There’s a conversation happening in European boardrooms this year that goes roughly like this: “We invested in artificial intelligence and ran the pilots. The pilots worked, but very little reached production. What went wrong?” 

The answers on offer are almost all technical:  

  • “The models weren’t good enough.”  
  • “The data wasn’t ready.” 
  • “The talent wasn’t available.”  
  • “The infrastructure wasn’t in place.” 

Some of that is true, but none of it is the root cause.  

The actual issue is that most organizations delegated a leadership question to an engineering department. Then they’re surprised when engineering returned an engineering answer. 

What’s actually being asked 

When a board approves an AI program, they believe they’re approving a technology investment. They aren’t. In reality, they’re approving a change in who—or what—makes certain decisions inside the company and under what accountability. This is a serious governance decision, and it belongs to the people who own governance, not technology. 

Consider what an agent actually is once it’s in production. At the essence, it’s an entity that makes consequential judgements at volume, without a named human doing so each time.  

It follows that, at first, someone must decide which judgements an agent may make. Someone must decide what happens when it’s wrong, because at times it will be. Someone must be accountable to a customer or an authority for an outcome that no individual chose. 

Those questions have obvious owners in every other part of a business. For instance, when a company gives a purchasing authority to a regional manager, nobody thinks it’s an IT matter. But when the same authority is given to a software system, it somehow becomes IT’s responsibility. 

The pilot trap 

This explains the pattern that has frustrated so many European executives this year, and it’s worth stating plainly because the diagnosis determines the cure. 

Pilots succeed because they’re technical exercises. They have a defined scope and a measurable output. They’re run by a small team and, critically, always have a human between the system and any real consequence. 

Moving to the production phase presents a very different challenge. Production removes that crucial human in the middle, which means that someone has to establish what happens when something goes wrong.  

That question usually has no owner. It escalates  and circulates endlessly, until eventually the project is described as “promising” and reduced in scope until the human is back in the loop. At that point the economics—the key selling point—disappear, and the program gets quietly deprioritized. 

Clockwise loop: a pilot with a human moves to production, the human is removed, no owner and endless escalation follow, scope is reduced, and the human returns, bringing the process back to the pilot stage.

What leadership actually has to do 

Three things, none of which can be delegated downward:

Decide the accountability model before the architecture. 

For each process where autonomous decisions are intended, name the accountable executive. This isn’t the project sponsor but the person who answers for the outcome.  

If that name is difficult to produce, it’s a tell-tale sign that the process isn’t ready. Don’t panic if that happens: finding that out now is considerably cheaper than finding it out after deployment.

Set the risk factor explicitly and in numbers. 

Every autonomous system has an error rate; you can’t get away from it. The question you should be asking isn’t whether it makes mistakes but whether its error profile is better than the human baseline it replaces.  

Additionally, can your organisation tolerate the shape of the errors an agent makes? What I mean is that autonomous errors will differ from human errors in kind, not just in frequency. For instance, a system won’t get distracted and accidentally enter data in the wrong cell, but it might completely misunderstand the purpose of the entire column.  

As you’ve probably already noticed, these are board, executive-level judgements. Engineering can’t make them and shouldn’t be asked to.

Fund the unglamorous layer.

There are several key factors that determine whether any of this succeeds:  

  • process integrity,  
  • data governance,  
  • system interfaces,  
  • decision logging.  

Let’s be honest: these aren’t too spectacular. They appear in no keynote. However, this is still where the money needs to go, and leadership has to protect that allocation against the persistent internal preference for visible novelty. 

Why this is Europe’s opportunity 

There’s a fatalistic view that Europe has already lost this because the models were built elsewhere. I think that gets the competition wrong. 

The models are available to everyone. What isn’t so widely available is the ability to deploy them into complicated businesses without breaking them. That requires exactly the governance seriousness that European enterprises are frequently mocked for, and that will turn out to be the scarce capability. Organisations that have spent decades learning to govern consequential systems have an advantage here, provided they recognise the problem as theirs.  

The companies that will be ahead in three years aren’t the ones that simply adopted the technology first. They are the ones whose leadership understood soon enough what was actually being decided.