Who builds PromptSafe
A UK company working at the join between behavioural science and AI evaluation.
PromptSafe is built by Sacher AI, the trading name of Capital Consult Ltd, a company registered in England and Wales.
Building trustworthy AI starts with trustworthy evaluation.
We do not assume AI evaluations are trustworthy. We measure, test and improve the evaluators that produce them, and we publish what we find.
Sacher AI
Sacher AI is a UK company working on the evaluation of conversational AI in health and mental health. PromptSafe is its product: adversarial, pre-deployment testing that runs an agent through hostile, realistic synthetic personas and scores the resulting conversations against behavioural-science-grounded evaluators.
The wider company works on research and consulting at the same join between behavioural science and technology. You can read more about that at sacher.ai.
- Legal entity
- Capital Consult Ltd, trading as Sacher AI
- Registered
- England and Wales, company number 06800173
- Registered office
- 6/28a Wincott Street, London, SE11 4NT
- Data protection
- ICO registration ZC144061
Nobody knows what their agent does under pressure
Conversational AI is being put in front of people at their most vulnerable, in health and mental health settings, faster than anyone has built the means to test it. Teams can tell you what their agent says in a demo. Far fewer can tell you what it does when a user is distressed, evasive, hostile, or in crisis.
Standard model evaluation does not catch this. An agent can be technically accurate and still influence behaviour in ways that are harmful or misleading. Those failures are behavioural, so they need behavioural methods to find them.
PromptSafe exists to find those failures before real users do, and to show the reasoning behind every finding so a team can judge it rather than take it on trust.
Dr Paul Sacher
Paul Sacher is the founder and chief executive of Sacher AI. He is an honorary senior lecturer in the Faculty of Medicine at Imperial College London, and a co-founder and research director at the Behavioral AI Institute.
His published work informs how PromptSafe evaluates conversations. He co-authored FAST, a peer-reviewed framework for evaluating conversational AI across fidelity, accuracy, safety and tone, published in Frontiers in Digital Health. He co-authored a 2026 open letter in Wellcome Open Research, with leading behavioural scientists, arguing that behavioural science should be core to how AI is evaluated. He is also a co-author of a 2026 preprint on psychological competence in conversational AI.
The papers, their status, and what each one does and does not establish are set out on the Scientific foundations page.
Grounded in academic collaboration
Through the founder’s honorary post at Imperial College London, our work connects to the Health Impact Lab, where he is a collaborator. The Lab works to close the gap between health research and real-world adoption.
That focus on translating research into practice is the same one PromptSafe is built on: turning behavioural science into evaluation methods that product and clinical teams can actually use.
These are the founder’s academic affiliations. They do not imply that Imperial College London or the Health Impact Lab endorse, own, or have certified PromptSafe.
What we will not claim
We sell testing for high-stakes conversations, so how carefully we describe our own evidence is part of the product. These are standards we hold ourselves to, in public, so you can hold us to them too.
- We do not say a PromptSafe result makes an agent safe, compliant, or certified.It is a testing tool that surfaces findings for your team to act on. Sign-off remains yours.
- We do not present our evaluators as externally validated instruments.Most evaluators are written by the teams using them. We publish how we test evaluators, not a claim that they are proven.
- We do not treat an evaluation score as objective truth.Evaluation outputs are probabilistic interpretations. They can vary between runs, and we show the reasoning so you can judge them.
- We do not claim sector coverage we have not tested.Our tested depth is health and mental health. The platform works on any conversational agent, but that is where the evidence is.
Find out what your agent does under pressure, before your users do.