Your AI runs on records you are trusted with — claims, transcripts, clinical notes. Relevant builds protection that travels with them through prompts, retrieval, training, and logs, so every record your AI touches can stay classified and accounted for.
Every store, prompt path, and index feeding your models gets scanned, and each field comes back tagged by sensitivity so the controls downstream know what they are handling.
Personal identifiers leave your training sets, prompts, and support transcripts before a model reads them, and the measurement of what the model can still do afterwards ships with the rule.
Production values stay in the vault while your models, test environments, and analysts work with tokens that behave like the real thing.
Encryption reaches the places AI systems store data in practice: embeddings, response caches, prompt logs, and model artifacts.
A deletion request travels the whole path: source record, derived embeddings, cached responses, and the backups behind them.
Teams that need realistic material for model training, demos, and load tests receive generated records carrying the statistical shape of production.
Every prompt, retrieval call, and agent tool-call gets checked against policy inside your own request paths, tuned on your own traffic in shadow mode before anything starts being blocked.
Your prompts, indexes, training sets, and logs come back as one inventory, with an owner and a sensitivity rating per store. You approve the scope before any code moves.
Relevant runs the working session, drafts the list of data classes, and writes it into your repository as the single reference every control reads from. You choose which classes get protected first.
We write each control into your repository and deploys it behind a feature flag, so it proves itself on live traffic before it becomes the default.
A test erasure runs end to end across source, index, cache, and backup. The routine we install issues a dated proof for every request after it, mapped to the GDPR, HIPAA, ISO 27001, and EU AI Act articles in scope.
The handover is the running system: controls sit in your repository, keys in your custody, deletion and evidence routines running, with the runbooks and thresholds behind them.
Every rule ships as reviewed code with tests, so your team changes a threshold or adds a data class without opening a vendor ticket.
Each redaction rule ships with the answer-quality benchmark that justified it, so the privacy and accuracy trade-off arrives as a number your product owner can discuss.
Controls sit at your own boundary, outside any single vendor’s console, so a move from one model provider to another leaves your obligations satisfied.
Each control emits the log line an auditor asks for, timestamped per request and retained on the schedule your policy sets.
Up to 50% faster delivery and up to 20% fewer bugs through AI-assisted development, including the re-identification and utility test suites that ship with your controls, with every release still passing senior code review.
96% employee retention. The engineers who built your protection layer stay reachable as your data classes grow, and the runbooks leave with your team at the end.
Quite a lot, if the work starts in the right order.
We begin with something small: a request id that follows a prompt through retrieval and back out again. Without it you can describe what the model read, but you cannot show it. Reviewers can tell the difference.
Then classification, masking on your most sensitive data class, key custody, and the deletion paths. Each stage finishes something an auditor can read, so an early review still has a result in front of it.
First controls reach production inside the opening weeks, and the order is what makes evidence accumulate early. Classification lands first, then masking on the highest-sensitivity class, then encryption and key custody, then the deletion paths.
Each stage produces its own artifact, so a review three weeks in already has something to read. Relevant delivers 99% of milestones on time, and the dates sit in the roadmap you approve.
Yes, and the clause narrows the job. It covers training. Three things sit outside it, and each has a fix.
It comes from coreference. One placeholder for every name turns two people into one, and the model loses track of who did what. Pseudonyms that stay consistent inside a document keep them apart, at no extra cost.
Summaries, classification, and routing survive redaction well. Matching a person across two systems depends on the identifiers, so that workflow gets tokenization with role-based unmasking.
Four things, all in places your licence cannot reach.
Tell us about your AI systems and your compliance scope, and our engineers will outline the controls, the evidence each one produces, and the team to build it.
AI
AI
Tech label
