AI Project Management – Adapting Project Delivery for AI
- Feb 20
- 7 min read
1. Insight
Artificial Intelligence delivery at project level behaves differently from traditional technology or business change delivery.
While AI programmes are governed through established programme management disciplines, the actual delivery of AI solutions takes place through projects and delivery teams operating in conditions of higher uncertainty, iteration and learning.
Project Managers, delivery leads and technical teams are therefore required to adapt how delivery is structured and managed in practice.
The fundamentals of good project management still apply. However, AI introduces characteristics that need to be understood and taken into account for successful delivery.
This Insight is intended to help experienced Project Managers understand how AI delivery works in practice, what is different, and how to adapt their approach with confidence.
Within the Orr Consulting AI Transformation Process, AI Project Management sits within the Deliver stage and operates within the structure and governance established by AI Programme Management.
2. Why This Matters
AI delivery often feels more complex and less predictable than traditional digital or business transformation delivery.
This is because AI delivery is not purely execution-driven. It is exploratory, iterative, dependent on data, sensitive to feasibility constraints and shaped by continuous learning.
In many AI projects, it is not known in advance whether the proposed solution will work as intended, whether the available data and capability are sufficient, or whether the resulting outputs will translate into reliable, usable and valuable outcomes.
While Agile approaches are commonly used to structure delivery, they do not fully describe how AI work behaves in practice.
AI delivery typically combines:
Agile delivery structures such as sprints, backlogs and coordination
Data science and machine learning lifecycles
Iterative experimentation, evaluation and validation
As a result, progress is achieved not by executing a fully defined plan, but by progressively reducing uncertainty and refining the solution.
AI project delivery is not linear. It progresses through iterative cycles of discovery, design and validation.
3. What This Means
AI does not require a new project management discipline.
It requires Project Managers to apply existing discipline in a delivery environment where work is less defined upfront and evolves through iteration.
There are four realities of AI project delivery that need to be understood, and four corresponding implications.
Reality 1 – Requirements Emerge
In traditional delivery, projects are based on defined requirements and known solutions.
In AI delivery, requirements are often hypotheses and the solution is not fully known at the outset.
Projects begin with questions, not complete answers.
What This Means for Project Delivery:
Projects should be structured around:
Clearly defined problems and objectives
Hypotheses to be tested
Success criteria that can be evaluated
Rather than attempting to define the full solution upfront, delivery should focus on progressively refining understanding.
Reality 2 – Delivery Is Explorative
AI delivery does not follow a linear path from design to build to deploy.
Instead, it progresses through cycles of exploration, testing and refinement.
Progress is achieved through experimentation, not execution alone.
These experiments test whether assumptions about the problem, the data, the capability, the delivery approach and the proposed solution hold true in practice.
What This Means for Project Delivery:
Projects should be structured around short, iterative cycles that combine:
Discovery
Design
Delivery
Within each cycle, teams:
Test assumptions
Build and evaluate solutions
Learn from results
Refine the next iteration
Reality 3 – Data Drives Feasibility
In traditional projects, delivery progress is often driven by system development.
In AI projects, progress is heavily dependent on:
Data availability
Data quality
Technical feasibility
These factors are often not fully understood at the outset.
Feasibility is discovered during delivery, not assumed before it. In some cases, this may reveal that the data, capability or delivery conditions are not sufficient to support the intended AI use case.
What This Means for Project Delivery:
Projects must:
Prioritise early data exploration and validation
Test feasibility before committing to full solutions
Continuously reassess assumptions as delivery progresses
Reality 4 – Adoption Is Uncertain
A technically successful AI model does not automatically result in a usable or valuable outcome.
Challenges often arise in:
Usability of outputs
Integration into business processes
User trust and adoption
A working model is not the same as a working solution.
What This Means for Project Delivery:
Projects must:
Involve users early and throughout delivery
Test outputs in real-world contexts
Design for usability and trust
Integrate business change into delivery
4. AI Delivery Modes
AI delivery does not occur in a single uniform way. In practice, organisations typically encounter three distinct modes of AI delivery, each with different characteristics, risks and delivery approaches.
The nature of uncertainty differs across these modes. In Build, uncertainty is often centred on whether meaningful patterns exist in the data and whether a reliable solution can be developed. In Configure, uncertainty is more often about capability fit, integration and workflow alignment. In Use, uncertainty is usually centred on governance, user behaviour, adoption and safe usage.
4.1 Build – Custom Development
This mode involves creating bespoke AI solutions designed around organisational data, workflows and requirements. It typically carries the highest uncertainty because feasibility, performance and value often need to be discovered through delivery.
