Thinking · November 2026 · Barking Cat series
Where does AI actually pay back for a retailer?
AI pays back in three different ways, depending on how much of the business has to change around it. Where little has to change, it saves hours and rarely moves the profit on its own. Where a lot has to change, it can move the profit, but only if the business changes too. And on the customer’s side, it decides whether an AI assistant can recommend you at all. Judging all three on one business case is, I think, part of why so much AI saves time and so little of it shows in EBIT.
The evidence is in two surveys. In IMRG’s Tech & AI Connect 2026 survey, 79.6% of UK retailers use AI for written content and 55.7% to make images; 6.8% use it for dynamic pricing, and 6.8% for one-to-one personalisation. (2.3% said they used it nowhere meaningfully, which at least has the merit of honesty.) McKinsey’s State of AI survey, published on 25 August 2026, found the same shape from the other end: 80% of 1,719 respondents say AI has made them personally more productive, and 6% attribute 5% or more of their EBIT to it.
My reading is that the two surveys describe one pattern: retailers have started with AI where little else has to change, which is where it saves hours without moving the profit. What follows is a way to sort AI investments so that each is judged on the return it can actually produce.
Three kinds of AI investment
The sorting is my framing, not IMRG’s or McKinsey’s.
| Kind | Examples | What changes in the business | What it pays back |
|---|---|---|---|
| Low change | Content, imagery, drafting replies for the service team | Almost nothing: a task that a person did, a model now drafts | Hours. On its own, it rarely moves the P&L |
| High change | Pricing, personalisation, availability | Who decides, what is measured, how data flows | The bigger return, if the business changes around it |
| Customer readiness | Product data an AI assistant can read: specifications, reviews, stock and price, structured and current | What the business publishes about its products | Whether you can be recommended at all |
Low change
This is the 79.6% bar. It is quick, because the business stays as it is: nobody’s incentives move, nobody’s decision rights move, and the data stays where it was. It saves hours, and an hour saved is not a pound saved. The hour becomes a pound only when someone redeploys it, by paying for fewer hours or by putting them into work that earns.
So do it, and judge it on hours, with a named plan for where the hours go. Do not put it in the business case as EBIT.
High change
This is the 6.8% bar, and it is where I think the bigger return sits. That is inference: McKinsey did not test this split, and its survey does not say which uses produced the 6%. Pricing changes who decides the price. Personalisation needs clean customer and product data flowing in real time. Both pay back only if the business changes around them.
The six assumptions in the Barking Cat problem apply in full here. Is anyone rewarded for using it? Is it judged on profit? Do the teams it touches share its target? Can someone act on what it recommends, quickly? Does the business have the skills to run it, or will they sit with the vendor? Will the benefit land on a line someone owns? The machine-learning pricing project in Why did our approved project never go live? failed on the fourth of those, not on the model.
Customer readiness
This is the one the customer decides. Across 168 UK retailers, IMRG measured AI assistants at 0.39% of site traffic, and across 120 retailers that traffic converted at 2.28%, close to organic search. It is small, and it behaves like real demand arriving by a new route.
What matters more is how an assistant shops. It works from what it can read: product data, price, availability, and what others publish about you. At one retailer, the names, dimensions and fabrics a customer needs before buying lived in more than 50 spreadsheets, and the team named it their largest challenge. The data was there. Kept that way, it is hard to keep consistent, hard to publish in a form a machine can read, and hard for anyone, human or agent, to rely on.
Trust is the other limit. ACI Worldwide and YouGov found in June 2026 that 19% of 2,080 UK adults would trust an AI assistant with everyday purchases, against 55% for a human expert. ACI sells payments technology and these are stated intentions, so treat them with care. A business case that needs customers to hand the buying to an agent has that gap to close.
What I believe about AI in retail
Three positions follow from the numbers, and I hold them as positions rather than findings.
AI services existing demand. It does not create it. Nobody asks an assistant to find a product they have no interest in, and the traffic that arrives converts at about the rate of organic search.
Agents expose commodity, not brand. An agent comparing products reads the data, the price and the availability. If what makes you different lives only in the advertising, it will not find it.
Adoption is an operating-model question, not a technology purchase. You can buy the AI. You cannot buy the incentives, measures, alignment, decision rights, skills and plumbing that let it pay back.
How to use it
List every AI initiative in the plan, live or proposed, and put each one in one of the three rows. Anything that will not sit in one row is probably two initiatives.
Then give each row its own business case. Judge low change on hours saved, with a redeployment plan and a date by which a cost or an output changes. Judge high change against the six assumptions, written down before approval, with the weakest one tested first. Judge customer readiness on one question an assistant can answer for you: pick ten products, ask the assistants your customers use about them, and check what they say against your own data.
The last test takes an afternoon. It is worth doing before anything else on the list.
Sources
- IMRG, Tech & AI Connect 2026 (retailer survey and site traffic data; in the project files)
- McKinsey, The state of AI in 2026: On the road to ROI, 25 August 2026, 1,719 respondents
- ACI Worldwide and YouGov, June 2026, 2,080 UK adults
Related: The Barking Cat problem; Why do our digital investments never quite pay back?.
Back to Thinking