Give it away to get ahead

On stage at DocCheck's excellent Breaking Med: High on AI event in Berlin in June, I made an argument that split the room. Whatever you're building, I said, the one variable that decides whether your organisation survives is the speed with which it learns, so gear everything towards maximising experimentation. Then came the part that made a few people lean forward and a few others fold their arms: healthcare providers and pharma companies should open up eligible internal tools, their half-finished apps and their workflows, and hand them to the people on the front line, practitioners, patients and carers. Not as a generosity play, but because it's the fastest way to learn.

The pushback was fair and serious: data sensitivity, liability, quality control. The agreement from the other half of the room was (about) just as loud. That split is the whole debate in miniature, and two signals just landed on the side I was arguing.

When the model under you gets swapped out every few months, the only thing that compounds are the loops your people and your tools keep teaching each other, and the firms that open those loops the widest will learn the fastest.

The model is not the moat

The first signal came from Satya Nadella, whose June essay drew around 28 million views. His framing: a firm now runs on two kinds of capital. Human capital is the judgement, relationships and pattern recognition of its people. Token capital is the AI capability it actually builds and owns. The mistake most companies make is treating "which model did we pick" as the strategic question, when everyone has the same frontier models by Friday afternoon. The durable advantage sits in the learning loops you build on top: the workflows, evaluations and accumulated judgement that improve every time the system is used.

He set a test I keep coming back to. You should be able to swap out a generalist model without losing the "company veteran" expertise baked into your system. If pulling the model erases what made you valuable, you didn't own anything; you were renting someone else's brain. Nadella reaches for an uncomfortable analogy: the way the first wave of globalisation hollowed out industrial economies through outsourcing, where the headline numbers looked fine while the tacit knowledge quietly drained away. He calls the loop a "hill-climbing machine", an asset that compounds because each improved workflow produces a better training signal than the last.

Why giving it away makes you faster

The second signal showed what that looks like when a cautious institution actually does it. In late June, Banco Santander open-sourced its AI-governance toolkit, starting with eleven repositories under Apache-2.0: a harness that stress-tests model guardrails against jailbreaks, a "mechanical governance" framework with hard gates for high-stakes decisions, a synthetic fraud-graph generator, a vendor-neutral client that switches between OpenAI, Bedrock and Gemini, even its own version of the run-an-agent-in-a-loop pattern. A bank handed its control layer to its competitors. The customer data stays private; the machinery is now public.

Why would it do that? The honest answer is not charity, it's speed. Banks have spent years trying to govern AI behind closed doors and shipped almost nothing; doing it in the open is the faster route to getting it right. Open code attracts the talent that improves it, signals internally that the tools are real, and gives regulators something to audit. A firm that already runs its governance in public doesn't just look more defensible, it deploys each new capability faster, because the regulatory conversation starts from demonstrated practice rather than a blank page.

There's an older name for this move. Twenty years ago Joel Spolsky called it commoditising your complement: make the thing sitting next to your product cheap or free, and demand for your product climbs. Microsoft did it to PC hardware to sell the operating system; Google funds free software to protect search. A bank handing rivals its governance plumbing is the same play. Drive the cost of the trustworthy-AI scaffolding towards zero, and the scarce thing you keep, the judgement and the private data, only gets more valuable. Openness is not the cost of the learning loop. It is the accelerator!

The honest catch

There's a sharper version worth sitting with. The economist Christian Catalini argues the moat is not even the learning loop; it's the verification underneath it, your ability to judge, approve and stand behind a result better than anyone else can. And he names the trap exactly: the loops a firm sees as sovereignty are the loops the labs see as their next training frontier. Open the wrong layer and you hand over the very judgement you were trying to protect.

That's the tension the Berlin room felt in its gut. So in healthcare the move is narrower than "open everything". Open the tooling and the workflow, and keep the verification, the clinical judgement, the human who signs off, firmly in your own hands. Give a ward team a half-built triage tool and let them break it, and you learn in a week what a closed pilot hides for a year. The call about what's safe stays where it belongs.

Here's a smaller experiment than open-sourcing your bank. Pick one internal tool you've been guarding, and this month hand it to ten people on your front line who didn't build it. Watch what they do to it. The discomfort you feel is the point; it's the sound of your loop speeding up.

💥 May this inspire you to open the thing you've been protecting, and learn faster than the people who kept it locked.