Designing for a System You Can’t Yet See
The organization that goes looking beats the one that executes the plan.
Most teams, the moment they decide to build something serious with AI, reach for the ritual they already know: the offsite, slides full of confidently drawn boxes and arrows, a roadmap with quarters marching across the bottom, and then the reorg to deliver it. Draw the destination, point the organization at it, march. It’s a familiar ritual – it’s worked for most of the things we’ve ever built – but pointed at AI, it runs almost exactly backwards.
The reason is what Ethan Mollick named the “Jagged Frontier.” The capability surface of LLMs is wildly uneven, and you cannot read its shape from the outside – the AI model nails the task you assumed was hard and fails miserably on the one you assumed was trivial, and the only way to find the edge is to walk up to it and test. So the target-state slide, for an AI-enabled product, is merely your best guess – not an actual plan based on knowledge and facts. You’re drawing a destination you’ve never actually seen.
And here’s the part leaders often trip over – and it sounds like a contradiction: even when you can’t draw the destination, your organization is still going to shape whatever you end up building. That’s Melvin Conway’s old observation (immortalized as “Conway’s Law”), the one software people quote – any group that builds a system ships a design that copies the group’s own communication structure. The seams in the product end up mapping the gaps in how people talk. The deliberate way to use that, popularized by Matthew Skelton and Manuel Pais in Team Topologies, is the inverse Conway maneuver: instead of letting your structure dictate the architecture by accident, you design the teams to produce the architecture you want and let Conway work for you. Want a modular product? Build modular teams with clean interfaces, and the product follows. It’s a genuinely powerful move – and it carries one quiet assumption that AI quietly demolishes. It assumes you can draw the architecture first.
Having thought about this for quite a while now, and having had the chance to discuss this with a group of leaders at a recent event in Hamburg, Germany, I believe AI shifts that assumption: In the AI era, you stop designing teams to match a target architecture, and you start designing them to match the discovery process itself. The org’s job is no longer to produce a system you’ve already drawn. It’s to find a system you can’t yet specify. And finding has a structure of its own – short loops, low latency between the person who discovers what works and the person who can ship it, tight coupling between experimentation and production. The inverse Conway maneuver doesn’t disappear; it changes what it’s aimed at. You’re still shaping the org to shape the outcome. But the outcome you’re optimizing for is the speed and quality of learning, not the fidelity of a blueprint.
A traditional product org can run discovery in one corner – a research team, a lab – and hand the findings down a chain to the people who build. With AI, that chain is where the value evaporates. By the time the insight clears three reviews and a quarterly planning cycle, the frontier has moved and the insight is stale. The thing you learned in March about what the model could do is a different thing by June. Learning that can’t be acted on quickly isn’t learning your organization actually has; it’s learning one person had.
For leaders in organizations, this means that we have to stop trying to write the target-state architecture before we’ve earned the right to – we simply don’t know it yet, and a confident wrong guess is worse than an honest blank. Go measure one number instead: how long does it take, in your company, for something one person discovers about what AI can really do to become how the company works? Weeks? Quarters? Never? Then go shorten that number, because it’s the only metric in this whole conversation that you fully control. Put the person who experiments and the person who ships on the same small team, in the same room, reporting to the same human. Give them a real sandbox and standing permission to use it – not a committee they have to petition. You’re not building the organization that executes the plan. You’re building the one that goes looking.
P.S. Curious about the best approaches on how to build a learning organization fit for an AI-driven future? My latest book “OUTLEARN – The Art of Learning Faster Than the World Can Change” might be useful.
@Pascal