Somebody answers for the input not being rubbish.
McKinsey's retail practice published a piece in January about what merchandising looks like once agents run the analysis, and I recognised lot of it from my own year. The number that stuck is 40. Merchants in their survey spend about 40 per cent of their time on low-value work, and McKinsey is specific about what that work actually is: consolidating data systems, repetitive spreadsheet work for performance reporting, and reconciling data across systems that do not talk to each other. Hand that to agents, the argument goes, and the merchant gets 40 per cent of the week back for strategy, for finding good products, for understanding customers and for negotiating with suppliers.
I would like that week back as well. But the reconciling does not stop being necessary just because nobody is sitting there doing it. So who checks that the same product carries the same cost in all four systems? At the moment it is the person whose 40 per cent is being handed back, which is why I read the rest of that article more carefully than I meant to.
Because in a business of this size the data sits everywhere at once. The ERP holds one version, the webshop another, the warehouse a third, and then there is the Excel file somebody built in 2022 that everybody still opens. The same product can carry a different cost, a different stock balance and sometimes a different name in each of them. A person notices, mostly, because a person remembers buying the thing and remembers what it cost. An agent does not notice. It picks one of the values and carries on, and what comes back out looks exactly like analysis.
Going through one business's files this month I found three different figures in circulation for the same month's revenue, none of them labelled with the basis it was calculated on, and all three called revenue. The highest sat 34 per cent above the true net number. It was caught by one subtraction on one row that did not come out. In the same set of files three product codes appear twice in the stock report, and the later row, the one with no location on it, is the one that wins when a machine loads it. Nobody decided that. It is just what the loader does with a duplicate.
Feed either of those into an agent and you get data slop back. Confident, fast, neatly formatted, wrong.
So no, I do not think the thing to worry about is that there will be nothing left for people to do. The responsibility moves, away from doing the work by hand and towards owning the data and the processes underneath it. Which product is this. What does it cost, landed, rather than what the invoice says. Which of the three revenue numbers is the real one and who says so. Is the stock report the primary source for status, or only for quantity. McKinsey's own survey has 61 per cent of merchants saying their organisation is not at all or only slightly prepared to scale AI across merchandising, and I suspect this is most of what that sentence is made of.
That is a job. It is a duller job than the one it replaces and it is the one that decides whether anything built on top is worth having. I have written before about how hard it is to get the decision logic out of a head, and that is the slow problem. The data underneath is the boring problem, and it fails much faster, usually in the first week, usually silently.
And once the data and the processes are in good shape you can build all sorts of things on top of them. That is the part where AI is actually worth something in a product business. Then the agent can go ahead.
Anyway. Back to the stock report with the three duplicate codes in it.
"Merchants unleashed: How agentic AI transforms retail merchandising", McKinsey & Company Retail Practice, January 2026. The 40 per cent of merchant time spent on low-value tasks (data system consolidation, repetitive spreadsheet work for performance reporting, and reconciling data across siloed systems), the up-to-40 per cent of time reclaimable by offloading that work to agents, and the 61 per cent of respondents describing their organisation as not at all or only slightly prepared to scale AI across merchandising are all from that article. Retrieved 21 September 2026. The three revenue figures and the duplicated stock-report codes are from Wren Ventures client work this month, anonymised: no client, no sector and no absolute figure.