There are still a surprising number of technology leaders who have not taken a clear position on AI in the workplace. The challenge is rarely access itself. While 93% of organisations are leveraging AI, only 7% have embedded governance frameworks. The difficult part is agreeing early on the principles that should guide decisions when trade-offs arise. For many organisations, that is where the real governance problem begins.
AI governance is often treated as a technical exercise, and the conversation starts with tooling, policies, implementation details or compliance frameworks. They are all important, but governance is only meaningful when an organisation has decided on what it is prepared to accept and what it wants to protect. This is why ethical dilemmas should be put forward to clarify priorities, uncover any assumptions and create a strong foundation for AI governance.
Why dilemmas work better than policy-first discussions
McKinsey’s The state of AI in 2025 report shows that 64% of respondents say AI is enabling innovation, but many are still in the experimentation or piloting phase, and governance concerns are likely playing their part. Many AI governance conversations fail because they begin at the wrong level. When discussions focus on model types or control frameworks, for example, it is easy for senior leaders to stay at arm’s length. But an ethical dilemma changes that. It places the board in a situation where they must consider whether something should be done and who would be responsible if the outcome falls short.
Starting with a practical dilemma also helps avoid governance that exists just in writing. It is easy to approve high-level principles that sound sensible. It is much harder, and much more useful, to apply those principles to a realistic scenario where cost savings, risk and organisational values are creating a conflict.
A scenario to put before the board
As an example of an ethical dilemma presented to the board, consider a situation where an AI vendor runs a proof of concept for a business and concludes that its product could replace 30% of the team’s headcount. Because the licence cost comes in below the equivalent staffing cost, the commercial case appears strong on the surface.
But dig a little beneath that surface, and it isn’t so simple. A comparison between licence cost and salary cost may appear initially persuasive, but it tells only part of the story. Decision-makers don’t have assurances about the quality of AI-driven outcomes, the overall reliability of the system, the impact on customers, or the organisational risks created by changing how work is completed.
The second question is what kind of work is actually being displaced. If the AI system is reducing repetitive, rules-based tasks, it may be a case that’s easier to support. But if it is replacing judgement-heavy work or inadvertently weakening the strong link between teams and customers, it may not be worth the investment.
A key third question is who remains accountable. Research shows 17% of businesses report no clear AI owner, therefore hindering pace, alignment and impact. If an automated process produces poor outcomes or introduces bias which potentially causes customer harm, responsibility does not sit with the software and instead with a responsible person. This means the board should be clear on who owns the decision, who monitors the outputs and who has the authority to intervene if something goes wrong.
The second-order effects
Cost savings may well be possible with AI, but deployments can also give rise to second-order effects. Removing headcount is a clear example of reducing internal capability in ways that may not be initially visible. Teams with fewer humans may find it much more difficult to respond quickly when conditions change, or even be hindered in their ability to recover from failure. Lack of human support may also lead to inadvertent errors that negatively impact the organisation’s resilience.
Trust is also key here. Staff will have watched closely how leadership decisions around AI are made and justified. Staff engagement and confidence are likely to suffer if the organisation appears to treat AI primarily as a route to labour reduction, rather than augmenting staff to help them complete tasks more quickly and efficiently.
There’s also the question of customers. In some functions, automation with AI may improve speed and consistency. But in others, it may damage the quality of service or create frustration if human capabilities disappear too quickly. Governance has to account for that broader impact.
Once the board has worked through a realistic scenario, the next step is to translate that discussion into action. In many cases, a small number of decisions can create useful momentum, enabling leadership to define which AI use cases need board visibility or executive approval.
Then, clear expectations around accountability can be set, alongside assessments for higher-risk use cases, particularly where employee roles or sensitive decisions are involved. Agreements on what must be measured, such as output quality or error rates, can be defined. Importantly, organisations can also agree on what they won’t do, even if the commercial case is appealing.
Where meaningful AI governance begins
AI governance is failing because the difficult conversations happen too late, once a tool has already been chosen or the business case is already in motion. Ethical dilemmas offer a better starting point by forcing leadership teams to confront any compromises head-on, and define where accountability and judgement should sit.
For organisations still finding their footing with AI, grounded conversations are needed for when efficiency, responsibility and trust do not neatly align. That is the place where meaningful AI governance begins.
