About
We are new. The people are not.
An applied AI research institute, based in Sweden.
What we are
900 Labs studies how AI agents behave with the people who rely on them, check them and try to steer them, and puts agents to work inside process-heavy businesses, measuring what they change against a comparison. We test agents for security before they go live. We publish research under a method that writes down what would prove us wrong before we start. One question, asked in the work we sell and in the studies we run.
What we are not
- Not a consultancy paid by the hour.Each engagement is scoped to the deployment, one process or several, and any stage can end in a no.
- Not a software vendor.We put agents on a process and run them. The nine open-source repositories are free.
- Not a fund.See Structure, below.
- Not an accelerator.“Applied” means we deploy what we study.
Team
Sam Rusani
Chief Executive
More than fifteen years building, investing and exiting across four continents. General Partner at Beyond the Bounds, where the portfolio is fully deployed. Based in Sweden.
Stuart Yeoman
Chief Investment Officer
Chief Executive of Farringdon Capital, with $2.75 billion under management. Degree in economics and financial services. Regulatory authorizations across several jurisdictions.
Dr. Pontus Inerup
Head of Medical Research
Dual board certified in psychiatry, 2024, and family medicine, 2012. Practicing psychiatrist. Owns a specialist care facility that is a formal clinical partner of 900 Labs. Engineering degree. Leads our studies on AI and mental-health safety, and is principal investigator on our evidence-synthesis work.
Benny Lundström
Head of AI
AI and machine learning systems architect. Has built production inference pipelines at scale. Founder of AI Hubben in Stockholm, building AI systems since 2016.
Venture Partner
John Trew
Venture Partner
Treleaven Capital. Deal sourcing and diligence in oil, energy, mining and logistics, and market entry across Africa and Asia.
Structure
900 Labs is an operating company. A separate vehicle for buying businesses that we would then transform is planned, and it is not formed. Nothing on this site concerns it, and nothing on this site is an offer to sell or a solicitation of an offer to buy any security.
How we keep advice and acquisition apart
We advise companies, and we would transform the businesses that vehicle buys. That is a structural conflict rather than a hypothetical one, so it has written rules rather than good intentions.
- Confidential information obtained in an advisory engagement is never used for acquisition sourcing, valuation or negotiation without explicit permission.
- Acquisition targets are sourced separately from the research network.
- Where we hold an ownership, commercial or financial interest in a technology, a vendor, a transaction or a recommendation, that interest is disclosed before the recommendation is acted upon.
- We do not engineer advisory conclusions to manufacture acquisition opportunities.
These were written before anyone offered us the chance to buy a client cheaply, which is the only time they are worth writing.
The Truth Floor
Every other part of our operating standard can be relaxed under pressure, with the deviation logged, reasoned, and reviewed afterwards against what happened. One part cannot. No cash-flow exception, no founder exception, no strategic exception.
900 Labs will not:
- tell a client we expect an outcome we do not actually expect
- represent weak evidence as strong evidence
- present correlation as causation
- hide material uncertainty because it makes a sale harder
- conceal a material conflict of interest
- knowingly sell work whose expected value to the client is negative
- manipulate a research conclusion because a commercial outcome depends on it
- continue defending a thesis after the evidence has materially changed
Models retrieve. Code calculates. Humans adjudicate.
Our pipeline uses multiple AI systems, on different model families, running separate retrieval lanes against the same question. They are retrieval and analysis tools, not evidence. Agreement between models raises nothing on its own, because two models reading the same underlying study are one source, not two.
Every quantitative transformation, every table and every figure is produced by version-controlled code rather than by a model. Every load-bearing judgment is made by a person whose name is attached to it.