By Markus Kopko, PgMP®, PMP®, PMI-CPMAI™
March 19, 2026
A 2024 Deloitte survey found that 47% of enterprise AI users made at least one major business decision based on hallucinated content. Not outdated content. Not imprecise content. Content that was fabricated by the AI model and presented with full confidence.
For project managers, that statistic should trigger an immediate question: How many of our project decisions are based on AI outputs that no one verified?
The answer, in most organizations, is uncomfortable. A 2025 Microsoft Research study of knowledge workers found that generative AI shifts the core cognitive effort from producing outputs to evaluating them, with workers reporting increased time on verification, accuracy checking, and source validation (Lee et al., 2025). Verification is not a minor add-on. It is becoming a primary workload. And the organizations that skip it are accumulating risk with every AI-assisted deliverable.
The Failure Data That Matters
The conversation about AI in project management tends to focus on productivity gains: faster drafting, automated reporting, predictive scheduling. Those gains are real. The Adecco Group’s 2024 Global Workforce of the Future survey of 35,000 workers across 27 economies found that employees using AI for routine tasks save an average of one hour per day.
But the failure data tells a different story. MIT’s Project NANDA reported in its 2025 GenAI Divide study that only 5% of generative AI pilots deliver measurable business impact. RAND Corporation found that over 80% of AI projects fail, twice the rate of non-AI technology projects. S&P Global’s 2025 survey showed that 42% of companies abandoned most AI initiatives, up from 17% one year earlier.
The primary reasons are not technical. The Informatica CDO Insights 2025 survey identified the top obstacles: data quality and readiness (43%), lack of technical maturity (43%), and shortage of skills (35%). Computing power and model selection did not make the list. The failures are human. They are failures of judgment, governance, and oversight that happen when organizations treat AI as a replacement for human decision-making rather than an augmentation of it.
Why Human Judgment Matters More, Not Less
There is a growing body of research that suggests AI does not sharpen human judgment. It dulls it.
A 2025 study published in Societies (Gerlich, 2025) surveyed 666 participants and found a significant negative correlation between frequent AI tool usage and critical thinking abilities. The mechanism: cognitive offloading. When people delegate reasoning to AI, they stop exercising the judgment muscles that decision-making requires. A 2024 MIT study confirmed this: users who relied heavily on generative AI produced less original work and retained less information, even when they believed the tool was helping them.
For project management, this creates a specific risk. Project managers make dozens of decisions daily that require contextual judgment: whether a risk is material, whether a stakeholder concern is political or substantive, whether a schedule variance signals a systemic problem or a one-time delay. These decisions require pattern recognition built from experience, organizational knowledge, and situational awareness. AI has none of these. It has statistical patterns trained on historical data.
PMI recognized this explicitly as early as 2023: “GenAI can support project management tasks in many ways. The more complex the task, the more human intervention is needed to result in high-quality outcomes.” The PMBOK® Guide, Eighth Edition, reinforces this position. Its principle of leading accountably places the responsibility for project outcomes with people, not tools. Its principle of building empowered teams means teams that exercise their own judgment, not teams that delegate judgment to AI. And its Governance performance domain requires the oversight structures that make human-in-the-loop operational rather than aspirational.
As a member of the Core Development Team for the PMI Standard on AI in Project, Program, and Portfolio Management, I can confirm: the human-in-the-loop principle is not an afterthought in the standard. It is foundational.
The Complexity Spectrum
Not all project management tasks carry the same judgment requirements. Understanding where human oversight is critical and where AI can operate with less supervision is the starting point for any practical approach.
On one end sit administrative tasks: creating agendas, capturing minutes, documenting resource usage, generating test cases. These are high-volume, low-complexity tasks where AI delivers clear productivity gains. Human review is still required, but the stakes of an error are low.
In the middle sit tasks requiring contextual interpretation: developing timelines from historical data, analyzing risks, aggregating project health reports into portfolio summaries. AI accelerates the work, but a project manager must validate the output against organizational context the model does not possess.
At the far end sit strategic decisions: portfolio prioritization, resource conflict resolution, go/no-go decisions, benefits realization assessment. These require stakeholder judgment, political awareness, and accountability that cannot be delegated. AI can inform these decisions. It cannot make them.
The pattern is consistent: as complexity increases, the requirement for human judgment does not decrease. It increases. AI does not replace the project manager at any point on this spectrum. It changes what the project manager spends time on.
