It is budget season. Here are the three questions I would ask before signing an AI line for 2027.
One. Whose gain is it?
McKinsey surveyed 1,719 organisations between May and June 2026. The result is worth reading twice.
80% of respondents say artificial intelligence improved their individual productivity. 37% attribute an effect to their organisation's operating result. And around 6% put that effect above 5% of their EBIT.
The gap is not a contradiction. It says the gain stays in individual hands and does not travel up to the collective. Everyone saves twenty minutes, and the company gains nothing, because those twenty minutes are not reinvested anywhere in particular.
The question to ask, line by line: does what we are funding here produce a measurable result at company level, or an improvement in individual comfort?
Both are legitimate. They are not defended before a board with the same arguments.
Two. Where will the overrun come from?
A survey of 1,636 technology decision makers in September 2026 found that 46.9% of organisations overrun their AI budget, 11.3% of them substantially.
What they do next is the interesting part. Among those reallocating inside the IT budget, 60.9% cut external providers and consultants first. And 23% go looking for the money in business budgets outside IT.
One trap explains part of these overruns. The unit price of models is collapsing, by a factor of 9 to 900 a year depending on the task. But consumption per task is exploding with reasoning and agents. The price falls, the bill rises.
The question to ask: if we overrun by 30%, where does the money come from, and who approved that in advance?
Three. What does regulation actually require in 2027?
Less than you hear.
Regulation 2026/1744, in force since 27 July 2026, postponed the European obligations applying to high-risk systems to 2 December 2027 for those in Annex III, and to 2 August 2028 for those built into regulated products. In Switzerland, the consultation on a future framework is only expected at the end of 2026, making entry into force before 2028 unlikely.
What does already apply are the transparency obligations, since 2 August 2026. Telling someone they are talking to a machine, marking generated content.
The question to ask: how much of our 2027 compliance budget funds a deadline that moved, and how much funds an obligation in force today?
The test that covers all three
A survey of 2,360 executives in early 2026 asked the following question: what will you do if your AI investments do not deliver the expected financial impact within twelve months?
6% say they will cut back. 70% will stay the course, 24% will increase resources.
A budget nobody would consider cutting, whatever the outcome, is not a budget. It is a conviction.
One last question, then, to put on the table. At what result, quantified and dated, would you agree to drop this line?
Sources
- McKinsey, The State of AI in 2026: On the road to ROI, August 2026, 1,719 respondents, fieldwork May to June 2026
- The Futurum Group, global survey of 1,636 technology decision makers, release of 10 September 2026
- Epoch AI, analysis of language model inference price trends
- Regulation (EU) 2026/1744 of 24 July 2026, postponement of obligations applying to high-risk systems
- Federal Council, decision of 12 February 2025 on Switzerland's approach to regulating artificial intelligence
- Boston Consulting Group, AI Radar 2026, January 2026, 2,360 senior executives
Related topics