Introduction: The Same AI, Two Different Outcomes

Artificial intelligence is becoming part of everyday business leadership. From preparing reports and analysing financial performance to developing strategies and evaluating investments, AI is changing how information reaches decision-makers.

But I have been observing something interesting.

Two leaders can use exactly the same AI technology and arrive at very different conclusions.

One uses AI to question assumptions, explore alternatives, validate information and strengthen decisions.

Another may use it to reinforce an existing opinion, turn assumptions into convincing presentations or support a business case that has not been properly tested.

Both can produce professional-looking reports. Both can present impressive charts and confident recommendations.

Yet the quality of the underlying judgment can be completely different.

This raises an important question: Is AI improving leadership decisions, or is it sometimes making weak decisions look more convincing?

The answer depends less on the sophistication of the technology and more on how leaders use it.

1. Evidence-Driven vs. Assumption-Driven AI Usage

In my experience working with commercial decisions, pricing, business processes and operational data, one principle has remained consistent: the quality of a decision depends on understanding the information behind it.

AI does not change that principle.

Evidence-driven leadership

An evidence-driven leader approaches AI with curiosity but also healthy scepticism.

Before accepting a recommendation, they want to understand:

• Where did the information come from?
• Which facts are verified, and which are assumptions?
• What information might be missing?
• What alternative explanations should be considered?
• What are the operational and financial consequences?

They use AI to explore possibilities rather than simply confirm their expectations.

Importantly, they bring their own business knowledge into the discussion. They understand that a recommendation may appear mathematically correct while being commercially or operationally unrealistic.

Assumption-driven leadership

A different pattern emerges when a leader begins with a preferred conclusion.

Instead of asking AI to investigate a problem objectively, they may ask questions that guide it toward the answer they want.

For example:

"Show me why this automation will reduce operating costs."

That question already assumes the automation will deliver savings.

A more balanced question would be:

"Evaluate whether this automation can reduce operating costs. Identify the assumptions, implementation costs, limitations, risks and alternative options."

The difference may appear small, but it changes the entire direction of the analysis.

AI should help leaders investigate whether an idea is right, not simply explain why they are right.

2. When Assumptions Become Facts

One of the risks of AI-generated business analysis is how easily assumptions can be presented as established facts.

Consider a project proposal that assumes:

• An automation will eliminate 50% of manual work.
• Every hour saved will translate into financial savings.
• Implementation will require minimal resources.
• Existing systems will integrate without difficulty.
• Productivity improvements will begin immediately.

AI can turn these assumptions into a convincing business presentation within minutes.

It can create executive summaries, financial projections, charts, timelines and recommendations.

But none of those outputs independently proves that the assumptions are correct.

This becomes especially problematic when estimates are repeated across reports until people begin treating them as verified information.

A forecasted benefit becomes an expected saving. An expected saving becomes a financial commitment. Eventually, the original uncertainty disappears from the discussion.

This is not necessarily deliberate manipulation. Sometimes it is simply a lack of understanding, insufficient scrutiny or excessive confidence in the technology.

However, when selective information is intentionally used to influence a predetermined decision, the problem becomes one of leadership integrity rather than technological capability.

3. The ROI Problem: When Impressive Numbers Hide Weak Assumptions

Return on investment is particularly vulnerable to assumption-driven analysis.

Imagine a company introducing an AI solution to automate administrative work.

The proposal claims the system will save 10,000 working hours annually.

At an assumed labour cost of €20 per hour, the business case presents €200,000 in annual benefits.

The calculation is straightforward.

But does the company actually save €200,000?

Not necessarily.

The employees may still be required for other activities. The automation may handle only part of the process. Human verification may remain necessary. Software subscriptions, implementation, maintenance and training introduce additional costs.

Time saved is valuable, but it is not automatically equivalent to cash saved.

A realistic business case must distinguish between:

Productivity improvement: Employees can complete more work in the same amount of time.

Cost avoidance: Additional hiring or expenditure can be prevented.

Realised financial savings: Actual expenses are reduced or measurable incremental financial benefits are achieved.

These are different outcomes and should not be presented interchangeably.

A credible ROI calculation should also account for the cost of achieving the benefit, the period over which it will occur and the uncertainty surrounding the forecast.

AI can calculate ROI extremely quickly. But the accuracy of that calculation still depends on the reliability of its inputs.

A mathematically correct ROI built on unrealistic assumptions remains an unreliable business case.

4. What Research Tells Us About AI and Decision Quality

The risks are not merely theoretical. Research increasingly shows that AI's impact depends on the task, the user and the level of human oversight.

AI can reduce performance when people trust it too much

A 2023 study involving Boston Consulting Group consultants, conducted with researchers associated with Harvard Business School and other institutions, found that AI significantly improved performance on tasks within its capability range.

However, when participants used GPT-4 for a business problem-solving task outside that range, their performance was 23% worse than that of participants working without AI.

The finding illustrates an important limitation: AI can be useful in one situation and misleading in another.

Even when users are warned that AI may produce incorrect answers, they do not always challenge those answers.

Leadership decisions play a major role in AI project failures

A 2024 RAND Corporation report interviewed 65 experienced AI practitioners.

Among its industry interviewees, 84% identified at least one leadership-driven issue as a primary reason AI projects fail.

Recurring problems included poorly defined business objectives, inadequate communication, insufficient data and deploying technology without understanding the actual business need.

This percentage reflects the views of the interviewed practitioners, not the measured failure rate of AI projects across all industries.

Nevertheless, the message is significant: many AI problems begin before the technology is deployed.

Confidence in AI can affect critical thinking

A 2025 Microsoft Research and Carnegie Mellon University study surveyed 319 knowledge workers and collected 936 examples of AI use.

