High-performing enterprises in the AI era

Investing. My wife and I run an investment firm, and we currently develop it within the thriving Swedish tech community, covering Germany and Europe at large. We organize investor gatherings, deliberately small, where everyone states their thesis, their stake and their ticket size, and then we go deeper than a conference room usually allows. As a smaller investor I cannot run the full diligence a large fund runs, so we lean in alongside top VCs. Product-market fit matters and team capability matters. The strongest signal is the founder, and the ones who have been through genuinely hard times, including privately, tend to carry a company further. People who have seen the dark side get up again.

Background. I was an experimental physicist rather than a theoretical one, which means lab work and every problem the theory does not mention. We shrank structures, pushed electrons through them at low temperature in a magnetic field, and measured the quantum transport effects, work that sits underneath quantum computing now, where the noise problem may or may not ever be solved. I also coupled neurons onto a silicon dioxide field-effect transistor, trying to measure an action potential and then to initiate one. And I did neuromorphic engineering at Caltech, using the transistor in its exponential region instead of as a digital switch.

Leadership. What I admire is a large vision paired with a hunger for detail. Mukesh Ambani came from oil, gas and retail, and went far enough into 4G, network build, vendors and handsets that any engineer would be in awe, because he would not rest until he understood. The vision was a digital infrastructure for 1.3 billion people, in a market where eleven carriers already competed. That combination is rare, and people confuse the detail half with micromanagement. Understanding where the money flows is not interference. A leader should be closest when the problem is deepest, then let the team run.

High performance. Two things together. Value, for shareholders and underneath them for customers, employees and the societies we operate in. And resilience, because performing today guarantees nothing: inflation goes through the roof, a technology arrives, a competitor appears, a regulator moves. High performance means learning and adapting faster than the competition, which no static design delivers.

Exponentials. A model that optimises itself is a hyper-exponential behaviour, and it raises the question whether one player's small advantage compounds into an unassailable one. We have several frontier labs rather than one, which helps. Publicly available training data is now largely exhausted, so the frontier moves to private data, orders of magnitude larger and sitting in company databases, and to synthetic data. Reinforcement from human feedback will thin out because there is simply too much to supervise, and reasoning engines that question and check themselves take over: system to system rather than system to human.

Organisations as organisms. An organism has parts that must work together, signals travelling through it, and something like an immune response. Enterprises are living in the same way, and we treat them as static. We draw an organisation chart, then reorganise, and before every reorganisation everyone already knows the next one is coming. Some companies reorganise so often it hurts them. What matters is capabilities, and matching the capabilities a problem requires to the ones we have. At Reliance Jio we hit this early and organised around processes instead of lines: acquire a customer, onboard, provision, bill, support, and then recruit into each process whatever capability it needs, engineer, UX designer, front end, back end, operations, finance, legal. The line organisation hosts the capabilities; the process is the dominant axis. Pushed forward, with a co-pilot at everyone's fingertips, the question becomes how large an organisation needs to be at all.

The data and AI foundation. In an organism everything is connected, and in enterprises we build silos, data silos above all. Unlocking that data is the first job. The second is digitising processes end to end, because a process made of people typing into one system, reading an email and typing into another cannot be automated at all. Then you can automate, and then smartify. ServiceNow became our connective tissue: one ticketing system across the IT estate and the customer domain, so a customer complaint and a server fault can be correlated, and the customer told what broke and roughly when it will be fixed, with the loop closed when it is. On top of a digitised process, a bot watches a person restart a downed antenna, and next time the person is not needed. That is a learning system, and it moves people toward the uncharted work. Recommendation engines decide whom to sell what, when, and through which channel, and a language model now also shapes how the offer is presented to you specifically. Foundation means more than technology: visualisation at scale so nobody waits for an IT person or an Excel sheet, democratised access with ownership and confidentiality intact, and sandboxes where people can experiment before anything scales.

Governance and the economics. Leaders should ring-fence a domain where experimentation is safe, then scale what works. Chatbots are the cautionary tale: we launched them, customers got so fed up that call centre volumes went up, and the whole category acquired a bad name. With conversational models now handling audio and video, the moment to try again has arrived, contained, with a sub-segment or friendly users. Regulation is not the blocker; Europe has too many rules rather than too few, and that should never be an excuse for standing still. On architecture, a vendor who wants all the data fed in so the magic can happen in a black box confines your intelligence to theirs, and I have rejected that. Data has to stay accessible and fluid, with standard APIs introduced whenever a system is touched anyway. Value appears only where something is used: serving a customer, serving an employee, automating a workflow. That makes the foundation an economic decision, the same one we make on a network. The full cost of a network buys empty capacity and is punishing until it fills. The decision to extend is judged on incremental cost. The first use case often carries the shared enabler, as the first cars carry the factory.

How work changes. People converge on three things: defining the future, setting the guardrails and alignment, and handling exceptions that grow rarer. The manual work in disconnected systems disappears. Companies too slow will die, the new ones will build connected from the start, and the transforming ones have to move fast enough not to be wiped away. Innovation still happens at the intersection of domains, so human creativity has room for years. And there has never been a better time to build something new: the foundational model train has left, and the platform it created unlocks the next S-curve, as cloud and mobile did before it.