AI gets stuck in the middle

Over the past months I have read one sceptical analysis of AI after another: companies are spending heavily, but the productivity gains are hard to find. I share that assessment entirely. I struggle to see real success stories in most companies.

The reason is almost always the same: these companies have not managed to adapt and adopt the technology across the organisation in a way that creates value. That has been a well-known problem of change and transformation management for decades, and it has very little to do with the technology itself.

Ethan Mollick made a related argument in the Economist in April, in an opinion piece I really liked. Companies, he wrote, treat artificial intelligence “as if it were just another piece of enterprise software”: they slot it into existing processes, give it KPIs and hand it to the IT department. Since the mandate of most IT departments is to minimise risk, while AI only pays off if people experiment and are allowed to fail, he calls this a “category error”. In his view, the remedy has three parts: direction from the very top, employees who are given real permission to experiment, and a dedicated team that collects what they find and feeds it back into the organisation.

I mostly agree with this, and I will come back to the part where I don’t.

But first, let’s see what the problem looks like from the point of view of a manager.

Easy to use, hard to deploy

Part of the problem is a misunderstanding that the technology itself invites, as AI is extraordinarily accessible, maybe more so than any technology at such an early stage: by now almost everyone uses a chatbot at work and at home and finds it easy to use. So it seems natural to assume that using it inside a business must be just as simple, but it isn’t. Behind the friendly interface sits a technology so complex that nobody fully understands how it works. Not even the people who build it.

The difficulty grows the closer you get to the core of a business model: that is where AI could create the most value, and also where it is hardest to introduce, because the core is, naturally, where processes have been refined over years, where roles and routines are settled and where the people inside the organisation are most competent. Changing activities, job descriptions and eventually culture is hard anywhere, but in the core it is hardest or even impossible, because the core is what the company is. In addition, I would strongly argue that technology should not replace one-to-one what people do. What people do needs to be understood in its context. Only then can one see whether and how it can be replaced or, more often, complemented by technology. Technology usually creates the most value where people and machines complement each other, and working out where that is takes more than a rollout plan with clear names and dates behind each line.

What happens in the middle

Picture a production manager who has been very good at his job for many years. He hears that AI is now supposed to reach into his processes.

What goes on in him?

In my experience, most people react to AI in a similar way: with a mix of curiosity and scepticism. Everyone has understood by now that the technology is powerful, and most managers even have a rough idea where it could make their job, their department or their company better. What it means for them personally is far less clear. Part of the uncertainty comes simply from not knowing the future. In the case of AI, nobody is an expert in this technology, so nobody quite knows what it can do, or what it will be able to do, or how far to trust it, and when. That part can be addressed, and there is no shortage of consultants offering to do exactly that.

The other part is harder to articulate for most people. Somewhere in the back of the mind sits the thought that if AI delivers enough, the company will not need all of his people any more. And if it doesn’t need all of his people, it may one day not need the manager on top of them either.

I don’t think this fear is very acute for most people today, except of course in industries that AI is already hitting hard, such as marketing or, once again, media. In a manufacturing company, though, nobody expects to be replaced next month. But the thought is there, and it shapes behaviour.

What I see most often is waiting.

It’s a kind of “freeze reaction”, and an uncomfortable one, because nobody knows what is coming. The uncertainty is then projected upwards: surely top management has a plan, they just aren’t telling us. That may be true. It may also be completely wrong. Mollick in his article describes the mirror image of this: employees who use AI quietly and keep it to themselves, because they fear punishment or don’t trust that the gains will be shared.

Either way, the people at the top lose sight of what is actually happening in their own organisation.

Three real examples, three different traps

Three cases I know well show how differently this can play out. None of them is a story of failure in the usual sense, which is exactly what makes them instructive.

The first is a company we work with, a very successful mid-sized manufacturer. Leadership put AI on the agenda early, and it formally appears in the strategy. The company tried various things: internal productivity tools and even an AI tool for its customers. The customer tool had to be switched off again, though, because it hallucinated and produced seriously problematic results. I would put that down to teething troubles of the technology; it would probably already work better today than a year ago when they first tried. The more important point is a different one: the tools were not placed where AI could create the most value in this particular business model. The company is doing many things right. It simply hasn’t found its value levers yet, and I suspect the same is true for most companies even if they have taken the topic seriously at a strategic level.

