What AI Actually Changed About Running a Store, and What It Did Not

By Brennan Lunin, Founder at BL AdWorks

I run a marketing operation using AI tools every day. I also watch a lot of AI projects fail, and they fail in a predictable way: somebody used an agent where a deterministic workflow would have shipped in an afternoon.

So here is the practical version, from someone who runs these things rather than sells them.

The thing to get straight first

AI is not the operator. It is the thing the operator points at a job.

Every useful application I run has a person deciding what the job is, what “good” looks like, and whether the output shipped. Remove that person and you do not get an autonomous marketing department, you get volume with no judgment, which is worse than nothing because it looks like work.

That is not a limitation to be engineered away next year. Taste is the input. The tools amplify whatever taste you bring.

Where it genuinely earns its keep

Drafting at volume. Ten ad creative variations, twenty subject lines, a first pass at product descriptions for a catalogue. Work that used to cost hours of a person’s day now costs a few minutes and a trivial amount in tokens. The human still picks the winners, and the picking is where the value was all along.

Boring transformation work. Reformatting exports, classifying support tickets, turning messy data into a table. Unglamorous and enormous.

Summarising things nobody reads. The weekly report that took two hours to build and got skimmed for thirty seconds.

That is the honest list, and it is a good one. It is not a small thing to get hours back every week.

Where it does not work, and where people burn money

When a deterministic workflow would have done it. This is the big one. If the task has a fixed set of steps and a fixed set of inputs, you do not want a model deciding what to do each time. You want an n8n workflow that does the same correct thing every time, for pennies, forever. Reaching for an agent because agents are exciting is how you end up with an expensive, non-deterministic version of a cron job.

The test I use: can I write down the steps? If yes, it is a workflow, not an agent. Agents are for when the steps depend on what the model finds, which is a much narrower set of problems than the marketing for these tools suggests.

Anything where being wrong is expensive and unverifiable. Generating a number you cannot check is the worst possible use, and it is everywhere. A confident fabricated statistic is more dangerous than a blank page, because you will act on it.

Customer-facing writing with no human in the loop. Not because it cannot write. Because it writes fluently and hollowly, and your customers can tell, even when they cannot say why.

What it did not change

It did not change what makes people buy. Nobody has ever bought a product because the description was generated efficiently.

It did not change the fundamentals of a store: shipping cost visible before checkout, images that answer the question, a buy box that works on a phone, an email flow that fires when a person does something. Those were the levers before, and they are the levers now, and AI makes it faster to build them, not less necessary.

And it did not change who owns the customer relationship. That still comes down to whether you have a list.

The stack I actually run

Named, because vague talk about “leveraging AI” is how people avoid saying anything.

Claude for drafting and analysis. n8n for the deterministic connective work, which is most of it. Klaviyo for the flows. The AI sits on top of that stack and speeds parts of it up. It is not the stack.

What I am not going to tell you

I am not going to give you adoption statistics or projections about what percentage of ecommerce will be AI-driven by some year.

That entire genre of statistic, when I have gone looking for the sources under it, turns out to be pages citing pages, with no study at the bottom. I would rather tell you what I actually run.

The one prediction I will make

The stores that win with this will be the ones that were already good at operations. AI is a multiplier, and a multiplier applied to a mess produces a bigger mess, faster.

Fix the store, build the flows, then point the tools at the parts that are slow. In that order.