EU AI Act risk intelligence
Margelis.ai turns weak signals — model incidents, vendor drift, regulatory movement, operational telemetry — into a single reproducible, euro-denominated risk figure. White-box from input to output. A human signs every number.
Estimated annual AI risk exposure — example
€1 420 000
band €0.9M – €2.1M · confidence 80%
EV-2026-031 · 14 signals · method v2.3 · reproducible from evidence file
Most AI risk tools return a traffic light or a 0–100 score you cannot audit. Margelis.ai works the other way: every number decomposes back to its evidence, and every step in between is inspectable.
The method, in one line
Many weak, noisy signals are quantified into one euro number a human signs — and that anyone can re-run from the evidence file.
01Signals
weak · noisy · many sources
02Model
White-box method
priced in euros · method v2.3
inspectable · documented · versioned
03Signed €
€1 420 000example
band €0.9M – €2.1M · confidence 80%
A human signs every number.
EV-2026-031 · 14 signals · reproducible
Model and vendor incidents, audit findings, regulatory movement, internal telemetry, contract exposure. Each signal is logged with source, date, and weight.
every signal carries a source receipt
Signals map to loss scenarios through a documented, versioned methodology — frequency, severity, and exposure assumptions are all visible and challengeable.
method card published per version
No figure reaches your board without expert review. The reviewer's name and the full evidence file ship with the number. Rerun it yourself; you get the same result.
human-in-the-loop · reproducible
The Digital Omnibus shifted key EU AI Act enforcement dates — but obligations for high-risk systems are coming on a fixed horizon, and supervisory expectations are forming now. Boards that can already state their exposure in euros enter that period negotiating; boards that cannot enter it reacting.
In force today
The earliest AI Act obligations already apply. Inventory gaps here are findable in any due-diligence exercise — by regulators, insurers, or acquirers.
2027–2028
Under the revised Digital Omnibus timeline, the heavy compliance load for high-risk AI systems arrives in this window. Quantifying exposure now turns a future fire drill into a budget line.
Every quarter in between
New vendors, new models, new incidents. A one-off compliance report ages in weeks. A reproducible number can be rerun whenever the estate moves.
One pilot engagement produces a complete, audit-ready package — not a dashboard subscription you have to interpret yourself.
Euro-denominated annual exposure with confidence band — one figure, defensible in front of auditors, insurers, and supervisors.
Every signal, weight, and assumption behind the number, in a form a third party can rerun and verify.
The versioned methodology document: what is modelled, what is excluded, and what would change the number.
The ranked list of interventions that reduce the number — each priced against the exposure it removes.
What the front sheet looks like
Estimated annual AI risk exposure
€1 420 000
band €0.9M – €2.1M · confidence 80%
Signals behind the number
Margelis.ai is built and operated by Rūpestėlis Holding UAB in Klaipėda, Lithuania — a holding that runs its own fleet of production AI agents alongside six veterinary clinics. The risk methodology was not written in the abstract: it is the same discipline we use to monitor, diagnose, and account for our own AI systems daily.
Data stays in the EU. The methodology is white-box by principle, not as a feature tier. And every number that leaves the house is signed by a human who can defend it.
EU-built · EU-hosted · human-signed
Each pilot is scoped to one business unit or AI estate, so pricing is agreed after a short scoping call. Write one sentence about your AI estate and we reply with a fixed-scope proposal.
No. A pilot is an engagement, not a SaaS deployment. We work from the signals you can share, and you receive a number, an evidence file, a method card and an action register.
Yes. Every figure decomposes back to its signals, the methodology is documented and versioned, and a named human signs it. A third party can rerun it and reach the same result.
Most platforms hand you a dashboard and a score to interpret yourself. Margelis.ai gives you one euro-denominated figure with the full evidence behind it — a result you can act on, not another subscription to manage.
A pilot delivers your first number in about 30 days, covering one business unit or one AI estate.
Yes. Because it is reproducible, you can rerun it whenever your AI estate changes — new vendors, new models, new incidents.
Data stays in the EU. The methodology is white-box by principle, and no number leaves without a human signature.
Rūpestėlis Holding UAB in Klaipėda, Lithuania — a holding that runs its own production AI agents alongside six veterinary clinics. The risk methodology is the same discipline we use on our own systems.
A pilot covers one business unit or one AI estate: signal collection, first euro figure with evidence file, and a board-ready briefing. Write one sentence about your AI estate — we reply with a scoped proposal.