Vigilens Joins the NORA Startup Network

– How did the Vigilens adventure begin?

– It began with a red eye. I saw the movie 2001: A Space Odyssey far too young, and HAL 9000 stayed with me: a machine that never once broke a rule, and still ended up doing the worst thing possible. That is the whole problem of AI governance in one character. It is also why the Vigilens logo is a lens, an eye that watches the system rather than the paperwork.

Professionally, I came to it from the other direction. I have spent around fourteen years in fintech and regtech, as a CTO and as a founder: 7 years in Oslo, then P2P lending in Saudi Arabia, then AML and KYC compliance software. I have built the tooling that lets a regulated company prove it is doing the right thing. But that work was always locked inside the financial industry.

Two things made Vigilens the obvious next build. First, most AI innovation today happens in small companies, and small companies have no access to governance tooling. Every serious platform in this market is enterprise-only, sales-gated, six figures. I sat through around thirty competitor demos; one required an NDA before they would even show me the product. Nobody offers a low-threshold self-serve front door, the way GitHub or Figma do. The EU AI Act is about to make that gap impossible to ignore, because the obligation lands on the small provider exactly as it lands on the large one.

Second, and this is the actual reason Vigilens exists: everyone else works outward-in. Vanta and Drata ask how to protect your systems from the world (threats, attack surface, security holes) and their AI modules inherit that shape as checklists. Vigilens works inward-out: how do we protect the world from your AI system, and how do we prove to an auditor that it does not behave outside acceptable bounds?

– What is the foundation of your approach to responsible AI?

– Four layers, and they map deliberately onto the Guaranteed Safe AI framework (Dalrymple, Yoshua Bengio et al., arXiv:2405.06624).

Classify builds a model of the AI system: What it does, who it affects, what scenarios it can enter. Controls is the safety specification. Evidence is the verifier: we pull real artifacts from GitHub, Jira, MLflow and Datadog and test them against the specification. Audit Pack is the proof certificate.

Two things about that specification layer matter. It is normally a regulatory framework (EU AI Act, ISO 42001, ISO 13485, GDPR, ISO 27001, NIST AI RMF, SOC 2) but it does not have to be. It can equally be a company's own internal AI policy, or an open risk database such as the MIT AI Risk Repository, which catalogues the ways AI systems have actually failed. Regulation is one kind of spec. Institutional memory is another.

The bigger difference is the unit of assessment. Everyone else certifies a company. We assess an AI system. A company-level certificate tells you an organisation has processes. A system-level assessment tells you what this specific model, in this specific deployment, is actually doing. That is the only unit you can reason about predictively, and it is what makes the World Model layer possible: once you have modelled the system and its scenario space, you can move from detect-and-react to predict-and-prevent, anticipating where a system could step outside acceptable bounds before it does. You cannot get there from a questionnaire.

And the output is deterministic. Every assertion returns one of four verdicts (PASS, CONTRADICTS, PARTIAL, INSUFFICIENT) pinned to a named artifact, rolled up by a fixed engine, and frozen as a record. Which is the honest answer to "why not just ask Claude or ChatGPT?" You can ask a general model whether you comply and get a fluent, well-reasoned answer that is slightly different every time you ask, sourced to nothing, and worth nothing in front of an auditor. Compliance is not an opinion about your system. It is a reproducible claim about a specific artifact, and it has to survive being challenged a year later. That is an infrastructure problem, not a chat problem.

We would be glad to host master's projects, and an industrial PhD is something we would seriously consider.

– How would Vigilens like to collaborate with academia?

The World Model layer is where we most want academic partners. Mapping an AI system's scenario space, formalising what "acceptable bounds" means for a given deployment, and reasoning about it well enough to predict a failure before it happens is genuinely open research. The GSAI line of work is the closest theoretical anchor we have found, and it is a research programme, not a product spec.

Three other areas we would open up:

1. Validation. We produce deterministic verdicts. How well do they agree with expert human auditors? That is a measurable question and nobody has answered it. We would happily put our engine in front of a proper inter-rater study.
2. Decomposition. Turning a regulatory clause into atomic, falsifiable assertions is the hardest craft problem in the product. ISO 13485 alone took us 31 controls and 98 assertions. There is real methodology to be formalised there.
3. Incident-to-control mapping. Connecting documented AI failures to the specific controls that would have caught them, which is the bridge between risk repositories and enforceable governance.

We can offer something concrete in return: a working platform, real frameworks, and anonymised assessment data from live systems. That is hard to get in academia. We would be glad to host master's projects, and an industrial PhD is something we would seriously consider.

– What can we expect from Vigilens in the future?

– The World Model is the headline, but others sit alongside it:

1. From point-in-time to continuous. Today an assessment is a run. The direction is governance that executes in the build pipeline, every commit, every retrain, every release, so a failure is caught as a build failure rather than as an audit finding six months later. Governance that does not run in the pipeline is not governance, it is documentation.
2. Agentic systems. Governance frameworks were written for models that produce outputs. Agents take actions, chain tools, and operate with far less human oversight. Almost nothing in the current standards landscape handles that properly, and it is arriving faster than the standards are.
3. Framework depth, driven by demand. We support eight frameworks today. We are deepening them one at a time, in the order real customers ask for, rather than claiming breadth we cannot defend.
4. Sovereignty. European AI governance running on European infrastructure, assessed against European rules, without customer evidence leaving the jurisdiction. That is increasingly not a nice-to-have.