Platform
Capabilities a regulated buyer can evidence.
Every capability below is framed around something your risk committee, auditor or regulator will ask you to prove. K-Lake runs inside your infrastructure, so the evidence stays there too.
Classification
Know what is sensitive before an AI sees it
K-Lake classifies every file it indexes as public, internal, confidential or restricted, and shows the evidence behind each label.
Rules then decide what search and AI may return at each level, enforced in the data layer rather than in the assistant, so a new assistant or a new interface inherits the same control.
A data protection officer can show which restricted content an assistant is able to reach, and evidence that the answer is none.
Discover
Know your exposure before an AI sees it
Most AI programmes in regulated firms stall at the same question from the risk committee: what exactly will the assistant be able to read?
Discover answers it from the index K-Lake has already built, with no second crawl. It breaks the estate down by file type, age, owner, language, size, storage tier and source, and shows how much of it is searchable.
Exposure signals highlight world-readable files, files with no owner, duplicates and anything stale beyond three years. Every figure links through to the files behind it, so a percentage in a board paper can be traced to named documents. Where coverage is incomplete, Discover says so.
Smart Actions
Control what an AI can see
Decide what reaches an assistant, file by file, with steps that run after every crawl in an order you set per source.
Redaction masks card numbers, email addresses, phone numbers, IP addresses, identifiers and your own patterns in what search and AI receive, with a tester to check a rule before it goes live.
A signed webhook tells your own systems when a file has been processed, so your SIEM or case management tool stays in step. Files that fail extraction are retried on a heavier engine. Your original files are never changed, and when you change a rule only that step re-runs.
Governed AI access
Every assistant signs in as the person asking
K-Lake connects to Copilot, Claude, ChatGPT, Gemini and your own agents through the open Model Context Protocol.
Its built-in OAuth 2.1 authorisation server federates sign-in to Entra ID, Okta, Google or Keycloak, so an assistant acts as the individual user and inherits exactly their file permissions. Assistants receive extracted text only, never the original files.
The connection is switched off until you enable it. Once on, the console shows who is asking, from which assistant and against which sources. Administrators can revoke a session, or hide a sensitive source from AI while keeping it searchable for staff. Every request is recorded, including the ones K-Lake refused.
Knowledge graph
Relationships you can evidence
Find the people, organisations, locations and agreements your documents mention, and how they connect.
K-Lake records the relationships the text actually asserts and keeps them separate from names that merely appear in the same document. That distinction matters when the output feeds an investigation or a diligence report.
Ask which files place three named parties together, or start from one document and see everything connected to it. Duplicate spellings of a name can be merged and unmerged. Every result is trimmed to the documents the reader is allowed to open.
Access diagnostic
Evidence for your auditor
Access in K-Lake is enforced in the database and fails closed, so a user sees nothing without an explicit grant.
When a user expects a file and cannot find it, an administrator can run a check that shows which permission, identity mapping or source setting is responsible. The reasoning that keeps a file hidden is visible to the people who have to defend the control.
Together with the record of every AI request and every file opened, it gives a security reviewer evidence to examine.
Next step
See it on your own data
The fastest way to judge any of this is a proof of concept on your infrastructure, using your documents. Talk to us about one, or read the trust centre first.