The Deadline
In August 2026, the EU AI Act’s obligations for high-risk systems bite in full. The fines reach up to 7% of global revenue, and the required paperwork is substantial: risk classification for every AI system you operate, technical documentation across all 14 Annex IV elements, a conformity declaration, and post-market monitoring that doesn’t stop after deployment.
For enterprises, this is an inconvenient cost of doing business — they buy $50k+ GRC suites and staff compliance teams. For small and mid-sized companies actually building on AI, the options have been worse: a consulting engagement to fill out forms, or a spreadsheet and a hope. Neither scales, and neither survives the question every deployer eventually gets asked: how do you know the update didn’t break anything?
What aissurance Does
aissurance is our answer: a Compliance-as-a-Service platform for the EU AI Act, built for SMEs, from €99/month. It handles the paperwork so teams can keep running their business:
- Risk classification engine — per-system classification against the EU AI Act tiers (prohibited, GPAI/systemic, high-risk, limited-risk, minimal-risk), with rule-match transparency showing which regulation clauses triggered each classification
- Evidence locker & QMS — immutable, versioned storage for compliance evidence, auto-filling 9 of the 14 Annex IV elements from discovery data
- Documentation generator — one-click EU Declarations of Conformity, Annex IV technical documentation, and Instructions for Use
- Post-market monitoring — Article 73 incident tracking and automated reminders, with 10-year retention management
- Crypto-shredding — per-tenant envelope encryption; destroying the tenant key permanently erases all data, making GDPR Article 17 erasure a cryptographic guarantee rather than a promise
Why NOMYO Ships It
Our position on AI infrastructure has been consistent: compliance can’t be bolted on after the fact. We built encrypted AI inference because sending sensitive data to AI providers as plaintext is a breach in progress — data protection has to be a property of the system. The GUT Personalmanagement story showed what that looks like for HR data.
aissurance extends the same thesis from the data plane to the documentation plane. The AI Act doesn’t just ask you to protect data — it asks you to know and prove what your AI systems are, what risk they carry, and what changed about them. That’s a continuous operational duty, not a one-time filing. Companies don’t need another standalone GRC tool; they need compliance that runs where their AI actually runs.
And that’s the part we already have.
The Router Connection
NOMYO Router is our open source transparent proxy for Ollama and OpenAI-compatible inference endpoints. It already sits between your applications and your models — routing requests, tracking which model runs on which server, translating between API formats, and measuring traffic.
The aissurance router plugin is a shipped component of NOMYO Router that turns the deployment into a compliance agent. It is inactive by default and stays inert until both gates are in place: a plugin configuration block in your router and an activated aissurance account. The router alone never transmits anything, and neither does the account alone. With both gates in place, the router periodically reports signed payloads (HMAC-SHA256 over TLS 1.3) to aissurance:
- Model inventory — every model with version, quantization, size, request counts, and latency, per endpoint
- Endpoint registry — all inference backends with health status and connection counts
- Traffic telemetry — request volumes, cache hit rates, error rates
- Conversation affinity — the topic clusters being routed (customer service, code generation, data analysis), which aissurance uses to infer use cases for risk classification
Zero engineering effort, no forms to fill out, no inventory spreadsheet that goes stale the week after someone updates it. The system that actually serves your models documents what it serves — which is why discovery data can auto-fill 9 of the 14 Annex IV elements instead of a human transcribing them. With more than 450 GPUs and compute nodes under router management, this is the same pattern we proved with the router itself: infrastructure should be transparent to applications and honest about itself.
Why Upgrades Are the Blind Spot
Discovery covers what you run. The harder question is what changed. Model behaviour is not frozen at deployment — a serving-stack upgrade, a quantization change, or a runtime bump can silently alter outputs while the model name stays the same. Our model change control analysis on aissurance shows how bad this gets: the same endpoint was bit-exact 45/45 under sequential dispatch and produced nine distinct outputs for the same request under concurrent serving — with nothing in the logs flagging the difference. A single sampler parameter left open silently took the server’s default and shifted the measured tail slope from 1.1 to 0.73. Nothing in the output looked wrong.
This matters for anyone operating agents on self-hosted or OpenAI-compatible infrastructure. Your Annex IV documentation describes the system as it was when you wrote it. Your escalation gates — the thresholds that route uncertain cases to humans — were calibrated against the old behaviour. An upgrade can break both quietly.
We have direct production evidence of why this matters. Our DERAG commission reconciliation runs on measured uncertainty: the hybrid-groundworks routing framework measures semantic entropy while the LLM reasons, and the entropy signal decides who decides — machine, flag, or human. If a serving-stack upgrade quietly shifts the entropy distribution, that escalation gate stops doing its job and nobody notices, because the outputs still look plausible. That failure mode has a name and a test: entropy-as-wrong-signal, measured as AUROC.
That’s what the aissurance model change control kit does. It works like CI for model upgrades:
- Record a baseline before the upgrade — a fixed calibration set against your endpoint under a fully pinned sampler signature, with no defaults left open
- Upgrade — swap the model, bump the server, change quantization
- Re-run and compare — every metric checked against its gate; a failed gate fails the run with a non-zero exit code, so it drops straight into your existing CI pipeline
- File the report as evidence — under Annex IV element 12, change control documentation, with the report’s SHA-256 digest binding the summary to the archived JSON
It measures what an upgrade actually moves — determinism, accuracy, surprisal and tail shape, AUROC of entropy-based escalation gates, escalation rate — and the gates are deliberately one-sided: an accuracy improvement never fails a run, a degradation does. The instruments validate themselves against synthetic data before they’re trusted on yours.
What This Means for Your Agents
For teams running agents on the router, the workflow is short: run selftest once, record a baseline before your next model or stack upgrade, and wire the compare step into your deployment pipeline. From then on, every upgrade produces its own change-control documentation as a by-product.
And this is where the pieces connect: the next step of the router plugin is to run the change control kit on a schedule and file the reports without an operator in the loop — turning point-in-time comparisons into continuous post-market monitoring, which is what Article 72 was asking for all along.
Compliance as a By-Product
The throughline across everything NOMYO builds is the same: the infrastructure that serves your AI should also protect it, observe it, and account for it. The same pipe that routes inference reports the inventory, measures the change, and files the evidence.
That’s the difference between doing compliance as a paperwork project and running it as a property of the system. aissurance is how we ship that to everyone who doesn’t want to build it themselves.
Learn more at aissurance.eu, or start with the free compliance quiz. Running NOMYO Router already? The plugin docs get you connected in minutes.