The researchers found that higher confidence in AI was associated with lower reported critical-thinking effort.

The study does not establish that AI inevitably reduces critical thinking. It does, however, highlight the importance of verification and human oversight.

When AI produces polished, confident answers, users may be less inclined to investigate further.

AI can also help less-experienced employees

There is an important balance to this discussion.

Research by Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that AI assistance improved productivity among novice and lower-skilled customer-support workers by approximately 34%, with much smaller gains among experienced workers.

This suggests that AI can help people learn and benefit from established best practices.

Therefore, the issue is not simply experienced versus inexperienced leadership.

The more meaningful distinction is between leaders who apply critical judgment and those who accept information without sufficient examination.

5. Why Business and Process Knowledge Still Matter

This is where professional experience becomes particularly valuable.

A leader who understands how a business operates can identify problems that may not be visible in an AI-generated report.

Consider an example from manufacturing or spare-parts pricing.

AI identifies two purchase prices for the same material and recommends adjusting the lower price to match the higher one.

On the surface, the analysis appears reasonable.

But someone familiar with the business may recognise several possible explanations.

The prices could relate to different purchasing quantities, packaging units, currencies, intercompany markups, supplier agreements or validity periods.

Without understanding these details, an automated recommendation could introduce new errors rather than correct existing ones.

The same principle applies to automation projects.

A report may show that a task takes ten minutes and occurs thousands of times annually.

AI may calculate substantial potential savings.

But an experienced process owner will ask whether those ten minutes represent active work, whether the process includes mandatory checks and whether the proposed automation actually eliminates the underlying activity.

AI can process information.

Business knowledge helps determine what that information means.

This is why I believe organisations should not underestimate employees who understand systems, customers, operations and the relationships between different business processes.

Their knowledge becomes even more valuable when AI is introduced.

6. AI as a Tool for Investigation, Not Justification

There is another leadership risk worth discussing: confirmation bias.

People naturally tend to favour information that supports what they already believe.

AI can unintentionally reinforce this behaviour.

A leader who asks for arguments supporting a preferred strategy may receive a well-structured explanation. If they continue refining the request in the same direction, the final report can become increasingly persuasive without becoming more accurate.

This can create the appearance of independent analysis when the output has actually been shaped by the user's expectations.

A stronger approach is to deliberately challenge the original proposal.

For example, leaders can ask AI to:

• Identify the weakest assumptions in a business case.
• Develop arguments against the proposed investment.
• Compare optimistic, realistic and downside scenarios.
• Explain which additional evidence would change the recommendation.
• Separate verified facts from estimates and untested claims.

This does not mean leaders should become unnecessarily negative or resistant to innovation.

It means that important decisions deserve more than one perspective.

Good leadership is not about proving a preferred answer. It is about being willing to discover a better one.

7. Five Practical Principles for Better AI-Assisted Leadership

Organisations do not need to make AI governance unnecessarily complicated.

Several practical disciplines can substantially improve the reliability of AI-assisted decisions.

1. Start with the business problem.

Clearly define what needs to improve before deciding how AI should be used. Technology should support the objective, not become the objective itself.

2. Separate facts, assumptions and estimates.

Every important business case should clearly identify which numbers are verified and which depend on forecasts or judgment.

3. Validate financial benefits.

Distinguish time savings from cost savings. Include implementation costs, ongoing expenses, risks and the conditions required to realise projected benefits.

4. Involve people who understand the process.

Employees with operational and domain expertise should participate in evaluating AI recommendations. Technical accuracy alone is not enough.

5. Encourage constructive challenges.

Leaders should welcome questions about the evidence behind recommendations, especially when reports appear unusually optimistic or support decisions already made.

These principles apply whether an organisation is introducing a sophisticated AI platform or simply using generative AI to prepare everyday management reports.

8. The Leadership Challenge: Accountability Cannot Be Automated

AI is becoming more capable, faster and increasingly accessible.

It can analyse large datasets, identify patterns, compare scenarios and produce reports that once required significant manual effort.

These capabilities create genuine opportunities.

But they also introduce a new leadership responsibility.

When a recommendation comes from AI, who is accountable for verifying it?

When an investment proposal contains unrealistic assumptions, who challenges them?

When projected benefits fail to materialise, who takes responsibility for the original decision?

The answer cannot simply be the AI system.

Technology can support the decision-making process, but leadership accountability remains human.

This is especially important when decisions affect employees, customers, investment priorities and long-term business performance.

An organisation that encourages employees to challenge assumptions is more likely to identify weaknesses before they become expensive mistakes.

An organisation that rewards only confident presentations and optimistic forecasts may overlook those weaknesses, regardless of how advanced its technology becomes.

Conclusion: AI Amplifies Leadership Judgment

AI is neither a guarantee of better decisions nor an inevitable source of poor ones.

Its value depends on how it is used.

Evidence-driven leaders use AI to expand their understanding, challenge their thinking and evaluate alternatives.

Assumption-driven leaders risk using the same technology to reinforce existing beliefs, produce misleading business cases or create confidence that the underlying evidence does not justify.

Experience and domain knowledge can help leaders recognise these risks, but experience alone is not enough. Openness, intellectual honesty, curiosity and the willingness to question one's own conclusions are equally important.

The future of effective AI leadership will not be determined simply by who adopts the technology fastest.

It will depend on who can combine technological capability with business understanding, critical thinking and accountability.

AI does not replace leadership judgment. It amplifies it.

And perhaps the most valuable AI leadership skill is not knowing how to generate a convincing answer, but knowing when that answer needs to be challenged.

Final Thought

Before approving the next impressive AI-generated report or investment proposal, consider asking one additional question:

“What evidence would prove that our assumptions are wrong?”

The answer may be more valuable than the report itself.