# Why do our digital investments never quite pay back? > 80% of survey respondents say AI has made them more productive; 6% attribute 5% or more of EBIT to it. Why the hours arrive and the profit does not. Source: https://barkingcat.ai/thinking/why-digital-investments-never-pay-back/ ## Thinking · November 2026 · Barking Cat series Because the business case assumes a business that does not exist. A digital investment pays back only when the business around it changes: how its people are paid, what they are measured on, who is allowed to decide, and where the benefit lands. When those stay as they were, the technology can be delivered in full and the return still never arrive. The evidence for the pattern is now unusually clear. In McKinsey’s State of AI survey, published on 25 August 2026, 80% of 1,719 respondents say AI has made them personally more productive. 37% report any EBIT impact at all, about where that figure stood a year earlier. 6% attribute 5% or more of their EBIT to it. AI is only the newest case of an older pattern in retail: the investment is delivered, and the return never quite shows up. ## An hour saved is not a pound saved The arithmetic is worth doing once, with illustrative numbers. Suppose a new tool saves ten people two hours a week each. Over 52 weeks, that is 1,040 hours. At an illustrative loaded cost of £30 an hour, the business case books a saving of £31,200 a year. Not one of those pounds has moved yet. The £31,200 becomes real only when the freed hours are redeployed, and that has to be a requirement of the business case, not a hope attached to it. Before the tool is approved, the case should say where the 1,040 hours will go and who will move them: either fewer hours paid for, or the same hours put into named work that earns. If the case cannot say that, it has not found a saving. It has found some time, which the working week will absorb without being asked. That is my reading of the McKinsey numbers, not McKinsey’s: an hour becomes a pound only when something else in the business changes. ## The business has a constitution Milton Friedman had the image for this in a Newsweek column in February 1973. A reader agreed that a regulator was performing badly but thought the answer was better people. Friedman’s reply was that the behaviour followed from how the institution was built: “I would like to have a cat, provided it barked.” A business behaves according to its constitution: how its people are paid, what it measures, how its teams’ targets fit together, who is allowed to decide, the skills it has, and where its money and data actually flow. A business case that assumes a different constitution is asking for a bark. It makes the same request of the customer. The case assumes customers will arrive through the channel the business built and buy what it planned. Online sales in Great Britain were 19.2% of retail sales in 2019, 37.8% in January 2021 with the shops shut, and 28.3% in September 2026, on the ONS series J4MC. A plan written in 2021 is worth checking against all three. ## Four explanations, and the one that leads When a return fails to arrive, there are four explanations on offer. The technology did not work: sometimes true, and the easiest to check. The wrong people were in charge, so the business changes the agency, the platform or the ecommerce director (speaking as a former ecommerce director, I can confirm this is a very popular option); in my experience it works only when the new person also changes the targets, the approvals or the systems. The forecast was optimistic: that explains a disappointing return, not a delivered project whose benefit appears in no line at all. The fourth is the one I keep arriving at: the business case assumed a business that does not exist. Two projects from my own record make the point. A machine-learning pricing model was approved and never went live, stopped by shelf labels, systems, a media plan and an unconvinced Commercial team, none of which was about the model. A payment routing project was delivered with £3m a year of savings in its business case, which never appeared in the ecommerce P&L. I agreed to that one. What would prove this wrong is a business whose six assumptions were written down and tested before approval, and whose investment still failed to pay back. ## Why boards keep asking for the bark In my experience, technology is sold on an improvement in conversion rate, and a business case is written to win approval. My reading, and it is a reading: buying technology is a decision a board can make in one meeting. Changing incentives, measures and decision rights is a year of uncomfortable conversations. So the business buys what it can buy, and hopes the constitution follows. ## What to do before the next approval Write down what the case assumes, as a matter of fact rather than as a risk register. There are six assumptions I most often see it make about the business, and five about the customer; [the Barking Cat problem](https://barkingcat.ai/thinking/the-barking-cat-problem/) sets them out, with the question to ask about each. Then test the weakest one first. Two outcomes are honest. Rewrite the case for the business you actually have, and expect what it is built to deliver. Or change the one thing that would stop the return, before you spend. And make the benefit checkable. Name the line it will land on and the person who owns that line as a condition of approval. Where the case counts hours saved, the condition is a redeployment plan: where the hours go, who moves them, and the date by which the cost or the output changes. Agree the baseline, matched for season, promotions and stock, the date you will read the result, and the result that would make you stop. The decision for the board is which of three things it is doing: approving the case as written, rewriting it for the business it has, or changing the business first. The question for Monday is simpler: which assumption in our current business case has nobody tested? This piece rests on one operator’s record and published surveys. It cannot tell you which assumption is costing your business most, and I have not found anyone who has measured which costs most across retail. ## Sources - McKinsey, The state of AI in 2026: On the road to ROI, 25 August 2026, 1,719 respondents - Milton Friedman, “Barking Cats”, Newsweek, 19 February 1973 - [ONS, Retail sales, Great Britain (J4MC)](https://www.ons.gov.uk/businessindustryandtrade/retailindustry/timeseries/j4mc/drsi), release of 23 October 2026 ## About the author Steve Webster is an independent adviser to the boards and investors of UK retail and consumer businesses from £20m to £1bn: an operator’s judgement on ecommerce, technology and AI, before the board commits the capital. Formerly Global Ecommerce Director at Barbour, with 27 years running retail P&Ls up to £240m. ## Contact steve@barkingcat.ai Barking Cat Advisory Limited. Registered in England and Wales, company number 17338003.