What Human-in-the-Loop Means in Practice
The term “human-in-the-loop” is used frequently but defined rarely. In project management, it means three specific things.
First, verification before action. No AI-generated output enters a deliverable, a decision, or a stakeholder communication without a qualified person reviewing it for accuracy, completeness, and contextual fit. This applies to every AI-assisted task.
Second, accountability stays with people. When AI assists in drafting a risk response plan, the project manager who submits that plan owns its content. When AI generates a cost estimate, the estimator who signs off is responsible for accuracy. AI shifts the nature of the work. It does not shift accountability.
Third, active quality monitoring over time. A 2024 mathematical proof (Xu et al.) confirmed that hallucinations are structurally inevitable under current large language model architectures. Retrieval-Augmented Generation reduces them by up to 71% (AIMultiple, 2025), but cannot eliminate them. Verification is not a one-time gate. It is a continuous discipline.
The Cost of Skipping Oversight
The consequences of skipping oversight are already visible. In early 2024, Air Canada was ordered to honor a fare discount its AI chatbot had incorrectly promised a passenger, a ruling that established direct corporate liability for AI-generated misinformation. Judges across the U.S. issued hundreds of decisions addressing fabricated AI-generated legal citations.
In project management, the equivalent is a cost estimate based on fabricated historical data, a risk assessment that identifies the wrong dependencies, or a stakeholder communication that misrepresents status because the model summarized selectively. These are the predictable outcomes of AI without systematic human oversight.
Making It Operational
Embedding human-in-the-loop into daily project work does not require a transformation program. It requires four decisions.
First, define which tasks require what level of review. Use the complexity spectrum as a guide. Administrative outputs get a quick check. Analytical outputs get substantive review. Strategic inputs get full human judgment with AI as one data source among several.
Second, assign accountability explicitly. For every AI-assisted deliverable, one person is named as reviewer and owner.
Third, build verification into the workflow. Add “AI output verified” as a criterion in your definition of done. Track it. Make it visible in your governance reporting.
Fourth, monitor for drift. AI models change. Data quality fluctuates. Verification habits degrade under time pressure. Schedule periodic reviews of how AI is being used and where shortcuts have crept in.
The Project Manager’s Role in the AI Era
Over 80% of AI projects fail (RAND). Not because the technology is insufficient, but because human judgment, governance, and oversight are insufficient. The tools will improve. The hallucination rates will decline. None of that changes the fundamental requirement: in project management, accountability is human. Decisions are human. Judgment is human.
The real question is not whether AI will replace project managers. The real question is whether project managers who use AI without systematic oversight will be replaced by those who do. Accountability is not an optional feature of AI-assisted project delivery. It is the job.
Key Takeaways:
- 47% of enterprise AI users made at least one major decision based on hallucinated AI content (Deloitte, 2024).
- 80% of AI projects fail. The primary causes are human: data quality, governance gaps, and insufficient oversight (RAND, Informatica CDO Insights 2025).
- The PMBOK® Guide, Eighth Edition, reinforces human-in-the-loop through its principles of leading accountably and building empowered teams, and its Governance performance domain.
- Frequent AI use correlates negatively with critical thinking. Cognitive offloading weakens the judgment skills project managers need most (Gerlich, 2025).
- Human-in-the-loop means three things: verification before action, accountability stays with people, continuous quality monitoring.
- AI does not replace the project manager on any part of the complexity spectrum. But project managers who skip systematic oversight will be replaced by those who do not.
- Making this operational requires four decisions: define review levels, assign accountability, build verification into workflows, and monitor for drift.
This article is part of a series leading up to our upcoming Webcast, “5 Steps to Integrate AI into Your PPM Practices: A Tactical Blueprint”, on June 24, 2026. Register now to secure your spot.
https://www.iil.com/your-ai-advantage-practice-habit-strategy/
Markus Kopko is a strategic project and AI transformation expert with over 25 years of experience in project, program, and portfolio management. He contributes to the Core Development Team of the PMI Standard on AI in Project, Program, and Portfolio Management and served on the PMI Review Team for the PMBOK® Guide, 7th Edition. Markus holds PMP®, PgMP®, and PMI-CPMAI™ certifications and is a trainer and content creator for International Institute for Learning (IIL).
PMP®, PgMP®, PMBOK®, and PMI-CPMAI™ are marks of the Project Management Institute, Inc.
