Elisabeth L’Orange

Ms L’Orange, you’ve put forward the argument: Stop co-piloting, start driving or, put another way, we have too many pilots. What exactly do you mean by that?

Elisabeth L’Orange Companies launch huge numbers of pilot projects, but very few ever make it past the pilot stage. The problem is that these initiatives are often developed in sandbox environments that simply aren’t designed to scale across an organisation. Pilot projects often lack access to the relevant databases and interfaces needed for broader deployment. As a result, there’s a great deal of activity, but very little progress when it comes to scaling the use cases that actually matter. Yet that’s exactly where organisations need to focus their attention if they want to make real progress.

Was there a particular moment when you realised this, or was it something that became apparent over time?

L’Orange There were a number of occasions when I was building pilot projects for companies and clients and realised they could never be scaled because they didn’t even have access to the necessary databases. I found that increasingly frustrating, given the amount of time and energy that went into building those pilots while knowing they were never going to deliver meaningful results. At some point, it became clear to me that organisations need greater visibility over their initiatives and should focus their investment on projects that create genuine business value. Many companies provide AI licences to large numbers of employees. Those employees may end up writing emails more quickly, but the impact on actual business outcomes is often negligible.

There are plenty of fascinating examples where AI is being used in ways that make you wonder whether the effort is justified. It is a case of using a sledgehammer to crack a nut.

L’Orange Absolutely, and I’m not excluding myself from that. I use AI to remove backgrounds from images in my presentations and for translations. Those are useful applications, but they’re not the AI use cases that really matter. The most important applications lie in the backend of the organisation, in its processes. Getting there requires the right foundations. Processes first need to be set up in a way that allows them to be automated, and data need to be structured so that AI can access them. I believe a great deal of value can already be created once processes are clearly defined. Many organisations have hired business process specialists or automation managers, roles that are increasingly being renamed Business AI Automation Managers, simply to automate processes first. Generative AI can then be added as the next step, as the icing on the cake. Without the cake, however, the cherry serves no purpose, and in many organisations, those foundations are still missing.

Do companies need to rethink entire departments or processes from the ground up?

Elisabeth L’Orange Absolutely. I believe AI-first companies are the future. Many of them will be built from the ground up and, in some cases, fundamentally reshape established industries. The real challenge lies with established organisations. Transformation is difficult at the best of times because it requires people to change the way they work. Introducing new processes means convincing everyone to do things differently. Questions of this kind go far beyond organisational structures alone. Even during the era of digital transformation, organisations achieved only limited success in many areas. Turning established businesses into genuinely AI-first organisations won’t be easy. AI can certainly be used to automate workflows and improve efficiency, but the biggest opportunities, in my view, lie elsewhere.

A 2025 report from MIT came to the surprising conclusion that 95% of companies are not generating measurable returns from generative AI. Does that match what you are seeing?

L’Orange Yes. One of the biggest problems is that companies try to tackle too many things at once instead of focusing on a small number of use cases and scaling them consistently. Interestingly, private-equity-backed companies are often far more successful at transformation because leadership teams have very clear objectives and are measured against them. Large organisations, by contrast, tend to be much more cautious. Significant investment goes into security and governance, even though many AI initiatives have yet to reach production, which can divert attention from implementation itself. For that reason, I see the main challenge not in the technology but in strategy and management. When incentives are misaligned and different stakeholders are pulling in different directions, meaningful progress becomes difficult. 

The report points to a similar pattern. A large share of generative AI budgets is being spent on sales and marketing because those areas produce visible results quickly. The biggest savings, however, are typically realised in the back office, whether through outsourced services, document processing or risk management. Are there any positive examples?

L’Orange One example is the way DKB has automated parts of its application processes. The further a process is removed from a critical lending decision, the greater the role AI plays. As decisions become more sensitive and business-critical, a human in the loop takes over. Even so, the entire application journey is supported by AI, while multiple systems monitor one another in parallel to ensure that none of them hallucinate. All of this happens in the background and remains invisible to both customers and employees. Banking and financial services offer a wide range of AI use cases, and the same is true of law firms, where AI is often used without customers or clients ever noticing it. 

Why, then, do companies continue to invest so heavily in visible applications, such as chatbots on their websites?

Elisabeth L’Orange Because there is still a great deal of uncertainty. I’m not a fan of website chatbots either, as the evidence suggests that hardly anyone actually uses them. Companies need a more data-driven approach and a more data-driven strategy, with decisions grounded in reliable metrics rather than gut feeling, especially when they have long-term implications. Many people have also grown tired of obviously AI-generated content. In the case of text, for example, the style, wording and semantics are often immediately recognisable. Nobody wants to read that any more and I would never use AI to write for me. That is why large US companies such as McDonald’s have announced that they will stop posting AI-generated images. Cosnova, the company behind the cosmetics brands Essence and Catrice, provides another interesting example. Despite generating around a billion dollars in revenue from products such as nail polish and cosmetics, the company has deliberately decided against using AI-generated content in marketing. The brand wants to show real skin and real people because authenticity is part of its USP. Behind the scenes, however, AI is everywhere. Product decisions are continuously checked against the company’s values and assessed for their relevance to individual markets.

