We Built Factories When We Should Have Tended Gardens

There is something peculiar about the way we manage innovation. We ask people to challenge assumptions, then give them a methodology to follow. We encourage experimentation, but expect commitments to outcomes before the evidence is in. We celebrate learning from failure, while quietly rewarding those who deliver what they originally promised.

We have built considerable machinery around this work: funnels, stage gates, accelerators, venture boards and portfolio dashboards. Much of it serves a useful purpose. Organisations need ways to allocate resources and make decisions about uncertain investments. Yet the machinery can become so absorbing that we lose sight of what it was meant to help us do.

The question I keep returning to is whether our understanding is improving. Are we discovering what is true, recognising what matters and changing what we do as a result? Or are we becoming better at moving ideas through a process whose assumptions remain largely untouched?

Perhaps innovation is a capability we need to cultivate, and our dominant model of management makes that difficult. We built factories when we should have tended gardens.

How the factory became our model

The factory is one of humanity’s great achievements. Standardising work and coordinating activities made it possible to produce goods at a scale and consistency that would once have been unimaginable. When the desired outcome is understood, reproducing it reliably is valuable. We want medicines produced consistently and critical infrastructure operated safely. Predictability matters when deviation has serious consequences.

This logic shaped modern management: define the objective, design the process, allocate resources and measure performance against the plan. The difficulty comes when we apply it to work whose purpose is to discover something we do not yet understand. In innovation, we may not know whether a problem is worth solving, whether customers will change their behaviour or whether the proposed technology will work in practice. The most useful question may emerge only after we begin.

Yet teams are often asked to define a solution, quantify its potential and establish milestones before receiving resources to explore it. Their progress is then evaluated against commitments made when they understood the problem least. Activity starts to serve the defence of an earlier understanding, even when new evidence should give us reason to revise it.

Imagine a team that secures funding, completes a prototype and delivers its pilot on time. Its dashboard is green. Along the way, however, it discovers that customers do not behave as expected. The technology works, but integrating it into existing operations creates more friction than value. Another problem uncovered during research now looks more consequential than the one in the approved business case.

What should the team do with that knowledge? Continuing with the plan may satisfy every governance obligation while producing little value. Abandoning it may release resources for something more promising. A system can record both outcomes without rewarding the distinction between them. If stopping looks like failure and delivery looks like success, the team has a reason to protect the project from its own learning.

Stage gates and business cases solve legitimate coordination problems. But mechanisms for governing investment cannot tell us, on their own, how discovery should happen. We can track an idea’s journey through a process in great detail and still know very little about whether our understanding of the world is becoming more useful.

What if learning were the unit of progress?

A team begins with a belief: a customer has an unmet need, a technology can perform a task, or a different operating model will improve the work. It acts on that belief and encounters evidence. Sometimes the evidence supports it; sometimes the picture becomes more complicated. Occasionally the team discovers it has been asking the wrong question.

The consequential part is what happens next. Does the evidence change the team’s understanding? Does that change influence a decision? Does the organisation adapt what it is doing? An experiment can produce information without producing any of these changes. Interviews become presentation slides, results become reports, and lessons learned become documents that never return to the work.

I am interested in the relationship between what we believe, what we discover and what we do differently. Making that relationship visible would give us another way to examine progress. It would also make it harder to treat the recording of evidence as proof that learning has happened.

Consider a team that believes customers abandon a process because it takes too long. It could commit immediately to automation. Or it could examine that assumption first. Suppose research reveals that customers are more troubled by not knowing what happens next than by the waiting itself. The team can now test a change that makes the process clearer, rather than faster. Its original explanation was wrong, but its next decision is better informed.

A project system might describe this as a revised scope or a deviation from the plan. A system organised around learning would recognise it as part of the work. We would ask what evidence justified the change, what uncertainty remains and whether the intervention helps. The revised plan would carry a reason we could examine.

Learning still has to lead somewhere. A team could spend years refining hypotheses without creating anything useful. The value lies in the decisions and action that become possible, including the decision to stop. Discovery needs to be an accountable part of delivery, with enough room to change the commitments that evidence no longer supports.

Conditions shape the work

A formal operating model never tells the whole story of an organisation. Two teams can follow the same process and achieve different outcomes because of trust, incentives, experience or what people expect will happen if they make a mistake. These conditions are harder to put on a dashboard, but they shape whether the process works at all.

