Organisations introducing AI into project delivery usually expect the difficulty to be technical. Which tools, which integrations, which data. Those questions do need answering, but they are rarely what stalls adoption. When AI assistance underdelivers in a project environment, the cause is more often that nobody stated what the AI was supposed to achieve, nobody owns the risk it introduces, and nobody can evidence whether it helped.
That is a governance failure rather than a technology failure, and it is worth naming as such, because the two require different responses. A technology problem is solved by procurement or configuration. A governance problem is solved by deciding who is accountable, what will be recorded, and when it will be reviewed.
The Nine Challenges
Working with project teams and PMOs, nine obstacles recur. None requires specialist technical knowledge to recognise and address.
- Resistance to change. People are often the greatest barrier, not the technology. PMI research indicates only around 20% of project managers report extensive or good practical experience with AI tools.
- Over-reliance on AI. The more reliable AI becomes, the less carefully teams check it. Careful review gives way to a quick glance, and errors that would once have been caught pass unnoticed.
- Data quality and availability. Project data is scattered across spreadsheets, tools and meeting notes. Fragmented input produces fragmented assistance.
- Tool integration challenges. Adopting on the basis of market interest rather than defined business need wastes resources, particularly where IT and Information Security are brought in after the decision rather than before it.
- Lack of clear objectives. AI is introduced because the technology is available. Without a stated objective, the value of the AI assistance cannot be demonstrated or improved.
- Accountability and transparency. Where a team cannot explain how a recommendation was reached, the decisions that follow are hard to justify to a sponsor or an auditor.
- Unapproved AI use (shadow AI). A 2026 Wakefield Research survey found two-thirds of office professionals had used AI at work knowing it was not permitted, because the unapproved tools were better than the approved ones.
- Ethical considerations and bias. AI learns from data, and where that data carries historical bias, the AI reproduces it. Fairness, privacy and human oversight all need deliberate consideration.
- Keeping pace with AI regulation. A project governed to today’s rules can be non-compliant before it reaches closure. EU AI Act obligations are phasing in, ISO/IEC 42001 is maturing, and sector regulators are issuing their own expectations.
What the Nine Challenges Have in Common
Read together, the pattern is consistent. Every one of the nine challenges resolves into a governance question that has not been answered: what is this AI assistance for, who owns the associated risk, what evidence will exist that a human reviewed the output, what has been disclosed to whom, and which obligations apply.
Answer those questions and most of the nine challenges become manageable within existing project governance. Leave them unanswered and the challenges compound. The organisation ends up carrying risk it has never assessed, while claiming benefits it cannot evidence.
Shadow AI illustrates the point. Prohibition is the intuitive response, and it addresses none of the underlying cause. People are reaching for unapproved tools because the approved ones do not meet the need, or because they are unaware that the approved ones already do. Both causes are governance matters, and neither is fixed by a stronger policy statement.
Where the Fixes Sit
Some of the fixes belong inside a single project and can be started immediately: documenting AI assistance objectives during planning, recording who reviewed each AI-assisted deliverable, naming the person accountable for the output. Others are organisational and belong with the PMO, IT, Information Security or whoever owns AI policy: data readiness, tool evaluation routes, regulatory monitoring. Identifying the right owner before committing to an action is usually the difference between a fix that holds and one that is quietly abandoned.
