.

The other difference is that each store does one thing only. Someone after one product shouldn’t have to wade through a department store to find it.

There’s one more thing, and it sits underneath all of it. We’re a Christian-led company. Renoo is a well-known Christian speaker and preacher here in Thailand, and our faith is what the business actually runs on. In practice that looks boring. You tell the customer the truth, especially when something has gone wrong, and you treat the people who work for you the way you’d want to be treated. I’d rather be judged on whether we actually do that than on saying it.

Nearly all of it comes from Google, through a mix of paid search, Shopping and the pages we write that get found on their own. After that, email to people who’ve already bought from us.

The stores are built for a buyer who knows what they want and is working out where to get it. That demand already exists and our job is to be standing in front of it when it shows up. We’re not trying to talk anyone into something they weren’t already considering.

Where do I wish it came from? Two places.

More of it from people already on our list. A customer we’ve paid for once is the cheapest growth available to us, and it’s also the easiest line in a budget to neglect, because new customers feel like progress and existing ones feel like admin.

And more of it from being named. Google traffic is rented. You pay again every month or it stops. I’d far rather be the store that comes up when somebody asks an assistant where to buy a thing, because nobody can outbid you for that.

That’s the honest gap. I know where the better money is, and the bulk of the budget still goes the other way.

We handed over the day-to-day running of the business.

There are AI managers now across marketing, websites, inventory, sales, customer service and finance. Each one has a written charter: what it owns, what it can decide on its own, what it has to bring to a person, and what it must never touch. They report into command centres, and a central command centre sits above those. The hard rule is that nothing irreversible and nothing public-facing goes out without a person approving it first.

The honest measure of whether that worked is this. Over the past year we’ve gone from 60-odd brands to more than 90, and from around 700 stores to more than 800. We still have the same 15 staff.

What we pulled back was automated customer replies running without a person watching them.

We’d moved support onto a system we built ourselves. From our side the replies looked completely normal. Sent and logged, sitting in the thread. They were never reaching the customer. We found out because customers told us, which is the worst way to find out anything.

So now the first response can be automated and anything needing a judgement call goes to a person. Support staff can pull a customer out of every automated flow the moment something looks off, and the system can’t argue with them.

The biggest change is not where they come from. It’s how much they’ve already worked out before they get to us.

People used to arrive with a question. Now they arrive with a shortlist and an objection they picked up somewhere. The questions reaching support are narrower than they were. They’ll skip "what do you recommend" and go straight to "will this exact one fit my situation".

Assistants are part of that, and yes, they do send us buyers. The awkward part is proving it. A lot of that traffic doesn’t announce itself, and a default analytics setup files a good share of it as direct. So plenty of owners are reading a report that tells them nothing is happening while something is.

What we changed is what the page actually contains. An AI agent can’t be charmed by branding. It reads facts and repeats them. So we write the dull things down. What the product is, what it fits, what it weighs, what it costs, what buyers said about it, and the questions people ask before they commit.

You’re writing for two readers now. The person, and the machine answering on their behalf.

What I’d tell them to do.

Pick the job in your business that repeats and has one right answer. Hand that one over and leave the rest alone for now. Before you do, write a page describing it, including what the thing is not allowed to do without asking you. That page is most of the work, and it’s the part people skip before wondering why the output is wrong.

Then write your facts down somewhere a machine can read them. Prices, measurements, what fits what, real answers to the questions you get asked on the phone. It’s dull work and it decides everything, because the assistant answering your customer’s question is reading something, and if it isn’t reading your page it’s reading a competitor’s.

What I’d tell them to ignore.

Anything that already works. The pull to automate a part of the business that’s running fine is strong and it’s a trap. All you’ve bought yourself is another system to maintain.

Ignore the tool list as well. I genuinely don’t care which model anyone is using. What matters is whether you’ve written down how your business makes a decision, because you can’t hand over a process you’ve never described.

And stop producing the content a machine now answers in one paragraph. Nobody clicks through for it any more. Put that effort into the pages where somebody is comparing two real products and needs to be sure.