*Note: I used an LLM as a thinking and drafting aid for this post. The views are my own.*
**TL;DR:** The field increasingly agrees that AI safety undervalues founders and operators. I've spent 12 years building the operational scaffolding for early-stage non-profits in other sectors, and I'm now learning the AI safety context firsthand. Here are four operational traps that catch new orgs in year one — governance, the "boring" foundations, coordination cadence, and hiring order — and why getting them right frees researchers to do research.
*Writing this to think out loud and invite correction — I'm newer to the AI safety context than to operations, so if I've misread how these dynamics play out in this field specifically, I'd genuinely value the pushback.*
I think this matters more than it looks, because of a quiet pattern: research orgs tend to be led by researchers. But founder skills and researcher skills aren't the same thing. So, in year one, the domain expert often ends up doing operations — struggling, and at the expense of the research only they can do.
I've spent 12 years building operational infrastructure for early-stage, resource-constrained organisations in other sectors — from a white canvas, where little or nothing existed. I'm now learning the AI safety context firsthand, working with a new AI safety community as it sets up. I don't have experience of AI-safety-org war stories. But the year-one operational traps are remarkably consistent across sectors, and they're worth naming:
1. No governance backbone. No board, no decision rights, no basic policies. It feels premature — until a funder, a hire, or a conflict forces the question and there's no structure to absorb it.
2. The boring “foundation” gets deferred. As fractional COO at a non-profit fund, I built the compliance and governance scaffolding from scratch — GDPR, AML/KYC policy with external counsel, board packs and decision-tracking. None of it was visible work. All of it was the difference between an org that could legally take money and hire people, and one that couldn't.
3. No cross-functional cadence. As soon as there's more than two people, work needs rhythm — who's doing what, by when, and how it connects. Without it, the founder becomes the single point of coordination for everything.
4. Hiring in the wrong order. Hiring for visible gaps instead of the highest-leverage one — which, for a researcher-founder, is often someone to own operations so they can return to the research.
None of this is exotic. It's exactly what incubators like Catalyze exist to address — they explicitly help founders with strategy and with finding co-founders who complement their skills. The structural fix is simple to state: let the researcher do research, and bring in someone whose job is to build the organisation around them.
That's the work I care about. If you're starting an AI safety org and the operational layer feels like a fog, I'm always happy to compare notes.
