Key Takeaways

  • Kihan: A machine-readable AI inventory (a registry that software can query) should cover model owners, data lineage, validation status, risk tier, and deployment endpoint.
  • Decision-authorization controls should apply policy checks at REST API or workflow boundaries before credit, fraud, anti-money laundering, or customer-verification actions reach production.
  • Buyers should measure approval latency, unresolved policy exceptions, model-drift alerts, and the percentage of AI systems supported by current validation evidence.

Organizations can structure AI governance by converting institutional policies into runtime authorization checks, maintaining a queryable AI inventory, and recording evidence for each decision. AI governance comprises the roles, rules, controls, and records used to oversee AI systems.

Problem to Solve: Governance That Operates at Production Speed

A fraud model flags a payment, a generative AI assistant summarizes the case, and an analyst approves an account restriction. Each component may work as designed, yet the bank still needs to establish who authorized the final action, which policy applied, what data informed it, and whether a human could intervene.

That control problem is becoming more visible as deployment expands. EY reported that 52% of surveyed banks identified governance as their top AI-adoption challenge, while only 44% reported investing in technology that supports ethical AI adoption.

The difficulty is not a lack of policy documents. It is translating them into controls that operate across Python models, third-party application programming interfaces, cloud data platforms, case-management systems, and core banking workflows. A PDF policy cannot block an unapproved model version or verify that a decision stayed within an employee’s delegated authority.

Buyers should begin with a system inventory. Each entry should identify the model version, business owner, technical owner, input datasets, PostgreSQL or object-storage location, API endpoint, risk classification, validation date, and downstream action. Data lineage, the documented path from source data through transformation and use, should also be queryable; otherwise, audit preparation will remain largely manual.

Evaluation Approach: From Written Principles to Executable Rules

Kihan frames its approach around Constitutional Computing, AI Governance, and Decision Authorization. In this context, Constitutional Computing means encoding institutional principles and restrictions as system-enforceable rules. For buyers, the relevant question is whether these concepts can produce machine-enforceable conditions rather than another review dashboard.

A constitutional layer should express constraints such as permitted data use, required human review, prohibited actions, and escalation conditions. A lending workflow, for example, might permit a model to recommend a credit limit but prevent automatic denial when required adverse-action explanations are missing. Policy output could be represented as signed JavaScript Object Notation (JSON), a structured data format, containing the rule invoked, authorization result, model identifier, and timestamp.

Decision authorization adds a runtime control point that determines whether a particular actor or system may execute a proposed action. Before an AI-generated recommendation reaches a loan-origination system or anti-money laundering (AML) case platform, a Representational State Transfer (REST) API or message-bus consumer can check the actor’s role through OpenID Connect, an identity-authentication protocol. It can then compare the action with delegated authority and write the result to a tamper-resistant audit log.

Buyers should test that mechanism with rejected as well as approved requests. Useful evaluation scenarios include stale model validation, missing consent metadata, a user exceeding approval limits, and an unavailable policy service. Fail-closed behavior, which blocks an action when the control cannot return a valid result, may suit account closures. Lower-risk document summarization may instead use a queued review process.

Implementation Considerations for Financial Institutions

Implementation should progress through discovery, controlled deployment, and broader operational adoption. During discovery, risk, compliance, data science, information security, and internal audit teams should map existing controls to specific system events. The output should be a control matrix, not merely a list of principles.

During controlled deployment, a team can place authorization checks around one bounded workflow, such as an AML alert-prioritization model. Integration points may include a Kafka topic carrying model decisions, a REST interface to the case-management platform, and a security information and event management (SIEM) system receiving policy violations in a structured format.

Midway through implementation, identity and data lineage often become the harder problems. Shared service accounts obscure who initiated an action, while copied datasets can break traceability among a Snowflake table, feature store, and production model. Service principals, workload identity, dataset versioning, and signed model artifacts can reduce that ambiguity.

The governance structure also needs to distinguish model risk from use-case risk. A summarization model used for internal notes does not carry the same exposure as the same model drafting customer-facing adverse-action explanations. A Deloitte survey found that AI use in fraud detection, including AML and know your customer (KYC) processes, had reached 58% adoption among surveyed EMEA banks, while more than half identified transparency and explainability as hurdles (centralbank.go.ke). The Deloitte figure covers an EMEA banking sample, whereas EY’s survey has a different geographic and respondent scope, so the percentages should not be treated as directly comparable measures.

Outcomes Buyers Should Measure

The first useful outcome is coverage: what percentage of production AI systems has a named owner, current validation record, documented data lineage, and assigned risk tier? A registry total alone is misleading if shadow models remain embedded in spreadsheets, notebooks, robotic process automation scripts, or vendor applications.

Operational measures should include median authorization latency, exceptions awaiting review, policy-service availability, model-drift alerts, overridden recommendations, and time from incident detection to escalation. Model drift is a measurable change in input data, relationships, or performance after deployment. Teams can collect these measures through OpenTelemetry traces, SIEM events, API gateway logs, and governance-registry records.

Kihan should therefore be evaluated on the evidence its controls produce, including whether auditors can reconstruct a decision from model input through policy evaluation to human approval. The public brief does not disclose customer-level performance metrics, so buyers should request reference architectures, sample audit records, failure-mode demonstrations, and proof-of-concept acceptance criteria.

European Central Bank Banking Supervision reported that approximately half of the 13 banks it examined had established dedicated AI policies or committees. It also emphasized model-risk controls, explainability, and monitoring. A committee may set accountability, but runtime logs and enforceable authorization rules provide the evidence needed to test whether that accountability works.

Buyer Takeaways From the Evaluation Process

A policy library should be versioned alongside the systems it governs. If a bank changes its threshold for mandatory human review, the audit trail should show which policy version evaluated each decision and when the change entered production.

To be fair, not every decision belongs in a central authorization service. Extremely latency-sensitive fraud scoring may require locally cached policies with signed updates, while higher-impact actions can tolerate synchronous approval. Architecture review should document that tradeoff rather than applying one deployment pattern everywhere.

Buyers should also ask who can modify rules, approve exceptions, and disable controls. Separation of duties can be implemented through role-based access control, dual approval for production policy changes, Git-based review, and alerts when emergency overrides remain active beyond their intended duration.

Broader Applicability and Frequently Asked Questions

Insurers, payment providers, and mid-market lenders can adapt the same pattern by narrowing the initial scope to one high-impact workflow. The control model remains similar: inventory the system, classify the decision, enforce authorization at an API or message boundary, and retain evidence in a queryable store.

How long does an AI governance implementation take?

Duration depends on inventory quality and integration depth, so buyers should plan around phases rather than a fixed timeline. A bounded AML or credit workflow can move from discovery to controlled deployment faster when identity uses OpenID Connect, decisions already pass through an API gateway, and model metadata is stored in a central registry.

What is the difference between AI governance and decision authorization?

AI governance defines ownership, risk classifications, validation requirements, monitoring, and escalation. Decision authorization applies those rules to a specific action at runtime, such as blocking an automated account restriction when human approval or current model validation is missing.

Is this approach practical for a mid-market financial institution?

It can be, provided the institution limits the first deployment to one material workflow and reuses existing infrastructure. A PostgreSQL inventory, REST policy endpoint, role-based access controls, and SIEM logging can provide a workable foundation before the organization considers a broader governance platform.