AI Governance and Risk Management with the NIST AI RMF

Why AI governance belongs at the start

Most companies discover they need AI governance after the fact: a sales team pasting customer lists into a public chatbot, a support tool that told a customer the wrong return policy, or an enterprise buyer whose security questionnaire suddenly asks how your product uses AI. Tepia helps you put governance in place before those moments, sized for a mid sized company rather than a bank with a dedicated risk department.

Governance here means three practical things. You know which AI systems you use and build, and who owns each one. You have assessed what could go wrong with each and decided what is acceptable. And the controls you decided on actually exist in the software and in daily practice, not only in a policy document. Enterprise customers, auditors and regulators increasingly ask for evidence of all three.

The NIST AI Risk Management Framework is a useful backbone because it is voluntary, widely referenced in US procurement and security reviews, and flexible enough for an organization of any size. Tepia does not certify compliance. It helps you build the artifacts and the software controls the framework describes.

The NIST AI RMF in practical terms

The framework is organized around four functions. Tepia translates each into deliverables a product owner and an IT lead can maintain without a full time governance team:

The framework is deliberately not a checklist, which frustrates teams looking for one. Tepia’s job is to turn it into a short list of documents and recurring tasks. The goal is a record you can show a customer or auditor, and a habit of reviewing AI changes before they ship, the same way you review security changes.

Risks specific to generative AI

NIST’s Generative AI Profile adds risks that traditional software reviews miss. Tepia focuses on the ones most likely to affect a business application, and pairs each with controls:

• Confabulation: the model states false information confidently. Controls include grounding answers in approved sources, showing citations and keeping a person in the loop for consequential output.
• Data privacy: personal or confidential data sent to a model, logged, or exposed in a response. Controls include field masking, retention settings and role based access to retrieved content.
• Information security: prompt injection through documents or emails the model reads, and tools with too much permission. Controls include treating model input as untrusted and scoping every tool narrowly.
• Harmful bias: uneven output quality or treatment across customer groups. Controls include evaluation sets that cover those groups and periodic review.
• Intellectual property: generated content that reproduces protected material, or unclear rights to output. Controls include provider terms review and human sign off on published content.
• Value chain: dependence on third party models and components you do not control. Controls include version pinning, fallback providers and vendor review.

Each control is something Tepia can build and test: an audit log, a masking step, a confirmation screen, a case in the evaluation suite. That is the difference between a policy and a governed system.

Evaluating AI vendors and model providers

Most AI risk enters through vendors, whether a model API inside your product or an AI feature switched on in a tool your staff already use. Tepia uses a consistent set of questions when reviewing providers for clients:

• Are prompts and outputs used to train models, and is that off by default under your agreement?
• How long are inputs and logs retained, and can retention be shortened?
• Where is data processed and stored, and can it stay in a chosen region?
• Is a business associate agreement available for health data, and which services does it cover?
• Which security reports, such as a SOC 2 report, can the vendor share?
• How are model versions announced, deprecated and retired?

The answers go into the AI inventory alongside each system, so the next review starts from a record instead of memory. For regulated environments, pair this with our security and compliance work and, for health data, our page on AI in healthcare apps.

How Tepia approaches AI governance and risk

Governance work follows Tepia’s six phase process, with a US based project manager coordinating your legal, IT and product owners.

1. Discovery: interviews with leadership, IT, legal and the teams already using AI, plus a review of current tools and AI features in your products. Deliverables are an Investigation Summary that doubles as the first AI inventory, an Interview Summary and User Stories for each governed system.
2. Design: the AI use policy, risk register format and approval workflow, plus wireframes for in product controls such as AI disclosure labels and human approval steps.
3. Development and Testing: controls built into the software, including audit logging, data masking, permission scopes and evaluation tests for accuracy and bias, verified through functional and non functional test plans.
4. Training: sessions for staff on the acceptable use policy and for reviewers on assessing AI output, using scenarios from your own workflows.
5. Launch: rollout of the policy and controls, with the risk register handed to named owners and the first review date set.
6. Support: periodic reviews as models, vendors and regulations change, plus issue resolution and updates to controls.

Discovery for governance work typically takes 1 to 3 months, often running alongside the first AI build. Read our process or talk to Tepia about your AI program.

Frequently asked questions