Data Safe AI.
We built the models before we built the platform. Our own models, trained on our own hardware for collective labour relations, served from inside the EEA, and never trained on your data.
Models first. Then the platform.
Most software in this market started as a system of record and added an AI feature later, usually by sending your documents to someone else’s model. Graylark took the other road.
Graylark began as research. The first problem was reading collective labour relations material the way a practitioner does: what an EWC agreement actually commits an employer to, which clauses matter when a restructuring lands, and how the answer changes from one country to the next. The models came out of that work, processing EWC agreements and consultation records long before there was a product to put them in.
The platform was then built around those models. Landscape, Change Proposals and the EWC Portal are the places where the models do their work, on records the user is already authorised to see. That order matters. It is why the AI in Graylark knows the difference between the Betriebsverfassungsgesetz and the EWC Recast, and why we can say exactly where your data goes, because we built every part of the path it travels.
We call the result Data Safe AI: models we build and run ourselves, so your data never leaves our umbrella and is never used to train anything.
What Data Safe AI means.
Our own models, insights connected to your records, and data that never leaves.
Our own models
Domain-specific models for labour relations, trained on Graylark’s own hardware. A larger teacher model, tuned by an automated experiment loop, distils into the model the platform serves. The training data is synthetic and encoded from legal frameworks, so when a framework changes the data is regenerated and the model retrained.
Insights connected to your records
Every insight is generated in context, scoped by country, from the records around it: the proposal, related agreements, legal advice your team has added and previous change proposals. The output is always a draft for a person to review, never a decision made for them.
Nothing leaves
Inference runs on Graylark-managed infrastructure inside the EEA. There are no calls to OpenAI, Anthropic, Google, AWS Bedrock, Azure or any other third-party hosted model in the customer data path. Customer data is never used to train Graylark’s models, and that is enforced in code as well as in policy.
What it does for you today.
GrAI is the suite of AI services inside the platform. Each one works on the documents and consultation records already in front of the user, and each returns a draft the user can edit.
Change Proposal Insights
A structured analysis of a proposed change: considerations, risks and regulatory context, drawn from the proposal, the country, related agreements, aggregated works council feedback and previous advisories.
Agreement Chat
Ask a specific agreement what it says. “What does Article 12 say about consultation timelines” is answered from the text of that collective or EWC agreement, in context.
Engagement preparation
A preparation assistant for an EWC meeting or consultation round, working from the agenda, the uploaded materials, the opinions and the feedback themes, so the open questions are summarised before anyone walks in.
Assisted authoring
Drafting help across the platform: a response to a works council question, a section of an agreement, the narrative of a report. Always a draft, always for a person to finish.
Translation in fourteen languages
Inline translation of a comment or a chat message, and full document translation of a DOCX, across the languages the platform supports.
Aggregated feedback summaries
A short summary of the feedback collected in a consultation round. Comments reach the model without submitter identifiers, and nothing is scored, classified or inferred about any individual.
How Graylark uses AI: inputs, outputs and controls for every feature
What it will never do.
The lines are drawn in the architecture, not in a policy document, so that a works council, a data protection officer or a regulator can check them.
No judgements about people
No scoring, ranking, profiling or attrition prediction of any employee. There is no code path that takes an employee record as input and produces a judgement about that person, and sentiment is anonymised before a model sees it.
No automated decisions
GrAI is decision support. Its outputs are drafts for human review and are not used by the platform to make decisions that produce legal or similarly significant effects on anyone.
No side doors
A user can only invoke an AI feature on records they were already authorised to see. Country and business-unit scoping carry through, and there is no separate AI permission model to misconfigure.
Nothing you cannot switch off
A master off-switch disables every AI surface for a tenant, with per-feature switches beneath it. Every AI call is recorded in an audit log, with the user, the feature, the model and the record it ran on.
Proven on every build.
Graylark Labs is our applied AI research arm, and every research track maps to something in the platform.
Safety, tested continuously
A benchmark aligned to the OWASP LLM Top 10 runs on every build. Prompt safety and response safety layers sit on every call, refusing instruction override attacks and filtering what comes back.
Evidence that writes itself
The evidence behind our security questionnaires is produced by the same engine, build after build, so what we tell your security team is what the last build measured.
Agents, inside the same walls
A production MCP server lets AI agents work with Graylark data only inside the same tenant boundaries and permissions as a person, through scoped, revocable tokens tied to an accountable identity, with every tool call recorded.
See it on your own material.
Bring an agreement or a live change. We will show you what the models make of it, and which of your records it drew on.