AI, Ethics and Accountability: Who Owns the Decision?
Six weeks after a supplier was appointed, it emerged that the recommendation had rested on a comparison which quietly dropped a qualification buried in one of the bids. The comparison had been produced with an AI assistant, and nobody had concealed that. The awkward part was that no single person felt they had made a decision. The assistant produced the analysis, the project manager tidied the presentation, the evaluation panel approved the conclusion, and the sponsor signed a paper that read as though the underlying work had been done.
Accountability had not moved. It sat exactly where it had always sat, with the people holding the authority to appoint a supplier. What had moved was the point at which anyone last looked properly at the source material, and nobody could now say where that point was. That is the real problem with AI in project decisions, and it is a practical problem before it is an ethical one.
Contributing to a decision is not making one
AI systems produce outputs: summaries, drafts, estimates, risk scores, ranked options, forecasts. A decision is an act of commitment made by a person who has the authority to make it and who will answer for the consequences. The confusion sets in because a well-formatted output looks like a conclusion, and accepting a conclusion feels nothing like choosing one.
Appendix X3 of the PMBOK® Guide treats artificial intelligence as an ordinary project consideration with a set of attached responsibilities: the suitability of the use case, data quality and permission, confidentiality, bias, transparency, validation of what comes back, and human oversight of what is done with it. None of that reads as a case for treating a system as a participant in governance. A tool can inform a decision and can carry none of the accountability for it.
Holding that line in daily work means keeping two things separate in your own head. What was produced is a piece of information with a provenance and a reliability. What was decided is a commitment made by a named person at a known moment. When those two collapse into each other, the phrase that starts appearing in reviews is "the system flagged it", which is a description of an input being used as though it were an authority.
How ownership drifts
Very few project managers decide to hand judgement to a tool. Ownership drifts through small acceptances that each look reasonable. An output arrives fluent, complete and confident, with none of the hedging a colleague would have added. It arrives quickly, often under time pressure, and it arrives without visible workings, so the natural check is whether it sounds right. Reviewers further along see only the conclusion, and every layer of the review assumes that the layer beneath it did the reading.
The second route is quieter still, because nobody chose to use AI at all. Scheduling, procurement, resourcing and reporting tools increasingly generate suggestions inside the workflow: a predicted completion date, a recommended sequence, a supplier flagged as higher risk, a set of tasks the system thinks are at risk of slipping. If you cannot say whether a figure in front of you came from a person, a formula or a model, you cannot say what has been checked, and neither can anyone who relies on your report.
Both routes end in the same place. The authority to decide remains formally clear while the basis for the decision becomes unreconstructable. Six months later, during a dispute, an audit or a lessons session, the honest answer to "what was this based on?" is that nobody knows.
What makes an AI-assisted decision defensible
The answer is not to check everything, which would remove the reason for using the tool and would fail on any project large enough to need it. The useful discipline is proportionate: identify what the decision actually turns on, verify that against the source, and record that you did.
A facilities programme replacing mechanical plant across eleven buildings offers a workable illustration. Condition surveys ran to several hundred pages, so the project manager used an AI tool to summarise each one and produce a priority order for the replacement programme. The order looked sensible and the steering group approved it. One building sat ninth because its survey described a boiler failure as intermittent, and the summary reproduced that wording without the surveyor's note that the intermittent failures had become weekly during the previous winter. The boiler failed properly in January, a teaching space was decanted for five weeks, and the programme lost its contingency.
The defensible version of that decision was never to read all eleven surveys in full. It was to read the two or three sitting either side of the funding cut line, because those were the only ones where being wrong changed anything, and to note in the paper that the ranking came from an AI-assisted summary of the surveys with those specific entries verified against the original text. That note costs a line. It tells the steering group what kind of evidence it is approving, which is the difference between a group that has accepted a recommendation and a group that has been presented with an answer.
When the analysis is someone else's
Accountability gets harder to talk about when the AI-assisted work arrives from outside: a supplier's report, a consultant's model, a partner's forecast. You cannot inspect their process, and pretending otherwise helps nobody. What you can do is ask what the output was built from, what the supplier verified, what its known limits are, and whether the answer to those questions is written down anywhere. An organisation that cannot describe how its analysis was produced has told you something useful about how much weight it will take.
Ethics enters at the same practical level. The confidentiality question is what data goes into the system, particularly personal, commercially sensitive or contractually restricted material, and whether anyone has actually confirmed that it is permitted. The fairness question applies wherever an output affects people: resource allocation, performance discussion, the exclusion of a supplier from a shortlist. The transparency question is whether the stakeholders relying on your recommendation know how it was produced, because a decision they would have questioned differently had they known is a decision taken on an incomplete picture. These are answerable questions with names attached, not a values statement appended to a project plan.
On a live project, the working test is short enough to use in the moment. Name the decision, name the person who owns it, and be able to say what was checked before it was taken. If any of the three is missing, the gap is worth closing before the output travels any further, because it becomes considerably harder to close after a decision has been acted on.
For a PMP® candidate, the position is more specific than it might appear. The current Examination Content Outline names artificial intelligence as one of the emerging trends used to test the relevance of the exam's tasks, and no task or enabler within it names AI. Technology appears as an example of an external business-environment change to be assessed for its impact, and ethics appears within the work of establishing project governance. The judgement being assessed is therefore the ordinary one: who holds authority, what information a decision rests on, and what the project manager does when the basis for a recommendation is unclear.
That judgement is easier to practise in structured discussion than alone, which is one of the things our PMP® Exam Preparation course is built around, since reasoning through unfamiliar situations improves with challenge in a way that reading definitions does not.
The technology will keep moving, and most project managers will use more of it next year than this year. The thing that does not move is the answer to who owns the decision, which stays with the person whose signature is on it, and who should be able to say, without hesitation, what they looked at first.
Interested in going further?
Deciding how much verification an AI-assisted output needs, and being able to defend that choice to a sponsor or an auditor, is the kind of judgement that benefits from working through realistic situations with experienced practitioners. Omega's PMP® Exam Preparation covers governance, decision authority and responsible use of project information alongside the rest of the syllabus.
The PMBOK® Guide Eighth Edition sets out its artificial intelligence considerations in Appendix X3, alongside the governance and risk material that accountability questions usually connect to.
Ad · Amazon affiliate link.
References
A049: Artificial Intelligence in Project Management: Practical Uses, Risks and Limits
A053: AI Governance for Project Managers
A050: AI and the PMP Exam: What Project Managers Need to Know
A051: Do You Need to Know Specific AI Tools for PMP?
A052: AI Governance for Project Managers: What to Check Before You Use It
PMP and PMBOK are registered marks of the Project Management Institute, Inc.
