For years, we treated innovation as something that needed to be carefully encouraged and structured. We created idea pipelines, innovation boards, internal sponsors and MVPs to help good ideas survive inside complex organizations.
And then AI agents entered the room.
Suddenly, innovation is no longer necessarily scarce. It can appear almost anywhere. Useful, enthusiastic, sometimes duplicated, sometimes half-baked, sometimes surprisingly good.
Welcome to the beautiful mess of agentic innovation!
The cost of building is shrinking
A few years ago, if someone had an idea for an internal tool, testing assistant or domain-specific accelerator, the path was predictable.
Someone would write a proposal, estimated budget, asked who will pay for it or approve it. Very often, the idea would die somewhere between “interesting” and “not now”.
Today, any motivated individual can often build the first working version before anyone has even noticed that the idea exists. That is both great and scary – great because it creates energy (people can now automate and improve things without waiting for a programme to approve them), and scary because energy alone is not a system.
At CROZ, we see agentic assets appearing in many corners of the organization. Some improve delivery, some support internal operations, others help with knowledge access. Some are born from customer needs, others from curiosity and enthusiasm.
I like that energy, and I must acknowledge one thing – the old innovation process is too slow for this new reality.
The new question is when to notice
The new problem is how to notice the right ideas early enough, without killing them too soon. This is a delicate balance.
If you (the organization) get involved too soon, you can slow everything down. A small creative experiment suddenly becomes a project, a project needs a sponsor who needs a plan which needs a budget. And the original spark disappears.
But if you get involved too late, you may end up with ten similar tools, unclear ownership, security questions and a growing pile of assets that someone will eventually have to maintain.
The organizations simply need a much better mechanism for rapid validation after something is born. Let people build, and then quickly ask the hard questions – the ones about value, risks, lifecycle.
Runtimes matter
Organizational governance is not the whole story. There is also a more technical issue that is easy to miss at first: runtimes.
Most of these new assets are agentic in nature. They do not just sit there as static tools: they use models, call APIs, they depend on prompts, memory, permissions, data sources, orchestration, logging, tracing and evaluation.
In the beginning, every team naturally builds its own little runtime around its own agent, and that works for a while. If every agentic asset brings its own runtime, we create fragmentation in many areas: handling identity, model routing, security assumptions, cost visibility, deployment patterns, different answers to the same governance questions.
We have seen this pattern before in IT – local optimization creates global complexity.
So the question becomes obvious: should every asset manage its own runtime, or do we need a shared agentic runtime layer? I think we do. Centralization is not always good, but some capabilities should not be reinvented every time. Identity, security, model access, observability, cost control, evaluation, deployment and lifecycle management are part of making agentic assets safe and useful in a real organization.
This is where platform thinking enters the story again. Or, at least, that is how we decided to approach the runtime challenge around agentic assets. My colleague Ivan Krnić wrote a great article about what we learned from building an agent control plane. Yes, that is an agentic asset as well.
The goal is to learn faster
Agentic innovation creates new skills, new offerings, new delivery capabilities, new technical perspectives and new enthusiasm. But if we do not adapt our processes, we will not harvest the value, we will only create a new kind of mess.
The better way to manage it means more freedom at the beginning, faster validation after the first version, clearer ownership, better lifecycle management and a pragmatic way to decide which assets stay internal, which become part of services, and which deserve a more serious business model.
This is not a solved problem. At least not for us, not yet.
But it is one of the most interesting problems we are working on right now. And I suspect many of you are facing the same thing, even if you use different words for it.
Finally, I would like to share some of the AI initiatives that made it through the first round of experimentation at CROZ. We have brought them together under the Carbon brand, showcasing how we use our own AI tools today to modernize, build and operate enterprise-grade systems and solutions.
