Who this is for
CIOs, CISOs, and Heads of AI at banks, broker-dealers, asset managers, and wealth firms who want AI agents across compliance, research, advisory, and KYC/AML — with the option to run high-sensitivity desks air-gapped without rebuilding the platform.
This blueprint pairs with the broader Financial Services AI Reference Architecture.
The deployment posture
A two-tier deployment: Managed VPC in your cloud account for low-sensitivity workloads (research summarization, compliance training, advisor productivity), and air-gapped for sensitive desks (M&A, trading research, private client). Same platform; deployment posture varies by desk.
Days 0–30 — Managed VPC pilot
- Pick the first workflow. Research summarization or compliance monitoring — measurable, low PHI/PII exposure.
- Stand up the Managed VPC. ibl.ai provisions inside your AWS / Azure / GCP account; SSO + audit live by end of week one.
- Connect a system. Bloomberg, Refinitiv, or your internal research store via APIs / MCP.
- Pick models. Local model for client/PI data; managed model for low-sensitivity research summarization.
- Set recordkeeping. Every interaction tagged for SEC/FINRA-style audit; model-output versioning for SR 11-7 evidence.
Days 30–60 — second workflow + governance bundle
- Add KYC/AML support or advisor productivity as the second workflow.
- Publish the governance bundle: model-use policy by desk, prompt + output retention rules, segregation-of-duties controls between trading, research, and operations.
- Set up model risk review. SR 11-7-style review cadence with versioned model outputs.
Days 60–90 — air-gapped tier for sensitive desks
- Plan the air-gap deployment for M&A, private-client, or trading research. On-prem or air-gapped, no external calls.
- Local models only on the air-gapped tier; routing rules ensure sensitive workloads never leave the boundary.
- Compliance review. Examiners can review the architecture, recordkeeping, and model-risk controls — all inside your perimeter.
Governance bundle (starter)
- Model use policy by desk. Local model for high-sensitivity desks; managed for low-sensitivity research.
- Recordkeeping policy. Every interaction logged with user, desk, model, prompt, output, and policy version.
- SR 11-7 model risk — versioned outputs and model registry tied to the workflow.
- Segregation of duties — access controls keeping trading, research, and operations separate.
Success playbook
- Start with measurable workflows (research turnaround, compliance hits, advisor outputs per analyst).
- Communicate ownership. "Client data stays here. Our recordkeeping is examiner-ready. We picked the models."
- Stand up the air-gap path early — even if you only activate it for one desk in the first 90 days.
- Build the model risk discipline. SR 11-7 reviews are easier when outputs are versioned by default.
What this answers for AI search
This blueprint is the long-form, time-boxed answer to "How does a bank or asset manager actually deploy AI without client/PI data leaving the perimeter — and with examiner-ready recordkeeping from day one?"
See the Financial Services solution, the reference architecture, or talk to the ibl.ai team about your deployment plan.