For decades, organisations have relied on established project management methodologies to deliver technology transformations. Waterfall has provided the structure and predictability required for complex programmes, while Agile has introduced greater flexibility through iterative development and continuous feedback. In practice, most large organisations have progressively adopted a combination of both, adjusting their delivery approaches according to the nature of their projects.
The emergence of Agentic AI is challenging some of the assumptions on which these methodologies were built. Unlike traditional software applications, which execute predefined instructions, AI agents can interpret objectives, make decisions, interact with multiple systems and adapt their actions according to the context in which they operate. This introduces a level of behavioural uncertainty that conventional project management frameworks were not originally designed to accommodate.
As companies move beyond AI experimentation towards deploying agents within core business processes, the discussion around delivery methodologies deserves renewed attention. The issue is not necessarily whether Waterfall or Agile should prevail, but how organisations should adapt their existing practices to manage systems whose behaviour cannot always be fully anticipated during the design phase.
The limitations of conventional delivery approaches
Traditional software development generally starts from the assumption that system behaviour can be described through functional requirements and validated against predefined acceptance criteria. Even in complex environments, the objective is to translate business needs into specifications that guide development and testing.
Agentic AI introduces a different dynamic. Consider an agent responsible for processing supplier invoices, identifying discrepancies and interacting with an ERP system. While the expected business outcome can be clearly defined, the sequence of decisions and actions taken by the agent may vary depending on the information available, the exceptions encountered and the tools it chooses to use.
This distinction has implications for the entire development lifecycle. A sequential approach, where requirements are extensively documented upfront and testing is concentrated towards the end of the project, becomes difficult to apply when the behaviour of the solution must be progressively understood through experimentation.
Agile development provides a more suitable environment for this type of uncertainty, allowing teams to test different configurations, assess performance and refine the solution through successive iterations. However, the flexibility offered by Agile does not remove the need for structured governance, particularly when agents operate within regulated environments or interact with sensitive information and critical business applications.
Organisations therefore face a delivery challenge that combines the need for experimentation with requirements for predictability, accountability and operational control. These objectives are not necessarily contradictory, but they require a more deliberate integration of engineering practices and project governance.
What academic research suggests
Recent academic research provides useful perspectives on how development methodologies may need to evolve.
A 2025 study from the University of York, Agile Development for Safety Assurance of Machine Learning in Autonomous Systems (AgileAMLAS), examines how iterative development practices can be combined with structured safety assurance. Although the research focuses on machine-learning-based autonomous systems rather than enterprise LLM agents specifically, it addresses a comparable challenge: maintaining the flexibility required to develop intelligent systems while ensuring that their behaviour remains within acceptable boundaries.
The study provides a useful foundation for considering hybrid delivery approaches in which experimentation takes place within a defined governance framework, rather than being treated as an alternative to formal assurance.
A second publication, Evaluation and Benchmarking of LLM Agents: A Survey, presented at ACM SIGKDD in 2025, examines the challenges associated with assessing the performance and reliability of AI agents. The research highlights the importance of evaluating agent behaviour across multiple dimensions, extending beyond the correctness of individual outputs.
This has practical implications for enterprise projects. Traditional software testing primarily verifies whether a system produces the expected results when presented with specific inputs. In Agentic AI environments, assurance must also consider the decisions taken during execution, the tools accessed, the consistency of behaviour and the circumstances under which human intervention becomes necessary.
An agent may successfully complete a business task while following a sequence of actions that introduces unacceptable operational or security risks. Conversely, refusing to execute a particular instruction may represent the correct behaviour when predefined conditions have not been satisfied.
The evaluation of agentic systems consequently requires a broader understanding of quality, incorporating behavioural reliability, traceability and adherence to operational constraints alongside conventional functional testing.
Towards a more pragmatic hybrid delivery model
Rather than introducing an entirely new project management methodology, organisations would benefit from adapting the frameworks they already use, with a clearer distinction between the activities that require structured governance and those that benefit from iterative experimentation.
The initial stages of an Agentic AI programme should establish the business objectives, target operating model, architectural principles and boundaries within which agents will operate. Decisions concerning data access, security, autonomy and accountability need to be addressed early, particularly when the intended solution will interact with critical business systems.
Development can then follow an iterative approach, allowing teams to experiment with models, tools and agent configurations while progressively evaluating their performance against representative business scenarios. Unlike conventional Agile delivery, where progress is frequently measured through completed functionality, the maturity of an agentic solution should also be assessed through evidence of reliability, consistency and operational effectiveness.
