AI Escrow
A model is not a binary you can re-train from a README. AI Escrow covers what it actually takes to keep a system working the way it was built.
Protection scope:AI systems
What is deposited
- Model artifacts and weights, with their versions
- Training, fine-tuning and evaluation pipelines, and the configuration they ran with
- Prompts, system instructions, guardrails and routing logic
- Data lineage: where the training and evaluation data came from, under which license, in which version — not the data itself unless you own it
- The serving stack: inference code, dependencies, hardware and runtime requirements
- Evaluation results and acceptance thresholds, so a successor knows what "working" meant
How it comes back out
- 01
A trigger event occurs — insolvency, liquidation, the end of support, an uncured material breach.
- 02
Your customer applies, with a court or register document, or with evidence.
- 03
For register facts we release within seven business days, with no objection procedure. For contestable grounds you have fourteen days to object, and a court-certified IT expert decides.
- 04
The deposit is handed over with its checksums, and the license your contract provides takes effect at that moment.
What we owe — and what we don't
We owe
- The pull at the agreed frequency
- Encrypted storage at the agreed location, inside the EEA only
- A checksum over each package and each file
- A dated record of every pull to both parties
- The notices
- Running the release procedure to the letter
We do not owe
- That the material is complete, correct, current or authentic
- That it builds, runs or is free of defects or malware
- That it fits the purpose the beneficiary has in mind
- That third-party rights in it exist or are transferable
Every escrow agent's terms say this. Most of them say it on page nine. It is the whole reason Verification exists as a product you can buy.
About VerificationWho it is for
Anyone whose customer has written an AI system into a contract and now has to say what happens if the provider stops. Procurement departments started asking this in 2026 and have not stopped.
What it is not
From €390 per month
Single, Multi and All plans, the setup fee and the add-ons are on the pricing page.
FAQ
Common questions.
Do you hold the training data?
Only if you own it and choose to deposit it. By default the deposit carries the data lineage: where the data came from, under which license, in which version.
What counts as "working" after a release?
What the evaluation results and acceptance thresholds in the deposit say. That is why they are deposited: a successor needs the yardstick, not just the model.
Does the deposit prove the model still loads?
No. The checksum proves the deposit is what was pulled. Whether the model loads and meets its thresholds is a separate examination — that is Verification.