Control Plane for a Bank's Agent Estate
Fourteen independently deployed models and agents were brought under one governed control plane in 22 weeks — turning EU AI Act readiness from a remediation project into a byproduct of normal operations.
The challenge
Over three years, different teams at the bank had shipped 14 separate ML models and LLM-based agents — fraud scoring, an SME underwriting copilot, a retail chatbot, collections prioritization, and ten others — each on its own stack, with its own logging, and no shared registry. With the EU AI Act's compliance deadlines approaching, that meant:
- No single inventory of which models existed, who owned them, or what risk tier they fell under
- Human-oversight logging implemented differently — or not at all — across the 14 deployments
- Audit evidence assembled by hand from whichever logs a given team happened to keep
- No consistent process for classifying a new agent's risk tier before it shipped
- Compliance and engineering working from different, frequently outdated documentation for the same system
A pre-project inventory took six weeks just to confirm all 14 deployments actually existed and were still in production.
How it works
One registry, one policy engine, fourteen deployments behind it
The goal wasn't to rebuild the 14 systems — it was to put a single governed layer in front of all of them:
- 01
A central model and agent registry built to capture ownership, data lineage, and intended use for every deployment
- 02
An automated risk-tier classifier mapping each system against EU AI Act Annex III categories
- 03
A policy engine enforcing human-oversight and logging requirements consistently, regardless of which team built the underlying model
- 04
Observability agents deployed alongside each of the 14 systems to stream decisions and overrides into a shared audit log
- 05
Existing deployments migrated onto the control plane in weekly batches, starting with the highest-risk-tier systems
- 06
A compliance dashboard generating audit evidence packages on demand instead of by request
- 07
A gating step added to the deployment pipeline so no new model or agent ships outside the registry going forward
What we built
Key capabilities
Single source of truth
Every model and agent in production is registered in one place, with ownership, risk tier, and lineage documented, not scattered across team wikis.
Consistent oversight, regardless of stack
Human-oversight and logging requirements are enforced by the policy engine, not left to each team's own implementation.
Audit evidence on demand
Compliance evidence packages generate automatically from the shared audit log instead of being assembled by hand per request.
Nothing ships ungoverned
New models and agents are gated at deployment — the registry is a checkpoint, not an afterthought.
Before vs after
What changed across the estate
- Governed deployments
- 0 of 14 → 14 of 14
- Audit evidence turnaround
- ~3 days → ~9 minutes
- Risk-tier classification
- Ad hoc → automated at registration
- New deployment gating
- None → mandatory registry check
Business impact
What it changed
22-week path to full readiness
All 14 deployments were inventoried, risk-classified, and migrated onto the governed control plane within 22 weeks — ahead of the bank's internal EU AI Act readiness deadline.
Audit turnaround from days to minutes
Evidence packages that took roughly three days to assemble by hand now generate in about nine minutes from the shared audit log.
No remediation backlog
Because every system was brought onto the same policy engine rather than patched individually, the bank entered its first post-deployment exam with zero open remediation items.
Technology stack
“Governance that lives in 14 different places isn't governance — it's 14 separate bets that nothing goes wrong. One control plane turned that into one bet, managed on purpose.”
Keep reading