AI Cannot Own the Decision. But It May Redesign the Decision Process.
What AI could change in project governance and product development — and where accountability must remain human.
Many project decisions are not delayed because the information does not exist.
They are delayed because it is dispersed across functions, based on different assumptions and assembled too late.
That is particularly true in complex product development.
Engineering, manufacturing, supply chain, quality and commercial teams may each hold valid parts of a decision. The difficulty is bringing those parts together before the organisation commits.
AI is already helping project teams prepare reports, summarise meetings, update risk registers and analyse schedules.
Useful.
But its more consequential role may be elsewhere: redesigning how decisions are prepared, challenged and coordinated.
The information is often fragmented
Consider a decision to freeze a product design.
Engineering may regard the design as sufficiently mature. Manufacturing may still see unresolved concerns about tolerances, testability or assembly. Supply chain may be relying on supplier capacity assumptions that have changed. Quality may be waiting for verification evidence. Commercial may have revised the expected volumes underpinning the original business case.
None of these functions is necessarily wrong.
Each holds part of the decision.
The problem is that the knowledge is distributed. Some of it is documented. Some sits in specialist systems. Some exists only in recent conversations, experience and professional judgement.
As a result, constraints emerge sequentially.
A manufacturing issue appears after engineering has recommended design freeze. A supplier limitation becomes visible after the schedule has been updated. A changed market assumption reaches the project after cost and capacity decisions have already been made.
Each new discovery triggers another analysis, another revision and often another alignment meeting.
This is one reason project boards spend too much time reconstructing the situation and too little time addressing the actual decision.
The board is not there to review a polished version of the project’s past. It is there to make decisions about its future.
From task automation to decision orchestration
Kris Johnson Ferreira and Jordan Tong describe a more advanced role for AI: supporting the orchestration of decisions across organisational silos.
Their argument is not that AI should make major enterprise decisions. It is that AI systems may help identify relevant questions, activate parallel analyses, connect outputs between functions and surface human knowledge that would otherwise arrive too late. They describe this as automating the orchestration of intelligence rather than automating the final decision.
That distinction matters for project management.
AI may move from helping the project manager prepare the meeting to helping the organisation prepare the decision that the meeting exists to make.
A related HBR article by Felipe Csaszar describes three ways AI can expand strategic decision-making:
searching a wider range of possible actions
creating richer representations of the situation
combining and challenging different perspectives.
The same three activities are central to decisions in product development.
The research does not yet demonstrate an end-to-end, AI-orchestrated NPD governance model. Ferreira and Tong explicitly describe the most advanced company initiatives as early-stage.
What follows is therefore a practical interpretation of where the evidence and emerging technology could take project governance — not a claim that this model is already established practice.
What could change at a design-freeze decision?
The traditional preparation for design freeze often revolves around a gate pack, a maturity assessment, a risk register and a recommendation from the project team.
These remain necessary. But AI could potentially improve three aspects of the process.
A wider option set
Project teams frequently narrow the solution space early.
Sometimes that is necessary. Capacity is limited, deadlines are real and not every alternative deserves detailed investigation.
But early convergence can also happen because the team lacks the time to develop and assess credible alternatives.
Before design freeze, AI could help the organisation examine options such as:
retaining the current design with additional verification
freezing only stable subsystems
accepting a temporary manufacturing constraint
changing a supplier or component
reducing initial scope
delaying the freeze to close a specific technical uncertainty.
Human experts would still determine which alternatives are feasible.
AI’s contribution would be to reduce the risk that the organisation selects from an unnecessarily narrow option set.
A more integrated view
Most project governance relies on a collection of snapshots:
the current milestone plan
technical maturity
unresolved requirements
cost development
supplier readiness
manufacturing readiness
quality evidence
commercial assumptions.
Each view may be accurate within its own boundaries. The decision, however, depends on the relationships between them.
A product can appear technically mature while remaining industrially immature. A supplier may meet its contractual milestone but still create unacceptable programme exposure. A schedule recovery may protect the launch date while transferring risk into validation or ramp-up.
AI could help connect these views and identify inconsistencies.
It might show that the engineering recommendation assumes a supplier date that no longer matches the procurement forecast. It could identify that the manufacturing assessment is based on an earlier design revision. It could flag that the business case uses volume assumptions that have subsequently changed.
This would not create an objective answer. The evidence could still be incomplete, outdated or misleading.
But it could expose the assumptions behind the recommendation before the project reaches the decision table.
