Trust, risk & governance in an AI-driven world
Attacks on people don’t need to get past our distrust. The gap between the trust we feel and what the evidence supports is enough. I give talks and training on how to see that gap and make it smaller – for people, teams and AI agents.
Ten sources with dependence 0.5 count as roughly 1.8 independent sources. The feeling says: confirmed ten times.
n_eff = n / (1 + (n − 1) · D). The felt curve is illustrative: familiarity grows with every repetition, evidence only with independent origins. A visualisation of the model, not a measurement.
Trust Is Engineered describes trust as a directed relationship with evidence, uncertainty and boundaries. For talks and training, I turn it into three questions you can ask next Monday.
One describes how we actually judge: familiarity, repetition, sympathy and time pressure. The other, what the available evidence supports. The difference is the calibration gap. It runs both ways: too much trust lets attacks in, too much distrust costs time, money and cooperation.
The goal is not more distrust, but a smaller gap.
New information changes more than one relationship. It travels on through connected people, sources and systems. Trusting service B does not mean trusting its subcontractors. Good processes decide where the influence stops.
Trust is not transitive.
Who is asking? Are they allowed to? Is it correct? A genuine voice answers only the first question. Even a familiar colleague or a helpful agent needs a scope for action that fits the action and its potential harm.
Permissions follow the action, not familiarity.
Research status: Trust Is Engineered is a research concept and mathematical model hypothesis, not yet empirically validated. Three derived papers are available as preprints.
Three situations from talks and exercises. Choose a request and think about what you would ask before approving it.
“Please transfer the money to this new account today.”
A familiar sender. Time pressure. Changed bank details.
Possible checking response: Call back on a number you choose yourself before approving.
“Send me the customer list. Management has approved it.”
A plausible reason. Sensitive data. A claimed approval.
Possible checking response: Establish approval and purpose independently. Share only what is required.
“The agent checked it. It can make changes in production.”
A convincing result. Automated checking. Consequential write access.
Possible checking response: Check usefulness per task, keep permissions narrow, approve consequential changes under control.
Every talk opens with a scene the audience knows and ends with a rule they can apply the next day. The modules work on their own or combined into a keynote.
Why do we trust – and when do we get it wrong?
Keynote, leadership groups, Executive Decision Workshop
What really protects against social engineering?
Team training, finance and HR, security awareness
How do we trust machines without going blind?
Agent governance workshop, IT and AI leads
20–30 minutes per module
For conferences, professional groups and leadership teams. One module or several as a keynote, with room for discussion.
Work through situations from your own day-to-day: practise follow-up questions, agree checking responses, clarify roles.
Executive Decision Workshop: turn one real approval in your organisation into testable rules. Preparatory call, about 90 minutes of workshop, a short results document.
| Typical awareness | With the model |
|---|---|
| Look out for spelling mistakes. | Look at the action being requested. |
| Check the sender very carefully. | For payments, call back on a number you choose yourself. |
| Be suspicious. | Be thorough at a few fixed points and relaxed otherwise. |
| One mandatory module a year. | Named reviewer roles and a team norm people can see. |
| Success means fewer clicks in the test. | Success means fewer wrong approvals and fewer rejected legitimate requests. |
Generative AI writes flawlessly; signals that cost an attacker nothing prove nothing anymore. In a randomised trial with about 19,500 employees, annual phishing training showed no measurable effect (Ho et al. 2025). That the right-hand column works better is a hypothesis to be tested in pilots.
On LinkedIn the model unfolds in twelve episodes, in German on Tuesdays and in English on Thursdays. Each episode is also a building block of the talks.
Follow the series on LinkedInI’m Johannes Rumpf: Chief Governance Manager in insurance, problem solver and part-time maths teacher.
For more than 20 years I have worked in IT, banking and insurance – from developer and system architect to information security governance. Today I turn regulatory pressure such as ISO 27001, DORA, NIS2 and the AI Act into controls that teams can actually run day to day. I have owned ISMS implementations in regulated financial organisations and supported audits.
In the classroom I meet the same question as in an audit: how do I know whether I can trust a result? Trust Is Engineered is my attempt at a testable answer. I put it up for discussion openly, because critique moves it further than agreement.
Part-time maths teacher
Fiducraft combines the Latin fiducia, trust, with craft.
Tell me briefly about the audience, the occasion and your question. I’ll come back with a proposal for a talk, training or workshop.
Write an email