Model risk becomes a compliance concern when a bank cannot establish which models influence decisions, who authorised their use or whether known weaknesses were addressed. A technically impressive validation report is only one part of the answer. The operating framework must control how model outputs reach lending, pricing, financial reporting and other decisions.

Scope matters. The PRA's April 2026 SS1/23 applies to UK-incorporated banks, building societies and PRA-designated investment firms with internal-model permission for regulatory capital. It excludes firms without that permission and third-country branches from these expectations, while noting that they may find the principles useful. Credit unions, insurers and reinsurers are outside its scope. Within in-scope firms, the framework extends beyond regulatory-capital models, including relevant vendor and financial-reporting models. The PRA publication page identifies the current version.

Start with decisions, then find models

A useful discovery exercise asks each business area which estimates, forecasts or scores materially influence its decisions. Follow the answers into production systems, vendor services and end-user tools. An inventory based solely on the model-development department can miss models purchased or embedded elsewhere.

Record the intended use and compare it with actual use. A tool developed to rank cases may later be used to close them automatically. That change can alter the risk even if the underlying mathematical method remains unchanged. Capture downstream dependencies so that a weakness in one model can be traced into the decisions and reports it influences.

Make model status meaningful

I recommend statuses that express permission, not just project progress. For example, distinguish development, approved use, conditional use, restricted use and retirement. Define who can grant each status and what production controls enforce it. A restricted model that remains available without warning has only a documentary restriction.

For every conditional approval, specify the permitted population, limitations, compensating controls, owner and expiry. Avoid an indefinite exception renewed without new evidence. Senior management should see the exposure dependent on overdue conditions, rather than only the number of open actions.

Challenge the full chain of use

Independent validation should address whether the method is appropriate, whether implementation is faithful and whether performance remains suitable for the intended decision. Compliance can test whether governance makes that challenge effective without attempting to replace specialist validation.

Ask whether validators receive the necessary data and documentation, whether unresolved findings reach the approval authority and whether commercial urgency can override restrictions without a clear record. Examine incentives and reporting arrangements. Independence is weakened when the person assessing a model also depends on its timely launch for their performance objectives.

Examine a conditional-use example

In a hypothetical bank, a credit model performs adequately for established customers but has limited evidence for a newly targeted segment. Management proposes to extend use immediately. A defensible decision would identify the uncertainty, assess materiality and consider a bounded pilot, additional review or a different decision method.

The approval record should explain the chosen safeguards and the evidence required before expansion. Monitoring should separate the new segment from the established portfolio. A stable aggregate performance measure could otherwise conceal deterioration where the model has the least support.

Connect remediation to decisions

Prioritise findings by their effect on use, not simply their age. A long-standing low-impact documentation weakness and a newly discovered material implementation error require different responses. Record whether the finding changes approval, pricing, limits, financial estimates or customer treatment.

When remediation is complete, verify the repair in the deployed version and check dependent systems. Recalibration in a development notebook does not prove that production consumes the corrected parameter set. Retain version identifiers, test results and approval evidence so the bank can reconstruct the relevant decision.

Report aggregate dependence

A board view should show where several models share data, assumptions, vendors or staff. Failures can be correlated even when each model has a separate owner. Pair the inventory with exposure measures, material limitations and the decisions required to reduce them.

The practical compliance question is whether the bank can demonstrate control over model use throughout its lifecycle. A coherent answer connects the model, its approved purpose, independent challenge, production evidence and an accountable response when conditions change.