Who Owns the Decision When AI Has an Opinion?

Imagine this: a manager is reviewing two candidates for an important position. Her AI-enabled system analyzes their information and recommends Candidate A. The manager preferred Candidate B—but the system has access to more data, identifies patterns she hasn't considered, and presents its recommendation confidently, so she chooses Candidate A.

Three months later, the hire isn't working out. Who owns the decision? The manager? The executive who approved the AI system? The company that developed it? Or the AI?

Now replace hiring with a customer credit decision, employee performance assessment, pricing recommendation, safety decision, investment proposal, or strategic forecast.

Suddenly this isn't a technology question, it's a leadership question and as AI moves from helping us complete tasks to influencing decisions, CEOs and leaders need to answer it.

When Advice Starts Looking Like Authority

We've spent considerable time asking what AI can do and I think leaders increasingly need to ask a different question:

What decisions should AI influence—and what decisions must humans continue to own?

Recent research reported by Harvard Business Review research, When Employees Are Held Accountable for AI-Generated Decisions, examined an emerging organizational problem: employees being required to communicate, justify and defend AI-generated decisions they didn't create and may not fully understand. That's an uncomfortable position.

Imagine telling a customer, "Your application was declined, but I can't really explain why" or telling an employee: "The system identified you as a performance risk."

The technology may have generated the recommendation, but a human being still faces the person affected by it and that's where accountability becomes real.

AI Can Recommend. Leaders Must Decide.

My concern isn't that leaders will trust AI too little, it's that we may gradually trust it without realizing how much judgment we've surrendered.

Research on "automation bias" predates today's generative AI and researchers have documented circumstances in which people follow automated recommendations despite contradictory information, in part because they believe the automated system has superior judgment.

Today's AI makes this challenge even more interesting because its recommendations can sound remarkably confident, but confidence isn't accountability.

NIST's AI Risk Management Framework explicitly calls for organizations to clearly define and differentiate human roles and responsibilities in AI decision-making and oversight. It recognizes configurations ranging from autonomous systems to AI acting simply as an additional opinion for a human decision-maker and for leaders, that suggests an important principle:

Never give someone accountability for a decision without also giving them sufficient understanding, authority and ability to challenge it.

Not Every Decision Needs the Same Level of Human Judgment

An AI recommendation about when to reorder office supplies is different from one affecting someone's employment, finances, safety or future. The higher the consequence, ambiguity or human impact, the stronger the case for meaningful human judgment and clearly defined escalation. I suggest leaders think about decisions in three categories:

🟢 AI Assist

AI provides information, analysis or options. Human: decides. Examples might include summarizing research, identifying trends or brainstorming alternatives.

🟡 AI Recommend

AI proposes a preferred course of action. Human: reviews, challenges and accepts or rejects the recommendation. This requires sufficient expertise to make human review meaningful.

🔴 AI Escalate

The potential impact, uncertainty or risk warrants additional human review. Human: pauses the process and escalates to someone with appropriate authority or expertise. The specific boundaries will differ by business and regulatory context. The important part is defining them before something goes wrong.

OECD's updated AI Principles similarly emphasize human agency and oversight, transparency, traceability and accountability based on people's roles and their ability to act.

The Learning Leader™ Doesn't Blindly Accept—or Reject—AI

There's another trap: some leaders distrust AI simply because they don't understand it, while others trust it because they don't understand it. Neither is particularly useful.

The Learning Leader™ stays curious so when AI recommends something unexpected, don't immediately ask: "Should I accept this?"

Ask:

• What information informed this recommendation

• What might the system not know?

• What assumptions are embedded in it?

• What evidence contradicts it?

• What are the consequences if it's wrong?

• Would I be comfortable explaining this decision to the person affected?

AI should challenge human thinking and humans should challenge AI thinking!

The opportunity lies in the interaction. NIST notes that human-AI combinations can sometimes improve performance—but can also amplify human biases under certain conditions. Human oversight therefore can't simply mean putting a person at the end of an automated process and calling it responsible.

The human has to be capable of actually overriding it.

Exercise: The AI Decision Audit

Choose three decisions in your organization where AI is already being used—or soon will be and for each, answer:

1. What role does AI play? Information? Recommendation? Decision?

2. Who makes the final decision? Name the role—not "the team."

3. Does that person understand enough to challenge the recommendation?

4. What triggers escalation? Define the circumstances before they happen.

5. Can we explain the decision? Could you explain what happened to an employee, customer, board member or regulator?

6. Who ultimately owns the outcome? If that answer isn't immediately obvious, you've discovered a governance gap.

AI Accountability and the Five Focuses for Agile Business Growth

🧠 Leadership & Mindset: Move from "AI gave us the answer" to: "AI gave us information. What judgment does leadership now require?"

🧭 Strategy & Direction: Decide deliberately where AI should assist, recommend or operate—and where human judgment remains essential.

⚙️ Execution Systems: Build decision rights, documentation, review and escalation into workflows rather than relying on individual discretion. NIST's governance guidance specifically recommends defining human oversight roles, training requirements and organizational risk tolerances for human-AI configurations.

👥 People & Culture: Employees need psychological safety to say: "The AI recommendation doesn't make sense to me." If challenging the system feels career-limiting, human oversight becomes theatre.

💡 Innovation & Growth: Good governance shouldn't prevent experimentation, rather it should make responsible experimentation easier because everyone understands the boundaries.

Questions for Your Leadership Team

At your next leadership meeting, try these:

• Where is AI already influencing decisions in our organization?

• Which of those decisions could materially affect people or the business?

• Who is accountable when AI contributes to a poor outcome?

• Can employees challenge an AI recommendation without penalty?

• What decisions should never be delegated entirely to AI?

• Where are we holding someone accountable without giving them sufficient authority or understanding?

• What would our board expect us to be able to explain?

Don't assume everyone will give the same answers since that discussion itself may reveal where work is needed.

Three Actions to Take Now

1. Map AI-influenced decisions: you can't govern what you don't know exists.

2. Clarify decision rights: for consequential uses, establish who recommends, who decides, who can override and when escalation occurs.

3. Build learning loops: review decisions where AI was right, wrong—or simply surprised you. HBR's practical guidance similarly recommends creating opportunities for employees to learn how AI-generated decisions are produced and communicated.

Don't wait for failure to learn how your human-AI decision system actually behaves.

The Bottom Line

AI is becoming extraordinarily good at having an opinion, but an opinion isn't accountability and outsourcing analysis should never quietly become outsourcing leadership. The organizations that navigate AI successfully won't be those where humans make every decision nor will they be those where AI does. They will be organizations where leaders deliberately understand which decisions belong where—and why.

And that's part of becoming The Learning Leader™:

• Curious enough to use AI.

• Knowledgeable enough to question it.

• Humble enough to learn from it.

And accountable enough to say:

"AI informed the decision. We own it."

because when the consequences arrive, leadership cannot point to the algorithm. Someone still must own the decision.

Jerome Dickey