AI safety in projects: What every project professional needs to know

AI adoption across organisations is accelerating. APM’s own research found that 70% of project professionals now say their organisation uses AI. Yet around 80% of AI projects fail to meet their intended objectives. Organisations are adopting faster than they are governing. Project professionals are directly in the path of that gap.

The regulatory landscape

The governance framework around AI is developing rapidly and has direct implications for project delivery. The EU AI Act, the most comprehensive AI regulatory framework currently in force globally, has been applying in stages since February 2025 and high-risk system obligations apply from August 2026. UK organisations with EU operations, EU clients or EU supply chains are directly in scope.

The UK Data (Use and Access) Act 2025 introduces a statutory standard of ‘meaningful human involvement’ for AI-assisted decisions that produce legally or similarly significant effects on individuals. This is not a soft aspiration. It is a statutory duty. A board member who approves an AI recommendation without reviewing the underlying evidence has not provided meaningful human involvement, regardless of what the governance record says.

The agentic AI challenge

Agentic AI systems, capable of planning and executing multi-step tasks with limited human intervention, have become one of the fastest-growing areas of enterprise AI investment. Governance frameworks, however, remain comparatively immature. Gartner predicts that more than 40% of agentic AI projects will be cancelled by 2027 due to escalating costs, unclear business value or inadequate risk controls. For project professionals, the challenge is not simply deploying autonomous systems, but ensuring that accountability, oversight and human judgement remain effective as decision-making becomes increasingly distributed between humans and machines.

In project environments, an agentic AI system that autonomously reschedules resources, sends stakeholder communications or updates risk assessments without human review is not merely an analytical tool,  it is an autonomous actor in your delivery environment. The governance question is whether your accountability framework and oversight architecture are designed for a world in which the system can act before the project manager has reviewed its output.

Lifecycle governance

AI governance is not a procurement checklist. It is a lifecycle responsibility. The risks associated with AI systems in project environments do not stop at deployment. AI systems drift over time as underlying data patterns change. Vendors update models without always notifying deploying organisations. Systems that performed reliably in test conditions may behave differently under live operational loads.

Project professionals who deploy AI tools in governance environments need ongoing monitoring processes, not one-time impact assessments. The DUAA’s meaningful human involvement requirement applies throughout the life of the system, not just at the point of procurement.

Meaningful human involvement in practice

The most consequential practical change the regulatory framework introduces is the standard of meaningful human involvement. This means a competent human must actively review AI outputs and exercise genuine independent judgement before decisions with significant effects are made. A rubber stamp is not meaningful involvement. A governance record that shows human sign-off without documenting the basis for that review will not satisfy the statutory standard.

For most project teams, this requires a change to governance documentation, not just to AI tools. The question ‘how did we reach this decision?’ must be answerable in terms that demonstrate genuine human judgement, not just human presence.

Practice checklist

• Map your AI regulatory exposure: Identify which EU AI Act risk tiers and which DUAA obligations apply to the AI systems your programme is using or planning to use.
• Apply the meaningful involvement test: For every significant AI-assisted decision: can you document the basis on which a competent human reviewed the AI output and exercised independent judgement? If not, the governance standard has not been met.
• Govern agentic AI separately: Any AI system that can take autonomous action in your delivery environment needs specific governance provisions, including defined scope, human override capability and audit trails.
• Build lifecycle review into your governance framework: Schedule regular reviews of AI tool performance and vendor updates. AI governance is a continuous process, not a one-time assessment.
• Update your governance documentation: Your project governance framework should be updated to reflect the statutory requirements introduced by the DUAA 2025. Do this before, not after, a governance challenge. 

 

You may also be interested in:

• What is artificial intelligence in project management?
• Visit APM bookshop: Directing Change: A Guide to Governance of Project Management, 3rd edition
• Join the APM AI and Data Analytics Interest Network for more conversations on AI and data