Cosnova has also warned about the risk of “brain fry” caused by excessive AI use. According to a BCG study, the brain can shift into a kind of energy-saving mode, reducing people’s ability to exercise their own judgement. That becomes a problem when processes evolve to a point where nobody fully understands them any more. How can organisations guard against that?

Elisabeth L’Orange The study showed that once people start chaining together multiple AI tools, they gradually switch off mentally. By the time people are using their second or third tool, they often stop thinking as critically. Marketing is particularly exposed to this, which means the issue quickly extends to HR and organisational work as well. A typical example would be someone planning a campaign and moving from one AI model to the next: using ChatGPT or Claude for text and images, Gemini to animate the visuals, Suno to create the music and Adobe to bring everything together. The workflow may be highly automated, but it can also be mentally exhausting, making it harder to stay focused and, over time, reducing people’s willingness to do certain things themselves or think them through independently. That’s the real problem. The brain is an inherently lazy organ. It consumes a great deal of the body’s energy (around 20 watts) and if it doesn’t have to exert itself, it generally won’t.

So the key is to use AI intelligently and focus primarily on backend applications?

L’Orange The greatest value often comes from applications that remain invisible to users, or from using AI as a sounding board. When I’m preparing a presentation, for example, I might ask AI whether there are additional studies on a topic, more background information or critical questions I should consider. It allows me to explore a subject in greater depth and challenge my own thinking. The same applies when I come across a news story and want more context. An AI model can be like the wise older sibling many of us wish we’d had, someone who always seems to have an answer.

A good example is Ikea’s chatbot Billie. Rather than laying off 8,500 customer-service employees, Ikea retrained them as interior design advisors. What began as a customer-service function evolved into a sales channel generating billions in revenue—the exact opposite of an invisible AI use case. So where does AI create the most value: behind the scenes, in the spotlight, or does it depend on the use case?

L’Orange It depends. Only a relatively small number of industries and use cases will be fundamentally transformed by AI, and customer service is definitely one of them. Many of the systems people interact with today no longer reflect the current state of the technology because companies often selected and implemented them three or four years ago. In large international organisations, procuring and rolling out a technology globally can take years, which means that many of the chatbots customers encounter today—and often find frustrating—are based on technology that is already outdated. Far more capable models are available today, which is why I believe customer service is heading for genuine disruption. Parloa, a start-up based in Hamburg, is a good example. Its voice bots have been completely rethought, rebuilt and retrained over the past few years and are now, in some cases, difficult to distinguish from a human interaction. Once systems like these are deployed at companies such as Ikea, we will start to see genuine disruption. Existing solutions are still limited by latency and performance issues, so many people haven’t yet seen what these technologies are truly capable of. In my view, that will change very quickly. 

Will AI really take over the routine work and leave people to focus on the more complex tasks?

L’Orange No. Almost every technological revolution has come with that promise, yet it has never really materialised. When household appliances such as washing machines and dishwashers were introduced, they were seen as revolutionary. Back in the 1960s and 1970s, many people believed they would lead to a 30-hour working week because so much time would be saved. The reality is that many people work 60 hours a week today. Every gain in efficiency has tended to create new expectations and new demands. The assumption that AI will primarily take over routine tasks is now being questioned as well. A Stanford study analysing 250,000 Anthropic conversations found that most interactions centred on agency, meaning the requests often involved complex tasks with decision-making implications Questions ranged from whether an employee should be dismissed based on their behaviour to investment decisions based on specific data sets. 

So AI is no longer limited to routine tasks. It is increasingly influencing business and personnel decisions as well.

L’Orange Yes, and AI is certainly not unbiased, given that some models have been trained on questionable data. Even so, the bot often becomes the third voice in the room. One of the most common AI use cases in Germany, Europe and the United States is relationship advice. 

At the same time, companies are pouring vast sums into AI, sometimes with surprisingly little cost control. Amazon, for example, reportedly exhausted its entire AI budget within a month, while GitHub’s move from subsidised pricing to commercial rates led to a sharp increase in costs for customers. Are companies beginning to understand what AI really costs?

L’Orange Not yet. In my view, that realisation will only come once companies are under pressure to generate profits. Many are still backed by venture capital, but public markets operate by different rules. Following an IPO, companies such as Anthropic eventually have to prove that they can make money, and given the scale of the current deficits, that won’t be easy. Some are still billions of dollars short of the revenues they would need, which is why prices are likely to rise. Token costs will increase, and organisations will start thinking much more carefully about where AI truly creates value. Many of the things we now take for granted, such as removing image backgrounds, may no longer be economically viable in the long run. Google and Microsoft are in a different position because they can absorb those costs across their broader product portfolios and integrate AI deeply into their platforms. Even so, both companies have started funding recent investment rounds with debt rather than financing everything from operating cash flow, marking a notable shift in direction. Ultimately, though, everyone is making the same bet. 

If AI is likely to become much more expensive, what does that mean for businesses? Should they be investing in AI or not?

