There is a version of the AI conversation where the bottleneck is always data. Your data is dirty, your data is siloed, your data is not ready, and here conveniently is a product that gets your data ready. It is a good pitch because it is half true, and because it is the half that can be invoiced.
I have spent four months inside an agentic build that did not start out as an agentic build, and I have come out the other side thinking the industry has the problem one level too low.
It is not data readiness. It is architecture.
Start with the policy, not the pipe
Write the policy first. The policy tells you what the customer experience is supposed to be. The experience tells you what the process is supposed to be. And the process tells you which direction your data is supposed to travel through the business.
Now the useful part. When data travels somewhere the process did not send it, that is not a data quality problem. That is a systems failure with a location and a name. You can point at it. You can fix it at root. You can hand the fix to an agent, because the agent now knows the direction of travel and can tell when something is off it.
That is the whole trick, and it is not a technology trick. Every step of it is a business modelling exercise that could have been done in 2019 with a whiteboard and a bad marker. The difference is that the work that used to take weeks now takes hours, which means for the first time it is actually worth doing properly.
The counter argument, from Cape Town
Hendrik’s response, and he is a supportive critic in the way only a friend can be: agentic is just another channel. Do not build a religion around it.
He is right, and I do not entirely want him to be. But his second point is the one that has stayed with me, because it is structural rather than philosophical. Ecommerce is the stepchild of most organisations. AI is the shiny tool. And shiny tools get handed to development, not to operations.
Which is a problem, because the policies live in ops. The rules live in ops. The context these models are starving for lives in ops. If AI lands in engineering and never reaches the people who know why the returns policy has that specific exception in it, you get a very expensive model producing confident nonsense at speed.
Rules and context. Hendrik’s line was that without both, you have absolute chaos. Stating the obvious, he said. The obvious is doing a lot of work in this industry right now.
Meanwhile, the money
None of this happens in a calm quarter. The consumer is already carrying mortgage debt, auto loans, credit cards and student loans. Tariff refunds have flattered a round of US earnings that would otherwise have read very differently. Walmart has called it the winter of discounts, which is a wonderfully corporate way of saying nobody is buying anything at full price until January.
And underneath, the financing. Shopify Capital reports in two ways, loans and merchant cash advances, and it is the loan number that makes the slide. Roughly four billion out the door last year on top of three billion the year before, with a default rate that has started to move. Not enough for an analyst to flag. Plenty for an SMB trying to turn stock into cash in a quarter where the only lever is price.
That is the context in which everyone is being told to have an AI strategy.
What this means for the merchant on the ground
Not a transformation programme. Compounding growth.
Three more sales a day. Twenty percent fewer errors over six months. Forty percent less contact into the support team over the same period. None of those make a press release. Stacked, over a year, in an industry running single digit margins, they are the difference between a business that survives its own Q4 and one that does not.
That is what the architecture buys you. Not magic. A series of small, boring, compounding wins that you can actually name and measure.
The bit I did not expect
I have called what I am feeling an anxiety of responsibility, and I mean it literally.
The SaaS world sprinted to market with AI because it had to build product. Brands had to prove it first, which made us look slow, and made us right. The anxiety is not about falling behind. It is that when someone hands you a toolset this capable, the expectation arrives in the same box. You get access to information so fast that the temptation is to act at the same speed. The actual opportunity is the opposite: thinking time, which is the one thing operators have not had in a decade.
Next Friday we get into the EU packaging regulations and the proposed two euro charge on cross border parcels, which is either a rounding error or a business model, depending entirely on what you sell.
Stick with us. We are amateurs at this.