More systematic challenge
Good governance requires more than compiling evidence in favour of the preferred option.
It requires challenge.
AI could review a design-freeze recommendation from several perspectives:
What would a manufacturing reviewer question?
Which technical assumptions remain weakly supported?
What would have to be true for the proposed schedule to work?
Which risks are being transferred rather than resolved?
What could make this decision look wrong three months from now?
Which alternatives were dismissed without sufficient evidence?
Csaszar refers to a field experiment involving 776 commercial and R&D professionals working on innovation tasks. AI-supported individuals matched the average quality of two-person teams working without AI, and AI-supported teams were faster and more likely to produce top-decile solutions. The researchers also found that AI reduced some of the silo effect between commercial and R&D participants.
That does not prove that AI can govern a complex development programme.
It does suggest that AI may help people work across functional perspectives and subject proposals to broader challenge.
The purpose is not to create an artificial expert whose answer replaces the people in the room.
It is to ensure that the recommendation has been examined more rigorously before those people are asked to approve it.
The project manager’s value shifts
As information gathering and document preparation become easier, the project manager’s value cannot be defined primarily by producing plans, compiling reports and chasing updates.
The stronger role is to design the decision process.
That means asking:
What decision is actually required?
Which alternatives should be considered?
What evidence is necessary?
Which functions hold relevant knowledge?
Which assumptions must be made explicit?
How should the recommendation be challenged?
What uncertainty remains?
Who has the authority to decide?
Who will own the consequences?
These have always been project-governance questions.
AI makes them more visible because the quality of its contribution will depend on the quality of the decision architecture around it.
A poorly defined decision does not become a good decision because AI can process more data.
It may simply become a poorly defined decision produced faster.
The role of the project manager therefore shifts from being a producer of information to being an architect of decision quality.
AI cannot own the consequences
There is also a governance risk.
An HBR research roundup describes an experiment in which managers reviewed documents containing deliberate errors. Among managers already working in organisations that listed AI agents as employees on organisation charts, presenting the source as an “AI employee” rather than an AI tool led them to identify fewer errors, escalate more issues and shift more responsibility away from themselves.
The strongest effect appeared within a specific subgroup, so the result should not be generalised too far.
But the warning is relevant.
How an organisation frames AI can affect how people understand their own responsibility.
An AI agent may be given access rights, operating limits and permission to initiate defined actions.
It cannot be given organisational accountability.
In project and product governance, the boundary should remain explicit:
AI may retrieve, compare, connect, identify, challenge, simulate and recommend.
Humans must prioritise, accept risk, commit resources and approve the decision.
A system may prepare the decision record. It cannot stand in front of the customer, board or organisation when the consequences emerge.
Start with one decision
The practical starting point is not an enterprise-wide AI rollout.
Choose one recurring decision where coordination is difficult.
Design freeze is one possibility. Others include gate progression, change approval, supplier readiness, technical-risk acceptance or launch commitment.
Map how the decision works today:
Where does relevant information arrive late?
Which assumptions remain hidden?
Where do functions use different definitions?
Which questions are repeatedly discovered during the meeting?
Which elements require contextual or expert judgement?
Who has the authority to decide?
Then identify where AI could strengthen the process without obscuring accountability.
The success measures should extend beyond time saved:
Were important constraints identified earlier?
Were more credible alternatives considered?
Were conflicting assumptions made visible?
Was less rework required?
Were fewer repeated alignment cycles needed?
Were the conditions attached to the decision clearer?
Did the project experience fewer late surprises?
That is a more meaningful test than counting the number of reports generated automatically.
The Escape take
The next meaningful use of AI in project management may not be another automated status report.
It may be a better-prepared decision.
In complex product development, AI could help connect evidence across functions, expose assumptions, widen the available options and challenge the recommendation before it reaches the project board.
It can also be incomplete, misleading or wrong.
AI may redesign how the organisation reaches a decision.
It cannot accept the risk, make the commitment or own the consequences.
That remains a management responsibility.
References
Csaszar, F. A. (2026). AI Is Revolutionizing Strategic Decision-Making. Harvard Business Review, September–October 2026.
Johnson Ferreira, K., & Tong, J. (2026). How AI Agents Orchestrate Work Across Silos. Harvard Business Review, September–October 2026.
Szpilman, D. (2026). Research Roundup: AI on Org Charts, Job-Hoppers’ Skillsets, Autorenewal Subscriptions, and More. Harvard Business Review, September–October 2026.
