Article · AI Governance

The future of AI governance: moving from principles to enterprise practice

Most organisations can state their AI principles. Fewer can say which AI systems they run or who owns the risk. Here is what closes that gap.

On this page
  1. Why principles stop short
  2. What enterprise practice looks like
  3. Using frameworks without being ruled by them
  4. Three patterns that undermine programmes
  5. A sensible first 90 days
  6. How to tell it is working

Most large organisations now have a statement of AI principles. It usually mentions fairness, transparency, accountability and safety, and it is usually sincere. The harder test is operational. Ask which AI systems are in use today, who owns each one, what could go wrong and what evidence shows the controls are working. In many organisations, those answers take weeks to assemble.

That gap is what enterprise AI governance has to close. Principles describe intent. Governance is the set of routines that make intent visible in decisions, records and accountability.

Why principles stop short

Principles matter because they set direction. The OECD AI Principles, adopted in 2019 and updated in 2024, remain a widely used reference for transparency, robustness and accountability. But a principle does not tell a product manager whether a new vendor feature needs a review, or tell an auditor where to find the record of that review.

Three things tend to be missing: a complete view of the AI in use, a named owner for each system, and a repeatable way to decide how much scrutiny a given system deserves. Without them, governance depends on who happens to remember to ask.

What enterprise practice looks like

The specifics vary by sector and scale, but mature programmes tend to share five working components.

An AI inventory that stays current

Governance starts with knowing what exists. An inventory should cover systems built in-house, AI features embedded in purchased software and models accessed through an API. For each, record its purpose, owner, data sources, affected people, vendor and lifecycle stage. Embedded features are the entry most often missed, so include procurement in the process that keeps the list up to date.

Proportionate risk assessment

Not every system needs the same treatment. A tool that summarises internal meeting notes and a model that influences credit decisions should not pass through the same process. A simple tiering scheme lets teams move quickly on low-risk uses and reserve deeper review for systems affecting people's rights, safety or finances.

Clear ownership

Each system needs a named business owner who accepts its risk, supported by technical, legal, privacy and security specialists. Committees are useful for setting policy and handling exceptions. They are a poor substitute for individual accountability.

Controls linked to evidence

A control that cannot be demonstrated is hard to rely on. Link each control to the risks it addresses and to the evidence it produces, such as test results, approvals and monitoring reports. Capturing evidence as work happens costs far less than reconstructing it before an audit.

Monitoring and incident response

AI systems change after deployment. Data shifts, models are updated and vendors alter their services. Define what will be monitored, who reviews it, what triggers a re-assessment and how AI incidents reach the existing incident process.

Using frameworks without being ruled by them

The main references serve different purposes, and it helps to be clear about which is which.

ReferenceWhat it isHow it helps
OECD AI PrinciplesIntergovernmental principlesShared vocabulary and direction
NIST AI RMF 1.0Voluntary framework, released January 2023A structure for risk work through its Govern, Map, Measure and Manage functions
ISO/IEC 42001:2023International management system standardRequirements for establishing and improving an AI management system
EU AI ActBinding EU regulationLegal obligations that depend on role and risk level

These are complementary. Many organisations use a management system to run the programme, a risk framework to structure assessments and a regulatory map to confirm that specific legal duties are met. Mapping one internal control to several external references avoids running a separate programme for each. Note that NIST has said AI RMF 1.0 is being revised, so check for the current version.

Three patterns that undermine programmes

The registry nobody updates. An inventory built as a one-off project is out of date within months. Tie updates to events that already happen, such as procurement, architecture review and release approval.

The committee as bottleneck. If every use case waits for a monthly meeting, teams will find ways around the process. Let low-risk uses follow a pre-approved path and send only the exceptions to the committee.

Governance apart from security. Governance teams often assess fairness and compliance while security teams review infrastructure, and neither sees the whole system. A model can be well documented and still be exposed through a weak architecture. Bring threat modeling into the review of higher-risk AI systems.

A sensible first 90 days

This sequence is our suggestion and not a formal standard. Adapt it to your context.

  1. Days 1 to 30: establish the facts. Name an executive sponsor, build a first inventory that includes vendor features, and identify the legal obligations that apply to you.
  2. Days 31 to 60: define the method. Agree risk tiers, assign owners, write a short policy and assess your highest-risk systems first.
  3. Days 61 to 90: make it routine. Connect the process to procurement and release management, start capturing evidence and report status to leadership.

How to tell it is working

Coverage of frameworks is a weak signal. Better questions for leadership are these. Could we produce an accurate list of our AI systems today? Does each high-risk system have a named owner and a current assessment? When a vendor adds an AI feature, how do we find out? When did a review last change a decision? A programme that can answer those questions has moved from principles to practice.

Sources

AI governance frameworkAI inventoryAI risk managementResponsible AINIST AI RMFISO/IEC 42001

Share: LinkedInXEmail

Talk to Cognexaa

Knowledge creates confidence. Technology makes it possible.

Talk to our team about how Cognexaa’s platforms can help your organisation approach AI governance, cybersecurity and digital trust with greater structure.

Or email us directly at sales@cognexaa.com