The second example is an old friend of mine, a middle manager leading a team of software engineers in a mid-sized software company. He saw his own role as acutely threatened, since it consists almost entirely of processing, analysing and formalising information. His response was to adopt AI as aggressively as he could: he started very early, at a time when running agents still meant setting up dedicated machines. At one point he worked with around 70 AI agents, which he first treated like additional employees. In fact, they made his team ten times larger. A few months later, he switched to a capabilities-based approach: instead of managing AIs as employees, he used them to complement human capabilities, with even greater productivity gains. And with exactly that success came a new problem, and a very hard one: his agents now write millions of lines of code, a quantity that nobody can fully review. All he and his team can still do is sample checks. For a software company, releasing code that no human has ever properly reviewed is a real risk. The twist: he had feared for his job, but the opposite happened. Top management saw what he had achieved, called it best practice and promoted him, so he could do the same in other departments, too.

The third example is a country manager of a large international group. Headquarters has, for now, largely blocked AI for decentralised units, such as the country organisation the manager runs, on legal and compliance grounds, as they are concerned about data security in particular in the use of customer data. The manager nevertheless (of course) uses AI for his own office work where no customer data is involved. But his and his organisation’s job is sales, and sales is definitely an area where AI could help tremendously. Yet because customer data is part of almost everything in sales, it is also where the legal and governance risks sit from the point of view of headquarters. What would be needed is a central initiative that says: sales needs AI support too, and here is the governance for how it works, who may experiment and in which format. I understand the manager’s frustration at seeing every day how much easier things could be, if only he were allowed to use the tools.

Where to experiment, and where not

Is the software team lead a role model or a warning? Both, I would argue. In his position I would almost certainly have done the same. And yet the unreviewed code is a risk his company now takes, and the people responsible for it in the end are top management, i.e. the CEO and ultimately the board (who, in many organisations I know, are quite detached from these day-to-day dilemmas).

The dilemma is very real for every company, though: leadership needs to be acutely aware of what both a too slow and a too fast adoption can cost.

This, by the way, is where I part ways with Mollick, or at least with the way his argument is often applied, which is that a change in culture is required. I have seen companies declare that from now on everyone should experiment with AI, and present that as an organisation-wide cultural shift.

I simply don’t think this works, and I actually think it’s the wrong approach.

It reminds me of the years when companies across many industries naively tried to roll out innovation culture and agile working methods across whole organisations. That was a mistake then, it is a mistake now.

In many departments and functions the job is to keep what exists running robustly and perhaps to improve it at the margin. Iterating, failing and learning has no place in a process that is business-critical, let alone one that is relevant to safety. The same holds for AI: it should not be rolled out everywhere, and not every employee should be told to experiment with it. It belongs in the roles and activities where it makes sense. Of course, everyone should think along and be allowed to suggest new applications, but not to execute all of them immediately.

To me, two things follow, and they are very different:

First, top management has to work out where the strategic value levers lie, and that is a top-down analysis and decision.

Second, governance needs to set up fields of experimentation, where initiative from employees themselves is wanted, encouraged, made visible and ultimately leveraged.

Both are necessary, but they differ in kind and goals, apply to different topics and people, and are executed in different ways.

Part of the job of leadership today is to decide where experimentation is not wanted, and to say so clearly. The software company rewarded experimentation without deciding where it belongs. The international group blocked it without saying where the way forward lies.

What remains for the individual

For the manager in the middle, there is unfortunately no single right strategy.

In some companies the aggressive early adopter earns legitimacy and, in the end, keeps the job. In a conservative, risk-averse culture, a cautious approach may be the wiser one. It depends on the company, on the specific industry and context, and it depends on the person.

What helps is what has always helped people in times of uncertainty: a structured way of thinking it through. What does this actually mean for me and my team? Which options do I have? What should I do now, and what should I not do yet? Am I in the right place, or do I need to change something? Or have I perhaps been worrying more than necessary?

None of this has to do with technology, and no vendor or rollout plan will help with that.

But judging by the conversations I have had over the past year or so, almost every manager is asking themselves these questions right now. Most of them on their own.

That, I think, is the first thing that needs to change now.

The questions will only get harder from here. Rather than providing ever more answers that the next model update from an AI firm will make obsolete, managers need to learn how to cope with and work through the tough questions themselves.

The sooner, the better.