KYC and customer onboarding agent. A production blueprint.
An 8-week KYC and customer onboarding agent on Google's Gemini Enterprise Agent Platform. ADK, A2A v1.2, Memory Bank, Agent Registry, Model Armor.
The outcome.
- What it does
Onboards a customer in minutes, not days.
Document intake (passport, proof of address, beneficial-owner declarations), sanctions and PEP screening across the providers your compliance team already trusts, and a risk score the analyst can defend. Borderline cases route to a human reviewer with the rationale already written.
- Who it serves
Compliance, onboarding, and the customer.
The CCO gets a defensible audit trail per applicant. The Head of Onboarding gets a queue measured in hours. The customer gets a single intake flow and a decision before the kickoff call ends.
- Business KPI
Time-to-decision and analyst capacity.
Median time-to-decision drops from 3.5 business days to under 4 hours on the clean cases. Analyst capacity reallocates from triage to the borderline tier where their judgement actually matters.
Components used.
Five platform components. Each one earns its place. Every version pinned at delivery.
ADK Python v1.32
Agent Development Kit. The agent is shipped as ADK code your team owns and deploys.
Memory Bank
Long-running context across sessions and between agent steps.
Agent Registry
Versioned registry entry with policy, owner, and rollback target. Written before cutover.
Model Armor
Prompt and response policy at the gateway. PII redaction and jailbreak filters enforced inline.
Eval harness
Golden-set evals wired to your CI on day one. Every PR ships with the regression delta.
IAM & Workload Identity
Per-agent service accounts. Scoped to the providers and stores the blueprint actually touches.
Cloud Trace + Logging
Per-step traces piped to the stack you already operate. The on-call rotation does not change.
Agent Runtime
Autoscaled, regionally pinned. First-class A2A and tool invocation.
Persona. Chief Compliance Officer, Head of Onboarding.
The CCO cares about the audit trail before they care about the speed. Every decision needs a defensible rationale, every vendor lookup needs a logged response, every override needs a named human. The blueprint produces all three by default, not as a retrofit.
The Head of Onboarding cares about queue depth and time-to-decision. The blueprint reduces the manual triage tier by routing the clean cases to auto-approval and the borderline cases to a human reviewer with the rationale pre-drafted. Reviewers spend their time on the cases that actually require judgement.
Both roles get the same registry entry, the same eval coverage, and the same rollback target. There is one source of truth per agent, and it is versioned.
Timeline. 8 weeks.
The 6-week pilot methodology, extended by two weeks to cover the additional vendor integrations and the regulator-grade audit binder.
Book the call. We bring the named SE.
Two business days to a named SE on the call. The first meeting is a working session, not a deck.