AI safety operations
I now want to use 13 years of product and operations experience to help reduce existential risk from advanced AI. My current hypothesis is that I can contribute most where a lack of senior execution capacity is limiting important AI safety work.
Problem statement
I want to work on reducing existential risk from advanced AI, particularly loss of control.
01
As advanced AI systems become more capable and autonomous, humans may become less able to monitor, constrain or intervene in their behaviour.
A loss-of-control scenario is one in which advanced AI operates outside anyone's control, with no clear path to regaining it. (International AI Safety Report, 2026)
The AI Security Institute tracks precursor capabilities including self-replication and strategic underperformance during evaluations. (AI Security Institute, 2026)
02
The concern is the potential scale and irreversibility of the outcome.
Severe loss-of-control scenarios could lead to the permanent marginalisation or extinction of humanity. (International AI Safety Report, 2026)
Experts disagree substantially about the probability. The severity and the lead time required for preparation are why the risk warrants serious attention. (International AI Safety Report, 2025)
03
Safety interventions depend on organisations being able to execute effectively.
AI safety organisations report shortages of experienced managers and operational leaders who can coordinate complex work and translate strategy into execution. (80,000 Hours, 2026)
Beacon describes the catastrophic-risk field as operationally underserved. (Beacon)
This is the bottleneck I want to work on.
Theory of change
My theory of change is that I can help reduce catastrophic AI risk by strengthening the organisations doing important AI safety work. I want to contribute where a shortage of experienced operational leadership is limiting their progress.
AI safety organisations need people who can manage teams, execute strategy and run complex programmes. Experienced operators with enough AI safety context remain difficult to hire. (80,000 Hours, 2026)
Strong operational capacity should help an organisation make better decisions and turn strategy into work that gets done. The operator's impact includes some of the additional effectiveness of the people and projects they enable. (80,000 Hours, 2026)
Removing operational bottlenecks gives specialists more time for their comparative advantage. Beacon describes infrastructure that lets new projects start within weeks rather than spending six to twelve months setting up independently. (Beacon)
The safety impact comes through work such as AI control, evaluations, governance, security or institutional preparedness. Apollo Research uses a similar causal approach to ask how interventions change whether dangerous systems are detected, constrained or deployed. (Apollo Research, 2023)
If these organisations become more effective and their interventions influence real decisions, dangerous AI systems should be more likely to be understood or constrained before causing irreversible harm. (International AI Safety Report, 2026)
Key assumptions
The role must be a real bottleneck, and I must add more value than the next-best candidate or use of my time. (80,000 Hours)
Improved processes must help important projects begin sooner or free specialists to do more valuable work. (Beacon)
Additional capacity has little value if the work it supports does not affect an important part of the path from advanced AI to real-world harm. (Apollo Research, 2023)
Research and evaluations only reduce risk when relevant actors use them before dangerous systems are developed or deployed. (Apollo Research, 2023)
Comparative advantage
My current hypothesis is that my comparative advantage lies in senior operations and programme roles where an important goal is clear, but the systems, processes and ownership needed to achieve it are not yet in place.
I have repeatedly worked on problems where there was no established playbook. At Skeepers, I built a new country operation from the ground up, including research, recruitment, partnerships and day-to-day systems. At GetFound, I took an untested idea, identified the key assumptions behind it and built a small-scale version that allowed us to learn whether the approach worked in practice.
A recurring part of my work has been creating the conditions for other people to do their most valuable work. At Smily, I clarified ownership across four areas, improved coordination between teams and built simpler ways of planning and reviewing work. This reduced delays by 30%. I also recruited and supported five Product Managers so that more decisions could be made by the people closest to each problem.
My default approach is to treat a proposed process as a hypothesis rather than assume it is the solution. At GetFound, I used more than 20 interviews to understand which assumption mattered most before investing further effort. At Smily, I introduced continuous feedback from users so that priorities could be adjusted based on evidence rather than internal assumptions.
I have often worked on problems that required people with very different perspectives to act together. At Banco Santander, I coordinated security, technology, finance, external providers and teams across Spain, Mexico and the USA while introducing new identity-verification and financial-crime systems under regulatory constraints. This kind of translation and coordination seems particularly relevant in AI safety, where research, engineering, policy and operations often need to work closely together.
More recently, I have focused on identifying where AI can remove repetitive work without removing important human judgement. At HomeExchange, I built reusable AI workflows used by several teams and introduced AI-assisted triage for support issues. I have also trained more than 300 employees at organisations including TAG Heuer, part of LVMH, and LHH to use internal AI systems, and built automations covering a large share of my own recurring operational work.
Selected work
Six examples of decisions I made, the work that followed and the measurable result.
Swipe to explore all six examples
Skeepers provides influencer marketing software for brands. I tested local demand, adapted the model and built the Spanish operation.
30+ brands, a four-person team and a playbook later used in Italy and Germany.
Read the case study Resource allocationA smaller Smily product generated less than 1% of revenue and used around 20% of ten engineers' time. Although we expected it to grow, the same capacity could create more revenue and customer impact in the main platform.
I recommended closing it, recovered engineering capacity and moved four support colleagues to the core team.
Read the case study Programme designCodeOp trains people for careers in technology. I recruited Product Managers from companies such as Amazon and Expedia, then worked with them to develop realistic case studies. This also helped participants build their professional network.
The revised course improved NPS. Separately, around 80% of the following cohort found relevant work within two to six months.
Read the case study Talent developmentWhile leading Qwist's banking-app team, I improved the wider management environment and created a structured transition from QA into Product for one colleague.
Overall team satisfaction rose by 40%; her individual satisfaction rose by 60% and she later became a Lead Product Manager.
Read the case study Testing assumptionsGetFound is a reverse recruiting platform. I built the smallest version needed to test whether candidates and employers would use the model.
600+ users, first placements and a CHF 2.2 million funding round.
Read the case study Human oversightFor Smily's property management software, I changed automatic replies into AI-generated drafts after users told us they wanted final approval.
More than 50% adoption in month one and up to two hours saved per day.
Read the case studyCurrent work and career capital
I am building field knowledge and testing it through applied work. Structured career programmes and conversations with practitioners help me update where I could add the most value.