AI saves real hours, but the money appears only when finance can see where they went

Picture a tool that saves each of ten people four hours a week. Together that is 40 hours, a full-time job. But it arrives as 48 minutes a day per person, and those minutes go into email and meetings. Payroll does not change, overtime does not change, and finance sees only a new software licence. That is why fewer than one in five finance chiefs get significant value from AI bought to save time.

Ikea decided where the hours would go. Its chatbot, Billie, came to handle 47% of customer queries, saving about €13 million, and between 2021 and 2023 Ikea retrained 8,500 call-centre staff as remote interior design advisers. The remote sales centres where they work brought in €1.25 billion last year. Next to that figure, the saving looks small.

A single team can make the same choice. Many suppliers give 2% off for payment within ten days, and an accounts payable team with slow approvals misses most of those discounts. If the hours that AI frees go into approving invoices sooner, a company that buys €5 million a year on those terms keeps €100,000 a year, on a line that finance already reads. Choose that line before the tool goes live; once it is running, the hours will already have gone into email.

The agents tried more than 1,200 sites and worked unsupervised for about 23 hours on each break-in. The whole campaign used about $5,000 of AI services, so each successful break-in cost $50 and each attempt about $4.

At $4 a try, an attacker no longer needs to choose targets, just as a spammer does not choose addresses. So the site that falls is often one that nobody looks after: at SharkNinja, it was a separate product support site.

List every website that carries the company's name, including those that vendors run, and close the ones nobody uses.

Ask an assistant how many loaves a bakery will sell tomorrow, and it gives the number it expects on an average day. Baking that number is right only if an unsold loaf costs as much as a missing one, and it rarely does. A loaf that costs €1 to make and sells for €3 loses €1 when it is thrown away, and €2 of profit when a customer leaves without one. Weighing those two costs against each other tells the baker to sell out on only one day in three, which means baking more than the average day.

Stock orders and shift rotas work the same way. The manager knows what each kind of mistake costs and the assistant does not, so put both costs in the prompt and ask how many to order or to staff.

The chatbot told a customer that he could claim a bereavement discount after his flight, although another page of the airline's website said otherwise. In February 2024, a Canadian tribunal held Air Canada responsible for all the information on its website.

So a correct policy page elsewhere protects nobody, and a customer chatbot should answer only from that page. Reports follow the same rule: in 2025, Deloitte partly refunded the fee for a A$440,000 report that cited research papers that do not exist. Before a report leaves under your name, someone should open every source it cites.