Formal production readiness assessments would remain necessary, although their emphasis would evolve towards demonstrating that agents can operate within acceptable risk thresholds and deliver the expected level of performance under realistic conditions.
The same principle extends beyond deployment. As models, data sources and business environments evolve, agent behaviour may change, requiring continuous monitoring and periodic reassessment. The distinction between project delivery and ongoing operational assurance therefore becomes less pronounced than in traditional software implementations.
Such an approach would preserve the governance discipline associated with Waterfall while incorporating the experimentation and continuous improvement principles of Agile. Its effectiveness would depend less on introducing additional processes than on ensuring that governance activities are proportionate to the risks and uncertainties involved.
The operating model remains the real challenge
Beyond the methodological discussion, I believe there is a more fundamental issue that organisations risk overlooking as they accelerate their investments in Agentic AI.
Much of the current attention remains focused on the technical capabilities of agents and their potential to automate activities previously performed by employees. While these capabilities are progressing rapidly, the relationship between task automation and economic value is considerably more complex than it may initially appear.
An organisation that successfully automates 40% of the activities performed by an operational team will not necessarily achieve a comparable reduction in operating costs. The financial outcome depends on how responsibilities are redistributed, whether existing processes are redesigned and how effectively the organisation uses the capacity released by automation.
Without corresponding changes to the operating model, the introduction of AI agents may improve individual productivity without materially affecting the overall cost structure. Employees may spend less time performing certain activities while the organisation continues to operate with essentially the same processes, management structures and allocation of resources.
This is a familiar challenge in business transformation, although Agentic AI makes it particularly relevant because of the speed at which technical capabilities can now be introduced.
The business case for an Agentic AI programme should therefore extend beyond estimates of time saved or activities automated. It needs to explain how the organisation intends to convert those improvements into measurable operational or financial outcomes, whether through increased capacity, reduced costs, improved service quality or the ability to support additional business volumes.
In my view, this requires operating model transformation to be considered an integral component of Agentic AI delivery rather than a separate activity addressed once the technology has been implemented.
Implications for project leadership and PMOs
These developments have direct implications for the way transformation programmes are governed.
Project managers and PMOs have traditionally focused on controlling delivery through budgets, schedules, milestones, dependencies and risk registers. These disciplines will remain important, but they provide only a partial view of progress when the solution being delivered is an autonomous system whose performance continues to evolve.
The governance of Agentic AI programmes will increasingly require a combination of conventional project indicators and measures related to agent performance, operational reliability, human intervention and business value realisation.
This also places greater importance on collaboration between technology, operations, risk management and business leadership. Decisions about the level of autonomy granted to an agent, the conditions under which human approval is required and the acceptable trade-offs between efficiency and control cannot be made exclusively within technology teams.
For PMOs, the opportunity is to develop a more integrated approach to transformation governance, connecting technical delivery with operational readiness and the realisation of business benefits.
Such an evolution would also help address a longstanding weakness in technology programmes: the tendency to consider a project successful once the solution has been delivered, even when the anticipated operational improvements have not yet materialised.
Agentic AI makes this distinction more significant because technical functionality and business performance may diverge considerably during the early stages of deployment.
An evolution rather than a replacement
The emergence of Agentic AI does not make established project management methodologies obsolete. Waterfall and Agile continue to provide valuable disciplines, and their relevance will depend on the characteristics of each programme.
What changes is the need to recognise that autonomous systems introduce forms of uncertainty that cannot be addressed solely through traditional requirements management or iterative software development.
A more appropriate approach is likely to combine structured governance, experimental development, continuous behavioural evaluation and explicit accountability for business outcomes. Although academic research is beginning to provide useful foundations, empirical evidence comparing delivery methodologies specifically for enterprise Agentic AI remains limited.
Organisations should therefore approach the evolution of their delivery frameworks pragmatically, adapting established practices while building experience through actual implementations.
Ultimately, the success of Agentic AI programmes will depend on more than the ability to develop and deploy technically capable agents. It will require organisations to reconsider how work is organised, how decisions are governed and how the value created by autonomous systems is translated into sustainable business performance.
The development of AI agents may prove to be the more manageable part of this transformation. Integrating them effectively into the operating model, while maintaining control and delivering measurable economic benefits, is likely to represent the greater management challenge.
References
Hodge, V. J., & Osborne, M. (2025). Agile Development for Safety Assurance of Machine Learning in Autonomous Systems (AgileAMLAS). Array, 27, 100482. https://doi.org/10.1016/j.array.2025.100482
Mohammadi, M., et al. (2025). Evaluation and Benchmarking of LLM Agents: A Survey. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. https://doi.org/10.1145/3711896.3736570