A culture of experimentation cannot be created simply by asking people to experiment. If failure carries a severe career cost, avoiding uncertain work is a sensible response. If annual budgets reward adherence to commitments, teams will learn to protect those commitments. If departments compete for resources, knowledge that could help the wider organisation may remain within local boundaries. The formal process can encourage learning while the surrounding conditions discourage it.

Complex adaptive systems offer a useful lens here: interactions and feedback can shape collective behaviour in ways that are difficult to understand by inspecting each component alone. The Santa Fe Institute’s work on emergent engineering explores interventions in adaptive systems, including the challenges of institutional design. Applying that lens to innovation is an interpretation, not evidence that a particular management approach will work.

Organisations are not literal ecosystems, and people are not plants. They have intentions, negotiate power and make choices no ecological metaphor can fully explain. Even so, the garden directs attention to conditions we might otherwise overlook. A gardener observes the soil and climate, considers what can thrive together, and adjusts as the garden responds. Growth cannot be prescribed in the same way as a production schedule.

This is active work. A gardener decides what to cultivate, protect, prune and remove. They cannot control every outcome, but they remain responsible for their interventions. The metaphor is useful because it asks us to attend to how a system responds, rather than assume that following the prescribed sequence will produce the desired result.

The wisdom of not forcing

There are echoes of this orientation in Daoist thought. Ziran is often translated as naturalness or spontaneity; wu wei as non-action. Both have richer and contested meanings. As the Stanford Encyclopedia of Philosophy’s discussion of Laozi explains, these ideas concern a way of relating to the world, rather than a simple instruction to do less. One interpretation I find useful is action sensitive to a situation’s conditions, without trying to force everything into an imposed form.

I would be cautious about turning that into a management technique. Philosophical traditions are not early versions of contemporary organisational science. Their value here is in challenging the assumption that more control necessarily means better leadership. Detailed plans do not remove uncertainty, and additional governance does not automatically improve a decision.

Sometimes the useful intervention is to remove a constraint or connect people who have been working in isolation. Sometimes it is to make evidence visible, stop an initiative or give a team enough autonomy to respond to something nobody could have anticipated centrally. Choosing among these interventions requires judgement and continued attention. Neither rigid control nor unrestricted autonomy removes that responsibility.

A question worth asking is what prevents a particular team from learning and adapting. The answer might involve a process, but it might also involve incentives, relationships or a decision that keeps being deferred. We need to understand the constraint before prescribing another method.

When intelligence becomes abundant

AI gives this question more urgency. Research, analysis, software development and documentation have demanded substantial effort. As some of that work becomes easier to produce, smaller teams can examine more alternatives and build more prototypes. The capabilities remain uneven, and the outputs still need scrutiny. But cheaper production changes where the constraints may lie.

When generating solutions becomes easier, deciding which problems deserve attention matters more. When software is easier to build, its fit with the people and systems around it becomes harder to ignore. When teams deploy agents independently, their actions still need to make sense together. An organisation can produce more work while becoming less clear about what that work is meant to achieve.

“Intelligence is abundant. Coherence is scarce” is a thesis I am working with, rather than a claim that expertise no longer matters. Domain knowledge, trustworthy evidence and the ability to act remain consequential. The point is that making one capability cheaper can expose a constraint elsewhere. More prototypes help when prototyping is the bottleneck. They help less when the organisation cannot choose among them or act on what they reveal.

We could use AI to generate more ideas for the funnel, write business cases faster and automate stage-gate documentation. Some of that would save useful time. But if the system rewards completion of predetermined activities, we may increase the speed at which teams produce evidence of progress without improving the decisions underneath it. We will have built a faster factory.

The opportunity I find more interesting is to reconsider what the system helps people notice and remember. Could it keep evidence connected to the beliefs and decisions it should affect? Could it help a team recognise when its next action no longer follows from what it has learned?

An operating system that remembers the reasoning

Imagine a team making its current beliefs explicit: the problem it is trying to understand, the assumptions its proposed intervention depends on, and what evidence would challenge them. As research and experiments produce observations, those observations remain connected to the beliefs. A decision retains the reasoning behind it, so someone joining later can understand why it was made.

