AI Will Speed Up Parts of Product Development. The System May Not Keep Up.
AI is moving from isolated development tasks towards connected workflows and bounded agents. But digital work, physical development and organisational decision-making will not accelerate at the same speed.
In two recent Escape articles, I argued that organisations should not start with AI in new product development. They should start with the decision they are trying to improve.
I also argued that hybrid product development is bigger than combining Agile and Stage-Gate. It is an architecture connecting governance, learning, assurance, integration and industrialisation.
These arguments now converge.
The next question is not simply where AI can be added to the existing process. It is:
What happens to the product-development system when AI can continuously generate evidence, perform parts of the work and influence decisions?
AI can already accelerate analysis, software development, simulation, requirements work and documentation. But it cannot eliminate hardware lead times, physical testing, tooling, supplier readiness or organisational accountability.
The result is not one uniformly faster process.
It is a development system moving at different speeds.
That may become more important than the productivity gained from any individual AI tool.
AI will not accelerate the system evenly
We are not yet operating autonomous product-development organisations.
Recent studies suggest that most AI applications remain focused on individual activities, particularly in the earlier parts of development.
AI can help teams analyse customer and market information, search and structure technical knowledge, formulate requirements, generate concepts, support simulation, propose test cases and prepare decision material.
These applications can create real value. But they do not yet constitute an integrated AI-enabled development system.
A review of AI applications across product development found that most use remains concentrated within individual phases and design activities. End-to-end integration across the development lifecycle remains limited.
A separate review of 89 studies on AI in product-platform development reached a similar conclusion. AI was mainly used for automation and optimisation. Collaborative AI roles were uncommon, particularly across requirements, system architecture and multiple technical domains.
In other words:
We have many AI use cases. We do not yet have a mature AI-enabled product-development system.
People still carry the context between activities.
They connect customer needs to requirements, requirements to architecture, and design decisions to manufacturing, suppliers, quality and the business case.
AI may improve individual outputs.
The wider system remains largely human-coordinated.
From assistance to controlled action
Most current AI use follows a familiar pattern.
A person initiates a task. AI produces an analysis, draft or recommendation. The person assesses the output and decides what happens next.
AI will increasingly perform bounded actions within defined workflows.
Rather than only identifying an inconsistency, an AI agent may investigate related information, prepare a proposed change, request missing evidence, initiate a workflow or escalate an issue when an agreed threshold is reached.
It is useful to distinguish three levels.
AI assistant. AI helps a person search, analyse, generate, compare or recommend. The person initiates the task, validates the result and performs the action. This is where most organisations are today.
Bounded AI agent. AI performs several connected steps within a defined workflow. It operates within specified data access, permitted actions, verification rules, approval points and escalation thresholds. The organisation still owns the workflow, the decision and the outcome.
Development-system orchestration. Multiple agents, data sources and development systems become connected across customer insight, requirements, architecture, engineering, software, configuration management, verification, manufacturing preparation, suppliers, quality and commercial information.
This is where AI begins to change the development system rather than merely improve individual tasks.
It is also where the challenge becomes substantially harder.
An agent operating within one controlled workflow is very different from an AI system influencing trade-offs across product architecture, supplier maturity, safety, industrialisation and commercial commitments.
The technical integration becomes more difficult.
So does the governance.
Robert Cooper’s recent work offers one concrete practitioner view of this direction.
In his AI-Powered Stage-Gate concept, AI supports work inside stages and contributes analytical input at gates. His Stage-Gate Agentic proposal goes further by envisaging AI agents orchestrating substantial parts of a development stage.
This is not a proven industrial operating model. The literature cited here does not yet provide robust evidence that agent-orchestrated stages improve outcomes in complex industrial product development.
But Cooper’s work exposes the central governance question:
What happens when AI can produce and update development evidence faster than the organisation can understand and act on it?
Cooper and Xinjin Zhao also argue that gates should focus less on reaching a supposedly perfect decision and more on increasing learning velocity through conditional decisions, rapid experiments, explicit unresolved risks and reviews triggered by material changes.
That direction is plausible.
It also creates a risk.
Faster experimentation can become a perpetual-go mechanism if the organisation never steps back and asks whether the product remains strategically, commercially or technically worthwhile.
Faster learning does not make governance less important.
It changes what governance must do.
The bottleneck will move
AI can significantly reduce the time required for information retrieval, data analysis, requirements review, software development, simulation, design-space exploration, test preparation, scenario development, documentation and decision preparation.
But it does not remove hardware lead times, physical prototype builds, laboratory capacity, durability testing, tooling, material qualification, production trials, supplier readiness, regulatory approval, organisational disagreement or stakeholder commitment.
Digital and analytical workstreams may therefore accelerate much faster than the physical and organisational parts of the programme.
This creates a two-speed development system.
Software may generate new releases faster than hardware can be verified.
Design teams may explore more alternatives than manufacturing can assess.
Simulation may produce more evidence than engineering leaders can review.
Requirements may be revised faster than architecture, verification plans and supplier commitments can be updated.
More frequent design changes may increase pressure on configuration management, industrialisation and the supply chain.
Management may receive more recommendations than it can meaningfully challenge.
The bottleneck moves.
It may no longer be the production of information. It becomes the organisation’s ability to validate the information, connect it across disciplines, prioritise what matters, synchronise the consequences and implement the resulting decision.
This matters because local productivity can hide system-level deterioration.
An engineering team may complete more analysis while creating additional work for verification.
A software team may release faster while increasing integration risk.
