I’ve been thinking a lot recently about what it takes to build an AI-native company, and more specifically, what it takes for NEXT Amsterdam to become one. I ran a lot of experiments with Hey, Astra!, our latest “startup” led by an AI Agent running on OpenClaw. What I’ve been coming back to, is that it’s incredibly hard to explain to an agent how you work. The agent misses the data it needs to operate, so it starts to hallucinate.
Innovation is a knowledge-intensive practice. Our IP lives in frameworks, playbooks, coaching patterns, client adaptations, and a decade of accumulated lessons. Which makes the question “how do we implement AI and AI agents?” both urgent and interesting.
What NEXT Amsterdam does (quickly, for the uninitiated)
We’re an innovation agency based in Amsterdam. For the last ten-plus years, we’ve helped corporates build new business models and coached startup teams through the messy process of validation. Going from a rough idea to something people actually want and pay for. Our frameworks (the NEXT Canvas, Lean Innovation, Innovation Accounting) have been applied at companies like DHL, Sony, and ABN AMRO, and we’ve put our thinking into three books authored by my partner Esther Gons. We also built GroundControl, software that operationalizes the whole methodology.
AI-assisted vs. AI-native: the distinction that matters
Greg Isenberg wrote something recently that I keep coming back to:
An AI-assisted company uses AI at the edges. An AI-native company redesigns the center.
An AI-assisted company asks: Where can we add AI to save time? An AI-native company asks: How should this workflow exist if agents are doing the first 80%?
That second question changes everything.
Most companies, including us until recently, are essentially illegible to machines. The knowledge is scattered. The process lives in people’s heads. The playbook is a PDF that nobody has fully read. The real answer to “how do we run a lean sprint with that corporate team at X?” is to ask Esther or Timan.
That doesn’t scale. And more importantly, an agent can’t run on what’s in people’s minds.
To become AI-native, we first had to make NEXT legible for agents. We had to turn ten+ years of institutional knowledge into something a model could actually work with.
That’s the first step. The hardest step. And it all starts with a wiki.
Step 1: Build the wiki
We took our cue from Andrej Karpathy’s approach to LLM-maintained knowledge bases — the idea being that raw sources are preserved, and an LLM incrementally builds and maintains a structured, linked wiki that compounds in value over time. One clean operating layer that an agent can read, reason over, and write back to.
We used OpenAI’s Codex to set it up and start feeding it.
We started with our blog posts, more than 100 of them, going back to 2013. Then we added our core playbooks: the NEXT Canvas, Lean Innovation, the Experiment Loop, thirty-plus individual experiment playbooks. Then Esther’s three books: The Corporate Startup, Innovation Accounting, and Open Innovation Works. Then the playbooks we’d written for actual clients, the DHL Start-up Lab intrapreneurship program across three years and three phases, coaching engagements, and the tech due diligence work we’ve been doing.
Every source sits in its own layer. Core IP is separate from client adaptations. A client-specific compromise doesn’t silently overwrite our recommended method. When we did something differently for DHL, the wiki records why, and whether we agree with the adaptation, and whether it’s safe to reuse.
The result is a knowledge base that understands not just what NEXT does, but how we do it, why we made certain calls, and what we’d do differently next time. It’s a lot of work, but this builds the basis future AI agents can work from.
The moment it clicked
Last week I had a coaching call with a new portfolio company. Community-led commerce platform, early stage, smart team. I uploaded the transcript into the wiki.
What happened next is the reason I’m writing this post.
Codex didn’t just summarize the call. It filled in the NEXT Canvas based on what was said. It diagnosed the startup’s stage — correctly identifying that they’re still in the Problem phase, because the customer segment and core problem shifted three times during the conversation, and there’s no evidence yet that the right customer has the right pain strongly enough. It then generated an initial question plan for customer interviews, drawing on the principles from The Mom Test that we’d ingested into the wiki: focus on past behavior, not hypotheticals; ask about their current workflow, not their interest in the startup’s features.
We’ve coached hundreds of startups. That pattern recognition — “this team thinks they’re ready for Solution, but they’re still in Problem” — normally lives in our heads. Now it lives in the wiki. And the agent used it.
That’s not AI-assisted. That’s the beginning of something different.
What this wiki makes possible next
This is just the foundation. But you can already see what becomes possible when the knowledge is clean and structured.
Automated coaching prep. Before every coaching session, an agent reads the team’s most recent canvas, their experiment log, and the comparable patterns from other ventures we’ve coached. It produces a briefing: where the team is on the NEXT Canvas, what the riskiest untested assumption is, and what we should push on in the next 45 minutes. The coach walks in prepared, not improvising.
Playbook customization at speed. A corporate comes to us wanting to run an intrapreneurship program. Right now, building a custom playbook takes weeks of workshops and document work. With the wiki, an agent can compare the client’s context against how we adapted the program for previous clients, surface the decisions we had to make, and generate a first-draft custom playbook in hours. We still review it. We still own it. But the starting point is weeks’ work done in an afternoon.
Experiment design on demand. A startup tells the wiki: “We’re in Problem stage. Our riskiest assumption is that independent beauty brands care enough about referral-based growth to change their workflow.” The agent maps to the right experiment type from our 30-plus experiment playbooks, designs the test, pre-fills the success criteria and timeline based on comparable experiments we’ve documented, and flags the failure modes we’ve seen before. The team doesn’t start from a blank page. They start from NEXT’s accumulated pattern library.
Why this matters for corporates
Most large companies have the same problem we did, just at ten times the scale. Years of institutional knowledge scattered across Sharepoint folders, Confluence pages that are likely unmaintained, slide decks with no author, onboarding docs that are six years out of date, and processes that live entirely in the heads of the people who’ve been there longest.
Making that knowledge legible to machines is the first, foundational step to becoming AI-native. And it’s the kind of work we know how to do: we’ve been helping organizations structure their innovation practice for over a decade.
If you’re a corporate trying to figure out where to start with AI in your innovation practice, this is the step most people skip. They buy the tools first. They should build the operating layer first.
That’s what we’re doing at NEXT Amsterdam. And we’re happy to help you do it too.