The agent
An AI agent does the work. A person decides.
A payroll agent prepares a contractor payment. Each step is checked against its agreed form and signed as it happens, and nothing is paid until a person approves and signs it.
EU AI Act · EUSAiR
Europe’s AI Act makes people answer for what their AI systems do. Enforcing it, and managing AI risk, rests on evidence anyone can check. Aqua Protocol builds that evidence into every step an AI agent takes.
EUSAiR final event · Bologna · 30 October 2026
One AI agent, one payment, three views of the same evidence.
The agent
A payroll agent prepares a contractor payment. Each step is checked against its agreed form and signed as it happens, and nothing is paid until a person approves and signs it.
The audit portal
The audit portal lists each agent session with its approvals and runs the integrity checks on its evidence.

The viewer
The viewer opens the evidence package and shows every step, who signed it and how the steps connect. It checks the signatures itself instead of trusting the sender.
Recorded on our test system with sample payroll data.
An AI agent prepares a payment, a person approves it, a system pays. Later an auditor or an authority asks what happened. The answer comes from logs the company running the agent keeps for itself.
Promises don’t travel between organisations. Proof does.
Aqua Protocol attaches signed evidence to every step an agent takes and every handover between systems. All steps together form one evidence package that travels with the data.
Darlene is a contractor. Her client asks an AI agent to pay her for 40 hours. Here is what Aqua records at each step, and what happens when something goes wrong.
Authorities can only enforce the AI Act on records they can trust, and companies can only manage AI risk on rules they can show were kept. Both need evidence that holds up outside the system that produced it. That is the layer Aqua Protocol provides.
For stand-alone high-risk AI systems, these duties apply from 2 December 2027.
What stays with you: which events matter, the oversight procedure, training, and the means to stop the system. Aqua supplies the evidence; your systems act on it.
Our reading of the regulation, not legal advice. What we report, and what we don’t.
In 2026 Aqua took part in EUSAiR, an EU-funded project that runs simulated regulatory sandboxes for the AI Act. With its researchers we worked through our payroll agent as if it were a high-risk system, starting with record-keeping and human oversight.
It showed us where signed evidence answers those duties directly, and where the system around Aqua has to do more. What we learned shapes this page.
Confidential computing proves which code ran on which hardware, but not what the AI did. Audit logs record what the AI did, but can’t prove where it ran or that nobody changed them afterwards. Our proof of concept with enclaive brings both into one evidence package.
Start small: one high-risk workflow, one evidence package your team, partners and auditors can check themselves.