A product team may generate more concepts while delaying the decision about which concept deserves scarce development capacity.
A programme may produce better dashboards while leadership remains unable to resolve the underlying trade-offs.
AI may compress the work.
It does not automatically compress the physical system, the organisation or the consequences of the decision.
A faster activity does not automatically create a faster development system.
This is why AI adoption cannot be evaluated only through task-level time savings.
The relevant measure is not simply whether one activity became faster. It is whether the total system became better able to learn, decide and deliver.
That requires attention to the interfaces between workstreams.
When one part of the system accelerates, leaders must ask what new demand is being placed on the next part.
More simulations create a need for stronger evidence selection.
More frequent software releases create a need for faster integration and verification.
More design alternatives create a need for clearer decision criteria.
More recommendations create a need for stronger prioritisation and decision discipline.
AI will not remove dependencies.
It may expose and intensify them.
Faster output is not the same as organisational capability
There is another important risk.
An AI-generated output is not necessarily equivalent to the human process that previously produced it.
A 2026 multiple-case study compared AI-generated sustainability criteria with criteria developed through facilitated company workshops.
AI produced broad and structured outputs quickly. But contextual relevance, prioritisation and legitimacy depended on company-specific information and stakeholder participation.
The strongest result came from combining AI-generated breadth with human insight and organisational context.
The lesson extends beyond sustainability.
A workshop does not only produce a document.
A design review does not only produce comments.
A planning process does not only produce a schedule.
These activities may also create shared understanding, constructive disagreement, organisational learning, ownership and commitment to the resulting decision.
Automating the deliverable may remove part of the process through which the organisation becomes capable of acting on it.
This does not mean that organisations should preserve every meeting or manual process.
Many should be simplified or removed.
But leaders need to distinguish between work that merely produces an output and work that also builds the capability required to use that output.
The relevant question is therefore not only:
Can AI produce this output?
It is also:
What understanding, alignment or judgement was created while people produced it?
What product leaders need to redesign
The immediate task is not to predict when autonomous product development will arrive.
It is to prepare the development system for greater AI involvement without losing control of the evidence, the product or the decisions.
Five design questions provide a practical starting point.
1. Where may AI advise, recommend or act?
These are different levels of delegation.
The boundary should depend on risk, reversibility and consequence.
An AI agent may be allowed to update documentation, initiate a review or propose a test. The threshold should be much higher before it can alter an approved requirement, change a supplier commitment or influence a safety-related release decision.
2. What counts as authoritative evidence?
An AI-generated answer should not become product evidence merely because it is clear and convincing.
The organisation must understand the source data, assumptions, uncertainty, traceability and verification required for different types of decisions.
3. Where will uneven acceleration create new bottlenecks?
Accelerating analysis, software or documentation may increase pressure on verification, manufacturing, suppliers, quality and leadership attention.
Every AI initiative should therefore be assessed across the end-to-end workflow, not only within the team receiving the tool.
4. How should governance respond to continuously changing evidence?
When evidence is updated more frequently, reviews may need to become more event-driven.
Material changes in assumptions, risks, test results, supplier maturity or the business case may justify a decision review before the next scheduled gate.
But event-driven governance still needs explicit stop criteria.
Otherwise faster learning can become continued activity without renewed strategic commitment.
5. Which capabilities and commitments must remain explicitly human?
AI may prepare a recommendation to start tooling, select a supplier, release a product or accept a residual risk.
It cannot carry the organisational accountability for that commitment.
Nor should organisations remove the activities through which engineers learn, functions align and leaders understand the evidence unless those capabilities are deliberately replaced.
The Escape take
The future of AI in product development is not primarily about giving every engineer an AI assistant.
It is about redesigning how trusted evidence, engineering work, assurance and decisions connect across the development system.
The organisations that benefit most will not necessarily be those with the largest number of AI tools.
They will be those that know:
where AI may generate evidence;
where it may recommend;
where it may act;
where it must be challenged;
who owns the resulting commitment.
Do not begin by automating the process you already have.
Begin by identifying where the development system will accelerate, where it will not, and which interfaces will become the new constraints.
AI can accelerate the development work. The leadership task is to redesign the system around faster learning without surrendering accountability.
Research references
Jonuschies, I., Siewert, L., Michaelsen, E. and Gericke, K. (2025). “Mapping AI Applications in Design: Support and Activities Across the Product Development Process.” Proceedings of the Design Society, 5, 2341–2350. DOI: 10.1017/pds.2025.10248.
Panarotto, M. and Cascini, G. (2026). “Artificial Co-Intelligence in Multi-Domain Platform Development: What Is Now and What Is Next?” Proceedings of the Design Society, 6, 83–92. DOI: 10.1017/pds.2026.10367.
Woofter, J., Schulte, J. and Watz, M. S. (2026). “AI-Assisted Leading Sustainability Criteria Development: A Multiple Case Study.” Proceedings of the Design Society, 6, 1671–1680. DOI: 10.1017/pds.2026.10525.
Practitioner and emerging perspectives
Cooper, R. G. (2025). Stage-Gate Agentic: The Coming Revolution in the New Product Process. PDMA Knowledge Hub.
Cooper, R. G. (2026). AI-Powered Stage-Gate: Supercharging Your Idea-to-Launch Process. PDMA Knowledge Hub.
Cooper, R. G. (2026). Stage-Gate: The “Official” 2026 Version. PDMA Knowledge Hub.
Cooper, R. G. and Zhao, X. (2026). New Product Governance in a Time-Compressed World. PDMA Knowledge Hub.
