Most organisations now have all the AI access they need. Far fewer can point to what changed because of it. That gap is not a technology problem, and closing it starts with a decision made before any tool is bought.
Value is not an effect of implementing AI. It is a decision that needs to be made before implementation begins.
That was the thesis I argued recently at Otwarte Forum AI, hosted by SGH Warsaw School of Economics: fifteen minutes, one argument. It is worth setting out in full, because the pattern behind it is not specific to that room.
Organisations are no longer struggling with access to AI. They are struggling with translating that access into something visible in their results. And that is not a technology problem.
The data is not new
Twenty-five years of technology transformations tell a consistent story: results rarely materialise at the scale of the effort invested. In KPMG’s 2005 global survey of project management, 86% of organisations reported losing up to a quarter of their targeted benefits across their portfolios. Two decades later, MIT’s 2025 report The GenAI Divide: State of AI in Business found that 95% of generative AI pilots delivered no measurable business return.
The pattern that repeats is the same across both: starting with the technology before understanding the problem, and before accounting for the non-technological factors that determine whether anything actually changes.
AI is not an exception to this pattern. It simply makes the consequences arrive faster.
Four questions to ask before you commit
Before committing to any AI initiative, there are four diagnostic questions worth asking, one for each of the lenses we use to read a system of work:
- Value: do we know what business problem AI is supposed to solve?
- Flow: can decisions and knowledge actually move through the organisation?
- Quality: are we validating results before the next budget cycle drowns them out?
- Experience: do the people who are expected to use this actually want to?
Most teams can describe what they have produced. Far fewer can point to what has changed.
A case worth recognising
Consider a B2B SaaS company: around €48M in annual recurring revenue, net revenue retention sliding, and delivery stuck at fourteen weeks against a target of eight. The question on the table was which AI tool would speed up coding. The answer landed quickly: GitHub Copilot, hundreds of engineers, a significant spend for a year, decision made in three weeks.
The tool did exactly what it was built to do. Delivery time: unchanged.
The constraint was cross-functional dependencies, not coding speed. AI accelerated what already existed. It just was not the problem worth solving. That scenario is not unique, and the people who recognise it usually recognise it from their own organisation.
Outputs and outcomes are not the same thing
The distinction between outputs and outcomes is operational, not semantic. Outputs are under our control: the feature shipped, the tool rolled out, the model deployed. Outcomes depend on whether something shifted on the other side, for the client, the business, or the people doing the work.
This is why value has to be decided first. If you cannot say what would change, and for whom, you have no way to tell whether the AI initiative worked, only whether it was delivered.
The question worth leaving any leadership team with is not how to implement AI. It is how to make it matter.
For the four lenses behind these questions, and how to use them to read your own system of work, see The Four Lenses.