Applied AI Research Lab · Sweden
We study how AI agents behave with people, and put them to work.
An agent’s code decides what it can do. The people around it decide whether it works: who relies on it, who checks it, and who tries to steer it. We research that side, and we deploy agents in process-heavy businesses such as accounting and logistics firms, with the gain proved against sites that have not started yet.
Who relies on it
ResearchA chatbot answering someone in psychosis
WorkRun: agents a team depends on every day
Who checks it
ResearchAI-assisted review of medical evidence
WorkProve: the gain, against sites not yet live
Who tries to steer it
ResearchWho may direct an agent, and who answers for it
WorkSecure: an invoice with a hidden instruction
What we do
Three ways we work. Each one can end in a no.
01 Deploy
Agents on one process, in three stages.
Start with one process, or several at once. Scope and price follow the deployment.
How Deploy works →02 Secure
Every agent is a new way in.
An invoice with one hidden line can read as an instruction, and an agent that takes it as an order pays the wrong account. We test agents the way an attacker would, before they go live: AI security reviews and agent red-teaming, with a specialist offensive-security partner.
What a review covers →03 Transform
The whole business, not one process.
For an owner who has decided to change the business: the same three stages, applied process by process, with the margin as the measure.
How Transform works →How Prove works
Measured, not asserted.
Sites go live in waves, with the order drawn by lot and the decision rule written down before the first wave. The sites still waiting are the comparison, so the gain is measured against them rather than against a forecast.
If the measured gain does not clear the rule, the work stops there. You keep the baseline and the comparison.
Research
Three research lines, one question: what happens between AI agents and people.
We take that question where it is hardest to answer and most costly to get wrong.
Line 1
How agents perform in real work
Our first Report concluded that the binding constraints on enterprise AI are organizational rather than algorithmic. The Agentic Enterprise tests that, with its kill criteria written down first.
- The AI Execution Gap900 Report · published March 2026
- Nearly half of executives pulled back on AI agents. Half of what?900 Note 001 · published 9 August 2026
- The Agentic EnterpriseEvidence review · protocol approved · registration on OSF next
Line 2
AI and mental-health safety
A chatbot answering someone in psychosis is one of the hardest tests of how an AI system behaves with a person, and a psychiatrist can judge what a safe answer is.
- How chatbots answer psychosis, in SwedishLed by Dr. Pontus Inerup · protocol in preparation
Line 3
AI-assisted evidence synthesis
Whether AI-assisted review of medical evidence holds up when people make every judgment that carries weight.
- Drug repurposing for treatment-resistant depressionDr. Pontus Inerup, principal investigator · grant application in preparation
From the next Report on, every 900 Report is preregistered, independently verified and published with its error rate.
Two kinds of data, kept apart
Research data
From our studies, including those with patients, under academic and ethical governance. It never feeds our products.
Deployment data
From the agents we run, under each client’s contract. It improves our agents, and it is never sold.
Nearly half of executives pulled back on AI agents. Half of what?
Two figures traveling as a single sentence turn out to come from two different surveys, of two different populations, with revenue thresholds an order of magnitude apart.
Read the Note →
Team
Who answers for the work.
Sam Rusani
Chief Executive
More than fifteen years building, investing and exiting across four continents. Based in Sweden.
Stuart Yeomans
Chief Investment Officer
Chief Executive of Farringdon Capital, with $2.75 billion under management.
Dr. Pontus Inerup
Head of Medical Research
Practicing psychiatrist, dual board certified. Leads our studies on AI and mental-health safety.
Benny Lundström
Head of AI
AI and machine learning systems architect. Founder of AI Hubben in Stockholm.
Credentials only. The full team, with our venture partner, is on the About page.
Where we are
What exists today.
A firm that intends to publish its error rate should be able to say plainly what it has built.
- 01A written method and operating standard.
- 02One Report and one Note published.
- 03The Agentic Readiness Diagnostic, built and audited against sixteen adversarial cases. Independent certification is next.
- 04Nine open-source repositories, public.
- 05Five named people, and what each has done.
One thing to know
Our published Report predates the current method. It was not preregistered and it carries no error rate. The next one will carry both.
Start with one process.
Write to us with the process that eats the most time. The first call contains our qualification questions rather than a pitch, and if the process is not worth measuring, we say so.
Or run the free diagnostic on one workflow. The diagnostic opens after independent certification. Tell me when it opens.