L’Orange The time for experimentation is over. Companies need to start quantifying real outcomes. Every AI automation project should be backed by a business case or, at the very least, clear KPIs. For me, the most important KPI is time to decision. We need meaningful metrics that answer a simple question: what value is AI actually creating? In many situations, having a person write their own emails may still make more sense than having a model do it for them. Organisations need to be disciplined when selecting use cases and focus on areas where AI can genuinely be scaled and create measurable value. Not every process can be transformed, and when employees strongly resist change, companies need to ask themselves whether forcing that transformation is really worth it. Some of the biggest opportunities may lie elsewhere. Pharmaceutical companies, for example, are using AI to uncover effects and applications for medicines that had not previously been recognised, effectively creating new products from existing ones. That’s where I would focus AI: where it helps companies rethink products and, ultimately, the business itself. Making that happen is enormously complex, however, and that complexity is often underestimated.

Not everything can be measured. Some processes may even look highly optimisable on paper, yet end up producing the opposite of the intended result.

L’Orange Exactly. The whole process revolution is bringing many of those things to light. The sudden transparency is revealing that many processes only worked because of human interaction, because of what are often called “glue workers”—employees whose performance may not look impressive on paper but who keep the organisation running. The real problem is that these are often the very people who end up losing their jobs during AI-driven transformation, even though they are central to how the organisation operates. In reality, those are the people who should be given AI tools, while the processes around them are rethought before AI is layered on top.

Different departments often end up doing their own thing. Sales buys one tool, IT builds something in-house, and someone in marketing quietly uses ChatGPT through a personal account. How can companies prevent that?

L’Orange Shadow AI is widespread, although the same could be said of shadow IT. Companies have always had people installing their own tools or finding their own ways of getting access. What stands out today is the gap between private and corporate use of AI. Adoption rates in the consumer space are almost twice as high, and while usage in the US continues to grow, adoption in German companies has barely changed between 2025 and 2026. A major reason is that many employees still don’t have access to AI at all. Over 60% have no model they can use for experimentation or day-to-day work, making widespread adoption difficult from the outset. Reaching agreement inside organisations is often part of the challenge. Risk, compliance and employee representatives don’t always see eye to eye on how AI should be introduced, and until those issues are resolved, adoption will inevitably remain limited. Some highly publicised failures in the early days also contributed to the hesitation. Companies uploaded their own intellectual property into models, for example to review patents, only to discover that the information had been incorporated into training data. Cases like that are obviously problematic. Many of these risks are manageable. If employees know what can and cannot be entered into a model, most of the concerns can be addressed. Enterprise versions are now available for virtually all leading models, giving organisations much greater control over how their data are handled. Central gateways add another layer of oversight, making it easier to control what leaves the organisation and what stays inside. I always say: if we can put people on the moon, we should be able to figure out which data a model is allowed to send outside the organisation and which data isn’t.

Which part of the organisation should ultimately own this? IT, a Chief AI Officer, or the business units themselves?

L’Orange  AI needs to be woven into the fabric of the entire organisation. Everyone has a role to play; it isn’t simply an IT issue. Responsibility needs to sit with the CEO and the leadership team because they’re the ones who decide where the company is heading and which use cases should be prioritised. At the same time, individual business units need enough freedom to develop and drive their own applications. People often debate whether a centralised or decentralised approach works better, but a lot depends on the size of the organisation. A company such as Deutsche Bahn is so large that many things simply can’t be managed centrally in any meaningful way. Smaller organisations have much more room to do that. Ultimately, though, companies tend to move faster when business units are given the freedom to drive their own AI initiatives within a clear framework. AI isn’t something that can be introduced by a single person or a single team. It has to become part of how the organisation operates.

Looking ahead, how do you think we will use AI in the future?

L’Orange I think we’re still using AI in a very fragmented way. Much of the focus is on writing emails and automating similar tasks, but I suspect that approach will become obsolete over time. As organisations evolve, what we’re more likely to see is a genuine partnership between human and artificial intelligence, which some describe as carbon intelligence and silicon intelligence. The real potential lies not in replacing one with the other, but in creating meaningful synergy between the two. The transhumanism debate in the United States has produced some very far-reaching visions of the future. I don’t believe we’ll all end up with chips implanted in our brains and simply upload knowledge. I do think we’ll be able to work much faster and more effectively with AI. The real opportunity, however, doesn’t lie in automating small, routine tasks. Personally, I think it’s a waste of time to focus on things like writing emails when there is so much greater potential elsewhere.

If you were starting another company today, what would it be?

L’Orange I’d be very interested in building some kind of AI operating system for businesses. Much of what I’ve seen and used so far still falls short, largely because it has been layered on top of existing processes rather than designed around AI from the outset. Creating something fundamentally new would require those processes to be rethought from the ground up. In some ways, it reminds me of the Amadeus flight reservation system. Platforms like that are so deeply embedded in global workflows that replacing them is enormously difficult. Even so, I’m convinced that, in the long term, businesses will need a new generation of systems built specifically for the age of AI.