Insight · AI governance

AI governance belongs in the design phase, not the incident review.

The most expensive governance work is the work you do after a system is already in the operating model.

By Hamza AQ Mohammed, PMP · November 13, 2026 · 8 minute read

Most AI programs do not fail because the model is incapable. They fail because the organization cannot explain what the system is allowed to do, who owns the decision, where the data came from, or what evidence proves the system is working as intended.

Those questions are easier to answer before procurement than after deployment. Once a vendor has been selected and a workflow has been built around a model, governance becomes a negotiation with sunk costs. Before that point, it is a design choice.

Start with decision rights, not the model

A governance framework that begins with model selection is already behind. The first question is not "which model?" It is "who is accountable when the system is wrong?"

For each proposed use case, define:

  • The business owner who is accountable for the outcome.
  • The technical owner responsible for the data, integration, and monitoring.
  • The risk or compliance function that can pause or reject the use case.
  • The human decision that must remain with a person, not the model.
  • The escalation path when the system produces an unexpected result.

If those roles are unclear in the design phase, they will not become clearer in production. They will simply be discovered during an incident.

Map data lineage before you commit to the use case

AI governance is often described as a model problem. In practice, it is usually a data problem wearing a model costume.

A practical lineage review asks four questions:

  1. Where does the data originate, and who owns the source of truth?
  2. What transformations occur between the source and the model?
  3. Which jurisdictions, contractual obligations, or retention rules apply?
  4. What evidence can the organization produce if a regulator, auditor, or customer asks how a decision was made?

If the answer to the last question is "we can ask the vendor," the organization does not control its own governance position.

Define the evidence before the vendor defines the demo

AI procurement frequently rewards the most convincing demonstration rather than the most defensible design. A demonstration shows what a system can do in a controlled setting. Governance requires evidence of how it behaves under real constraints.

Before approval, define the evidence the organization will require:

  • Acceptance criteria tied to the business process, not just model accuracy.
  • Test cases drawn from the organization's actual data and edge conditions.
  • Documented failure modes, fallback behavior, and human override paths.
  • Logging and retention sufficient to reconstruct a decision.
  • Security, access, and residency requirements expressed as enforceable obligations.

The point is not to make procurement adversarial. It is to make the standard explicit before commercial pressure narrows the conversation.

Use the regulatory frameworks as design inputs

Frameworks such as ISO/IEC 42001, the NIST AI Risk Management Framework, the EU AI Act, and the UK's principles-based approach are often treated as compliance documentation to be completed at the end of a program. That is the wrong sequence.

Used early, they clarify design questions:

  • Risk classification: What level of scrutiny does this use case require, and why?
  • Human oversight: Which decisions require a human in the loop, and what does that person see?
  • Transparency: What must be documented, logged, or explained to a user or reviewer?
  • Accountability: Who monitors the system after deployment, and what triggers remediation?

These are not abstract obligations. They translate directly into architecture, reporting, and operating roles. The earlier they are answered, the cheaper they are to implement.

Put governance into the statement of work

A governance framework that lives only in a policy document has no operational force. It becomes real when it appears in the contract.

For implementation partners, the statement of work should specify:

  • Acceptance criteria and evidence required for each milestone.
  • Data lineage, retention, and deletion obligations.
  • Security and access constraints, including subcontractor obligations.
  • Monitoring, incident reporting, and remediation responsibilities.
  • Handoff requirements so the organization can operate the system without the vendor.

This is also where an independent reviewer adds value. If the same firm designs, builds, and certifies the system, the organization has purchased an opinion, not assurance.

Test the operating model, not just the technology

A successful pilot proves that a model can produce an output. It does not prove that the organization can govern the output at scale.

Before scaling, test the operating model:

  • Can the business owner explain how the system is performing?
  • Can the risk function retrieve the evidence it would need in a review?
  • Can operations identify and escalate an unexpected result?
  • Can the organization change the model or vendor without rebuilding the process?

If the answer is no, the pilot has validated a demo, not an operating capability.

What a board should ask before approval

Boards and executive committees do not need to evaluate model architecture. They do need to ask whether the organization can govern what it is approving. A short set of questions is usually enough:

  1. What decision does this system influence, and who remains accountable for it?
  2. What data does it use, where does that data come from, and what are the restrictions?
  3. What evidence will demonstrate that it works in our environment, not the vendor's demo?
  4. What happens when it fails, and who can stop it?
  5. What will we be able to prove to a regulator, auditor, or customer after deployment?
  6. What must be true for us to expand, pause, or retire the system?

If those questions cannot be answered before approval, the program is not ready for approval. It is ready for design work.

The governance dividend

Governance is often framed as friction. In practice, well-designed governance reduces rework, shortens procurement disputes, and makes it easier to scale a use case that has already been proven.

The goal is not to slow AI adoption. It is to make the first production deployment the start of a durable capability rather than the beginning of an expensive correction.

Learn about HAQM's AI Governance & Risk engagement

Bring a specific AI decision to the conversation.

We will help you frame the governance question, the evidence required, and the next executable step.