Custom AI development may include classical machine learning, retrieval-based generative AI solutions, orchestration of foundation models, or other bespoke AI applications built around organisational data and workflows. It is:
Model, application and data-led
Experimental and iterative
Higher uncertainty
Profile:
Longer timescales
Higher technical delivery risk
Feasibility not guaranteed
Delivery Focus:
Iterative Discover–Design–Deliver cycles
Early feasibility validation
Managing uncertainty
4.2 Configure – Vendor Solutions
This mode involves implementing AI capabilities that already exist within established vendor platforms or enterprise systems. The primary challenge is usually not building the AI itself, but configuring it appropriately, integrating it effectively and ensuring expectations are realistic. It is:
Platform-led
Configuration and integration
Vendor-defined capabilities
Profile:
Moderate timescales
Medium delivery risk
Delivery Focus:
Validating capability against use cases
Managing integration
Aligning expectations with reality
4.3 Use – Generative AI Adoption
This mode involves enabling people to use generative AI tools directly in day-to-day work. It usually requires less technical build effort, but creates greater emphasis on governance, guidance, behaviour and safe usage at scale. It is:
User-led
Rapid adoption
Minimal technical implementation
Profile:
Shorter timescales
Low technical delivery risk
High governance and behavioural risk
Delivery Focus:
Governance and controls
User guidance and training
Monitoring and safe usage
The first step in AI project delivery is identifying the mode of delivery. Each mode requires a different approach to planning, risk management and execution.
5. The AI Delivery Cycle
Once the mode of delivery is understood, the next question is how delivery should be structured in practice. Across all three modes, delivery follows a form of Discover–Design–Deliver, but the depth and emphasis of each stage varies.
The following outlines how the cycle applies across each mode.
5.1 Build – Custom Development
In custom AI development, Discover–Design–Deliver is iterative and forms the core delivery mechanism.
Discover:
Clarify the problem or hypothesis
Explore data and constraints
Define success criteria
Design:
Select approach or model
Design experiments or solution options
Plan how to test and evaluate
Deliver:
Build and test the solution
Evaluate outputs against success criteria
Capture learning and insights
This cycle is repeated multiple times.
5.2 Configure – Vendor Solutions
In vendor-led AI delivery, the Discover–Design–Deliver cycle is applied in a more structured and predictable way.
Discover:
Clarify business requirements and use cases
Assess vendor capability fit
Identify integration points and constraints
Design:
Define configuration approach
Design integrations and workflows
Align solution with business processes
Deliver:
Configure the solution
Integrate with existing systems
Test functionality and performance
5.3 Use – Generative AI Adoption
In generative AI adoption, the cycle shifts away from technical delivery towards enablement, governance and behavioural change.
Discover:
Identify high-value use cases
Understand risks and constraints
Assess suitability for generative AI
Design:
Define usage guidance and guardrails
Design prompts, templates and workflows
Establish governance and assurance controls
Deliver:
Enable access to tools
Train and support users
Monitor usage and manage risks
5.4 The Full Technical Lifecycle
AI project delivery also needs to be understood within the context of the full technical lifecycle. For custom AI development in particular, delivery does not end at initial deployment. Ongoing monitoring, evaluation, tuning, version control, operational support and, where relevant, retraining or reconfiguration are essential to ensuring that solutions continue to perform safely, effectively and reliably in live use.
6. Benefits
Effective AI project delivery is characterised by:
Evidence-based decision points that allow use cases to progress, pivot or stop based on what is learned
Structured but flexible delivery approaches
Iterative cycles that reduce uncertainty
Early and continuous validation of data and feasibility
Close collaboration between technical and business teams
Strong focus on usability, adoption and value
Project Managers play a key role in:
Maintaining structure and control
Enabling iteration and learning
Ensuring alignment between delivery and outcomes
7. Risks
Projects that apply traditional delivery approaches without adapting to AI realities often experience:
Over-defined plans based on weak assumptions
Late discovery of data or feasibility issues
Rework and delays due to unrealistic expectations
Poor alignment between delivery approach, solution capability and business need
Technically successful outputs that are not usable or not adopted
Weak governance or uncontrolled usage in generative AI deployments
These are not simply delivery issues. They reflect a mismatch between delivery approach and the nature of AI work.
8. Final Thoughts
AI does not replace the fundamentals of project management.
It changes how those fundamentals are applied in practice.
A structured approach provides confidence not by removing uncertainty, but by managing it through controlled experimentation, validation, adoption planning and evidence-based progression.
For experienced Project Managers, the key is to recognise the delivery mode, understand the dependency on data and feasibility, structure work to reduce uncertainty, and ensure that technical progress translates into usable outcomes.
Project management ensures delivery is structured. Iterative discovery, design and delivery ensure that the right solution is built.
This Insight is part of the Orr Consulting AI Insights Library — structured thinking for AI transformation leaders and decision makers.
9. Call to Action
If your organisation is undertaking AI delivery and would like to structure projects more effectively to manage uncertainty and improve outcomes, we would be pleased to support you.
Subscribe to Orr Consulting to receive occasional emails with practical AI Insights and updates.