
A company can publish responsible AI principles and still leave its employees unsure about what they are allowed to do.
It can also have an approval process that is followed carefully while asking too few questions about the people affected by the technology.
That is why I find it useful to discuss responsible AI and AI governance together. The terms overlap, and organizations use them differently. For business planning, I use a practical distinction: responsible AI describes the qualities and outcomes we want our AI use to uphold; governance describes how we assign responsibility, make decisions and check whether those intentions are being carried through.
This is a working distinction, not a universal definition.
Principles need a place in the workflow
The OECD AI Principles address themes including transparency, robustness, human-centred values and accountability. They provide a reference for discussing what trustworthy AI should involve. Source: OECD AI Principles.
For a business, the next question is what those themes mean in a particular workflow.
If you value transparency, what information should the user receive? If you value accountability, who responds when an output is challenged? If you value fairness, which people or situations need to be considered in evaluation, and who has the expertise to do that work?
A statement of intent does not resolve those questions on its own.
Put the distinction into an ordinary business example
Illustrative scenario: An e-commerce technology company provides an AI feature that drafts product descriptions from a retailer's catalog.
The responsible AI discussion includes accuracy, misleading claims, accessibility and whether important product details are lost or invented.
The governance discussion includes who approves the feature, which source fields it can use, who reviews generated descriptions, which claims require specialist approval and what happens when a retailer reports an error.
Suppose the draft adds “allergen-free” even though the catalog contains no supporting information. The immediate task is to correct or remove that description. The broader review asks how the claim reached publication and whether other descriptions need checking.
Those are connected pieces of work. The principle gives the team a reason to care about accuracy. The process gives it a way to respond.
Three different kinds of evidence
I would separate evidence of intent, evidence of operation and evidence of results.
Intent: The company has committed to reviewing sensitive claims before publication.
Operation: There is a defined reviewer, and records show that drafts are being sent to that person.
Results: The team has examined whether the review detects the kinds of mistakes that matter in this use case.
Each answers a different question. A principles page supports the first. It cannot, by itself, establish the second or third.
This distinction is useful when teams discuss maturity. Otherwise, an organization may award itself credit for having a policy while leaving the effectiveness of the practice unknown.
Avoid treating governance as a paperwork project
A process can become so cumbersome that employees cannot use it in daily work. That deserves attention too.
I would ask whether staff can find the rules, whether the approval route is understandable and whether there is a sensible way to ask for an exception. The answer may be a clearer instruction or a better escalation route, rather than another document.
For a low-consequence drafting task, the process may be straightforward. A system that influences hiring or takes actions in customer accounts deserves a more substantial examination. “Low consequence” should still be checked against the actual data, access and use.
A documented example of turning principles into requirements
Microsoft publishes a Responsible AI Standard with requirements for its own organizational practice, including impact assessments. It is a useful example of expressing principles through a defined process. Its existence does not establish how another company's system performs, and adopting a similar document would not establish that either. Source: Microsoft Responsible AI Standard v2.
For a B2B team, I would begin with one use case and ask: what quality matters here, what decision or control supports it, who owns that work, and how will we know whether it is happening?
That keeps the conversation connected to the product and the people using it.
Apply responsible AI principles to your customer communication with the free Responsible AI Claims & Buyer Trust Check. The AI Claims & Messaging Review connects product promises to evidence, limitations and clear approval responsibilities.