Migrate AI Pilots to a Shared Bank AI Platform: A 2026 Implementation Guide
Migrate in phases ordered by materiality. Inventory every pilot, triage by consequence, autonomy and human oversight, translate each pilot's controls into one policy set enforced at runtime, and cut over only when the workload carries its permissions and evidence with it. Treat runtime enforcement and evidence as entry requirements for the platform, not as work to add after cutover.
Sequence the cutover in phases
- 1
1. Register and classify
Complete the inventory and assign each pilot a risk tier based on consequence, autonomy and human oversight.
- 2
2. Set platform entry requirements
Publish the control requirements every workload must meet: owner, scoped permissions, runtime policy, evidence capture.
- 3
3. Migrate a low-risk wave
Move supervised, low-consequence workloads. Run the new policy set in parallel with pilot controls and compare outcomes.
- 4
4. Reassess and gate
Reassess each use case at cutover and after material changes. Review evidence output with model risk management before advancing.
- 5
5. Migrate higher-risk workloads
Move agentic and consequential workloads with stricter enforcement, human approval and deeper evidence, then retire pilot environments.
Normalize pilot-era controls into one policy set
Data access
Pilots used broad service accounts or copied datasets. Common policy: per-workload data scopes, with redaction rules for sensitive fields.
Tool permissions
Pilots wired tools directly with shared credentials. Common policy: each workload receives an explicit allow list of tools and actions, re-approved at cutover.
Human approval
Approval was informal or absent. Common policy: consequential actions escalate for human approval, with thresholds set by risk tier.
Retention
Pilots kept prompts and outputs inconsistently. Common policy: one retention schedule covering prompts, outputs, embeddings, logs and feedback.
Usage limits
Pilots had no consistent ceilings. Common policy: rate limits per workload and per tool.
The supervisory baseline has moved
Many migration plans still cite SR 11-7 and OCC Bulletin 2011-12 as the model risk baseline. That is no longer current. On 17 April 2026 the OCC, Federal Reserve Board and FDIC issued revised interagency model risk management guidance (OCC Bulletin 2026-13). It clarifies principles, sets a risk-based approach, and rescinds prior issuances, including OCC Bulletin 2011-12 and OCC Bulletin 2021-19. It describes practices that are tailored and commensurate with size, complexity and extent of model use, and it is expected to be most relevant to banking organizations with over $30 billion in total assets. Confirm with your model risk management function how it applies to AI workloads at your institution.
On 4 November 2026 the OCC also issued Bulletin 2026-54, a supervisory framework for examiners evaluating bank use of super intelligence (SI), including agentic SI. It is examiner guidance, not a rule, as of November 2026. Several of its themes bear directly on a pilot consolidation: use-case inventory, materiality, human oversight, and aggregate exposure from shared providers, data, permissions and infrastructure. A shared platform concentrates that exposure, which is why the platform's controls and evidence become the subject of review. Accountability for managing risk generally remains with the bank even when vendors supply the platform or models.
Inventory and triage the pilots
Start with a register, not a platform design. For each pilot, record a named owner, the use case, the models and prompts in use, the data it reads and writes, the tools and connections it can call, who is authorized to use it, its degree of autonomy, and the human oversight in place. Diffuse ownership is a recognized governance risk, so a pilot without an accountable owner should not migrate until it has one.
Then triage. The OCC framework says risk classification can rest on potential consequences, transfer of agency and degree of human oversight, rather than technical sophistication. A simple retrieval assistant for internal policy documents and an agent that can initiate payment actions may use the same model, yet they belong in different phases. The readiness criteria below are practitioner judgment, not regulatory requirements.
- Named accountable owner and registered use case
- Documented data inputs, including data created through prompts, outputs, embeddings and logs
- Enumerated tool and connection permissions
- Defined human approval points for consequential actions
- Retention and logging behavior known and reproducible
- Reassessment completed for the move to the shared platform
Keep evidence continuous through cutover
Examiners will ask what each migrated workload was allowed to do and what it did. Answering requires evidence generated at the point of action, not reconstructed from application logs afterward. Per workload, retain the registered inventory record, the policy version in force, and the permissions granted, with a history of changes. Per interaction and tool call, retain the policy decision (allow, redact, block, rate limit, escalate), the identity and workload involved, and any human approval and its outcome. These artifacts are a practitioner design, and their depth should scale with use-case risk, since examiners are told to tailor reviews to materiality. Map them to the NIST AI RMF functions and to your model risk management documentation so reviewers can follow one trail.
The cutover itself is the moment most likely to break continuity. Preserve pilot-era records, tie them to the same workload identifier as the platform records, and avoid a gap in which a workload runs without enforcement or evidence.
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate-for-human-approval decisions on individual interactions and tool calls. That enforcement generates audit-ready evidence mapped to frameworks including NIST. The same approach has run in production for more than 18 months at a regulated enterprise, and Trussed has 100+ models under one policy plane. Insurance, a regulated financial sector, is where Trussed is deepest. Trussed does not trace end-to-end agentic workflows or provide model lineage or dataset provenance, so those needs stay with your existing model risk tooling.
Pilot to shared platform: migration path
Label: Inventory
Description: Owner, use case, data, tools, autonomy, human oversight
Label: Triage
Description: Rank by consequence and autonomy, not model sophistication
Label: Normalize
Description: One policy set replacing per-pilot controls
Label: Cut over
Description: Re-scoped permissions, runtime enforcement, evidence from day one
Plan your pilot migration around enforcement and evidence
Review your pilot inventory, risk tiers and target policy set with a Trussed expert.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.