AI agents might help organise evidence, surface contradictions and retrieve relevant learning from earlier work. They might also support prototyping or analysis. The people remain responsible for interpreting what they see, negotiating competing interests and making trade-offs. Better access to information does not settle what matters or determine whose interests should prevail.

As understanding changes, the system could help teams notice which priorities and activities need another look. Across the organisation, learning from one team might inform another without requiring both to use an identical methodology. Shared understanding could develop through connections between problems and evidence, as well as through reporting lines.

There are difficulties we should take seriously. Evidence is rarely unambiguous, and people disagree about what it means. Incentives and politics can outweigh information. AI can introduce errors or make a prevailing assumption look more credible than it deserves. Documenting beliefs can become another administrative task, particularly if teams have to perform their thinking for a dashboard.

Any useful implementation would have to earn its place in the work. It should make consequential reasoning easier to examine and reuse, with less effort than it creates. Otherwise, we will have added another reporting system to an organisation already struggling to act on what it knows.

The hypothesis behind Duyên

In Vietnamese, duyên carries associations with affinity, connection and the conditions through which relationships and possibilities arise. The word has different shades of meaning; what interests me here is how something valuable can emerge through encounters nobody could fully plan. A question meets evidence, a person encounters another perspective, and an opportunity begins to take shape.

Duyên explores what an innovation operating system might look like if it paid closer attention to those relationships. Its starting hypothesis is that teams making assumptions explicit, connecting evidence to decisions and adapting their work accordingly can make better innovation decisions than teams supported mainly by systems that track activities and milestones.

That is a proposition we can test. Do teams identify invalid assumptions earlier? Can they stop unpromising work with less wasted investment? Does learning travel usefully between teams, and do promising interventions become part of actual operations? Better documentation would be insufficient. We would need to see differences in decisions and outcomes, and understand what contributed to them.

There are good reasons the hypothesis might fail. Experienced teams may already do this effectively without additional technology. Recording beliefs might create more overhead than benefit. Leadership, trust and expertise may matter much more than the software. Organisational incentives may overwhelm improvements in how information is connected. These possibilities belong in the work from the beginning.

Duyên’s development should be subject to the same demands it makes of innovation teams. We need to meet real teams, understand their constraints, build interventions and examine what changes. We need to be willing to discard features, and perhaps the premise, when the evidence gives us reason to. Designing a definitive system before learning what helps would repeat the mistake this essay questions.

Factories still have their place

The gardening metaphor has limits. Organisations must serve customers, deliver products and meet obligations. Not every decision warrants an experiment. Once a team understands how to deliver something valuable reliably, standardisation can protect that value. A proven service needs a different discipline from a team searching for a viable business model.

The difficulty is recognising when to change modes. Practices that make execution efficient can restrict exploration when applied indiscriminately. Equally, an experiment should not remain an experiment forever because its creators prefer discovery to implementation. Something learned has to become something useful, and that transition brings responsibilities of its own.

Cultivating learning should make commitments better informed. It does not excuse a team from making them. We still need to decide when the evidence is sufficient to act, what risks we are prepared to take and what would cause us to reconsider. Those judgements cannot all be delegated to a process or postponed until uncertainty disappears.

Learning to tend

I suspect our attachment to process reflects how uncomfortable uncertainty becomes inside an organisation expected to justify investment. A process offers a path we can describe and defend. It makes an uncertain destination feel more manageable. The garden offers less reassurance: we have to pay attention, act, and remain prepared for a response we did not anticipate.

This may be a more demanding form of management. It asks us to distinguish work that needs more time from work whose assumptions no longer hold. To know when intervention helps and when its intensity becomes part of the problem. To revise our understanding without giving up responsibility for what happens next.

If AI makes intellectual production more abundant, I suspect this capacity will matter more. We may need to spend less effort increasing the volume of ideas and more effort understanding which deserve attention, what the evidence is telling us and what prevents us from responding. Duyên is one attempt to explore whether technology can help with that work. It is too early to know how much of the answer belongs in software.

We have spent decades developing machinery for innovation. Some of it remains essential. But before improving the machinery again, I would like us to ask what we are trying to grow, what conditions it needs, and whether our way of managing is helping it take root.

---

Duyên is an emerging experiment within Nya, exploring how humans and AI agents might help teams learn, coordinate and act under uncertainty. Its starting hypothesis is that making beliefs, evidence and decisions visible and connected can improve innovation decisions. The work is to discover whether that is true.