We have priced the hard part at zero

On SAP's second-quarter call in July, Christian Klein said the AI transformation "doesn't require massive operational investments like the cloud transformation did". I have read that sentence several times. A large European company said out loud what a lot of budgets already assume: this change is the cheap one.

A few of the things I read last week:

• On the same call, Klein called the rise in R&D costs temporary, tied to first investments in AI tools and AI experts. (SAP Q2 2026 earnings call)
• A World Bank background paper put AI adoption among frontier firms in emerging markets at about 40%, tracking prior digital capability rather than size or money: 63% among digital-first firms, 16% among the rest. (World Bank / IFC)
• Vivian Lee, Linda Mantia and Jon McNeill write that what once took a large team a year can now be built in weeks by a handful of people, while incumbents are held back by "siloed data, legacy workflows, and rigid roles". (HBR)
• Golo Henseke at UCL found 12% of workers across 35 European countries using generative AI at work, from under 3% in Bosnia-Herzegovina to about 25% in Luxembourg, shaped by job content and by practices such as employee voice. (UCL, arXiv)
• Ethan Mollick, in his summer guide to the tools: "Working with these systems is more like managing than it is chatting." (One Useful Thing)

Read together, the cheap line stops looking like a fact and starts looking like a decision somebody made.

Cheap compared to what

I want to be fair to Klein, because in one narrow sense he is right. You do not need to rip out a data centre to use a model. Set against the cloud transition, the technical bill really is small, and anyone who has lived through one of those programmes will feel the relief in that sentence.

But the cloud bill was never mostly technical either, and we knew it at the time. What made those years expensive was the arguing about who owns which process, the retraining, the decisions that had to move. We are now being told that none of that applies. The tool is cheaper, so the change must be cheaper. I do not think that follows. If anything the second part gets bigger, because a model reaches into judgement rather than infrastructure, and judgement lives in people and in how we have agreed to work.

What bothers me is what "no massive operational investment" turns into further down the building: no budget line. And a change with no budget line has no owner, no plan and no date. It becomes something people absorb on top of the job they already have, which is usually another way of saying it will not happen.

What you cannot buy in a quarter

The World Bank paper is the most useful thing I read last week, and almost nobody will see it. It looks at frontier firms in emerging markets, and what decided who adopted AI was not size and not money. It was whether the firm already had digital capability: cloud, usable data, engineers, standards. Companies founded after 2015 were above 50%; those founded before 2000 were at 11%.

What predicts whether AI lands is a stock of organisational capability built up over years, and there is no quarter in which you can buy it. That stock is the investment Klein says is not needed. It is not needed at SAP because SAP already has it, which is a very different thing from it being free.

Henseke's European numbers say the same from the worker's side. A spread from under 3% to around 25% across 35 countries is not a story about who was issued a licence. What moved it was the content of the job and how the place is run.

The strongest case against me sits in the same HBR piece. Lee, Mantia and McNeill describe AI-native startups building in weeks what used to take a large team a year, and for them the change genuinely is cheap, because there is nothing to change. But the same authors say incumbents are slowed by siloed data, legacy workflows and rigid roles, and their first instruction is to redesign processes before automating them. The startup case does not tell us this is cheap. It tells us the cost is the organisation you already have, and that is the bill nobody is writing down.

Benedict Evans added a smaller worry last month: nobody knows what tokens will cost in two years, so even the cheap part sits on an unsettled rate card. And as we wrote on Friday, the market has stopped paying extra for AI skills.

If you work in L&D, HR, or transformation

This usually reaches us as an enablement request with no money attached, and I understand why "it will cost real time and take two years" is not a welcome slide. Two things anyway. Take the next AI project on your list and write one line of cost that is not tool cost: the days of people's time needed to redesign the work, review the output and change who decides what. Even a rough figure turns an assumption into a proposal someone can argue with. Then borrow from the decision-rights research Lindy Greer, Jennifer Jordan and Maxim Sytch published this month: across more than a hundred companies, those rights get written once by senior people and then left alone. Take one workflow an agent is about to touch and get the people who do that work into the room where the rights are set.

The provocation

Find your AI budget this week and look at how it splits. What share is buying tools and tokens, and what share is buying the time of the people who have to rebuild the work around them? If the second number is zero, we have not decided this change is cheap. We have decided not to pay for it. So who would you have to convince, and what would you put in front of them?