By industryFinancial services: one control layer for BCBS 239, FRTB and DORABuilt for BCBS 239, FRTB and DORASee how
Solutions / AI Governance

Your AI does the work.
Your signature carries the liability.

Supervisors treat that signature as evidence of effective challenge. It only holds if you can reconstruct what the AI read.

EU AI Act Art. 14GDPRSOC 2ISO/IEC 42001
Book a 30-minute briefing Read the CRO brief →
88%
of organizations use AI in at least one function
McKinsey, n=1,993, July 2025
20%
have a mature governance model for AI agents
Deloitte, n=3,235, September 2025
€15M
or 3% of global turnover
Art. 99(4) — the tier covering deployer and oversight duties
2 Dec 2027
High-risk obligations apply
Deferred from Aug 2026 by the Digital Omnibus

A signature is not the control. Oversight is.

The overseer must be able to understand the system's capacities and limitations, remain aware of automation bias, correctly interpret its output, and override, disregard or stop it.
EU AI Act, Article 14(4) — paraphrased

That is a capability standard, not an approval standard. Nobody can override output they cannot trace back to its sources.

What a supervisable run looks like.

Every link in this chain exists today. The audit event is the artefact, not the agent.

Lakekeeper — available todayAgent runModel · task · promptScoped credentialShort-lived, per requestCatalog decidesDeny by defaultAudit eventIdentity · policy · timeAppend-only. Emitted as the data is used — not reconstructed afterwards.

What changes with Lakekeeper.

Deployable today — access control and a durably-recorded, append-only audit trail.

Today
With Lakekeeper
Access
Agents inherit a service account with broad standing access to whatever it can reach.
Supports Art. 26 deployer duties
Deny by default at the catalog — table and view level for tables, dataset level for files in object storage.
Attribution
A log that shows a query ran, not who authorised it or under which policy.
Supports Art. 12 record-keeping
Every read and write is an append-only event, tied to the identity that made it and the policy applied.
Credentials
A static key in the agent config, valid until someone remembers to rotate it.
Supports Art. 26 · SOC 2 access control
Short-lived scoped credentials, vended per request through pylakekeeper. No standing access.
Scope
Governance stops at the warehouse; the documents agents actually read sit outside it.
Supports Evidence across both
One authority over structured tables and files registered as generic datasets alike.

Building now, with design partners

In development
AI Contract object in the catalog

A first-class catalog object declaring which agent, which model, which task, which sources and which outputs — adding agent identity, run ID, model version and prompt/output hashes to every audit event.

Lakekeeper does not just manage data.
It manages liability.

Book a 30-minute briefing

Bring whoever owns AI risk and whoever owns data governance.

Sources: McKinsey, The State of AI (Nov 2025); Deloitte, State of AI in the Enterprise (2026); Regulation (EU) 2024/1689 (AI Act) Arts. 14 and 99, as amended by Regulation (EU) 2026/1744.

← Back to all solutions