From Projects to Learning Systems

Artificial intelligence is making it less costly to produce analysis, hypotheses and prototypes. That does not automatically make organisations more innovative.

A hypothesis can now be formulated in minutes, a prototype in hours, a simulation in an afternoon. Research that once took days can sometimes be completed in a morning. Yet companies are not learning as quickly as model capabilities are growing.

The data help frame this tension. The Stanford AI Index 2026 ↗ estimates that 88% of surveyed organisations used AI in at least one function in 2025, up from 78% in 2024; agent use, however, remained at an early stage. In McKinsey’s 2025 global AI survey ↗, 62% of respondents say their organisation is at least experimenting with AI agents. Twenty-three per cent report scaling them in at least one area, but in no individual function does the share exceed 10%. These are findings from separate surveys, with their own definitions and samples: they describe a broad picture, not the condition of every company.

Adoption, then, is a weak measure of transformation. It tells us how many people have access to a tool; it does not tell us how quickly an organisation turns what it discovers into decisions, changes and outcomes.

This paper’s thesis is that the next competitive advantage in enterprise AI may come less from ever more capable models and more from better organisational context and faster learning cycles.

A foundation model may have broad knowledge of the world. It does not, by itself, know how a particular company works.

It does not know what “customer” means in that organisation, which revenue definition is authoritative, which KPI takes precedence when two systems return different figures, who owns a decision, or which exceptions apply to a process. It cannot know which actions require approval. To act usefully, it needs context.

Context is more than a longer prompt. It is the information, instructions, tools, memory and constraints that allow a system to interpret a request and act within intelligible limits. Anthropic’s discussion of context engineering ↗ describes the move from designing a single instruction to designing the overall information environment. In its report on enterprise AI ↗, OpenAI identifies integrated access to enterprise data, codified institutional knowledge and an understanding of organisational context among the factors associated with maturity.

The convergence matters: available intelligence is important, but its value depends on what a system can know, interpret and do inside an organisation.

For years, the enterprise knowledge base has primarily been a place to store procedures, policies, presentations and documents. It serves people who search, read and interpret. An agent expected to take action also needs shared meaning, explicit relationships and applicable rules.

Microsoft’s GraphRAG research ↗ illustrates why retrieving isolated passages may not be enough when answers depend on relationships spread across documents. In parallel, Microsoft describes Fabric IQ and ontologies ↗ as ways to provide people and agents with shared context about business entities, relationships and rules. These are examples of directions being explored by industry, not evidence that a universal solution already exists.

A progression helps distinguish the layers: data record what happened; information organises it; knowledge gives it meaning; context explains why it matters to a particular problem; intelligence interprets the context; action changes a situation. Learning connects the outcome to future experience.

Ontology is often treated as a technical artefact. It can instead become an explicit representation of the organisation: entities such as customer, product, boutique, market, campaign, order, employee and supplier; the relationships among them; and the properties, rules and measures that give them operational meaning.

This structure can make practical questions explicit. Who is an active customer? How is revenue calculated? What makes a customer eligible for a private event? Which threshold triggers an action? When the answers remain implicit or diverge across systems, AI inherits that ambiguity.

In luxury, the example is immediate. If “active customer” means one thing in CRM, another in retail and another in e-commerce, a clienteling assistant may produce inconsistent recommendations. The problem is not necessarily the model; it may be that the organisation has not agreed on a shared definition. Ontology does not replace cross-functional discussion; it makes the outcome of that discussion legible and reusable.

This paper proposes reading a mature AI environment as seven connected capabilities. Governance runs across every layer; learning returns what the cycle has produced to the knowledge base.

This is not a purely linear chain. The knowledge base feeds context; context guides intelligence; action produces outcomes; evaluating those outcomes updates knowledge. Governance defines access, accountability, controls and limits at every step.

If AI reduces the cost of analysis, prototypes, simulations and code, an organisation can test more hypotheses with less effort. The advantage shifts from finding one perfect idea to learning faster through disciplined experiments.

The proposed cycle is: context → hypothesis → experiment → evidence → decision → action → learning → new context. The last step is decisive: what is learned returns to the knowledge base and makes the next cycle better informed.

A traditional project is often organised primarily as a delivery mechanism. An AI-native project can also become a structured learning system. An experiment that fails to confirm a hypothesis creates value when it rules out an unpromising path quickly and at low cost. The project then becomes a way to reduce uncertainty, not merely a sequence of activities to complete.

The table does not claim that every traditional project is unlike every AI experiment. It is a lens for asking which parts of work might be redesigned as the cost of testing a hypothesis falls.

AI does not make the project manager redundant; it shifts the centre of the role. The traditional remit coordinates scope, time, budget, resources, dependencies and stakeholders. The emerging role also coordinates people, AI, data, knowledge, agents and governance.

We can describe this as orchestration: designing the conditions in which different forms of intelligence work together. People retain judgement, accountability, ambiguity management and leadership. AI can contribute analysis, synthesis, generation and simulation. Organisational knowledge provides institutional memory and rules. Agents perform actions within defined workflows and permissions.

This change has entered the professional conversation too. In June 2026, the Project Management Institute published a standard for AI in portfolio, programme and project management ↗. The existence of a standard does not mean every organisation has adopted this model; it does signal an emerging professional language for managing AI in project work.

The question, then, is not only whether AI will enter project management. It is what a project becomes when experimentation costs less and learning can be one of its principal outcomes.

In this model, governance is not a control added at the end. It sets the boundaries within which experimentation and autonomous execution can take place responsibly. What can an agent do on its own? What may it only recommend? Which actions need human approval? What data can it access, which systems can it change, and how are its actions recorded and evaluated?

This is controlled autonomy: enough freedom to accelerate, and enough structure to preserve accountability and the ability to intervene. Clear rules can reduce operational uncertainty and make it easier to experiment within shared limits. They do not eliminate risk or replace case-by-case judgement.

This paper proposes a management relationship to bring these factors together:

Innovation speed = Context × Knowledge × Experimentation × Governance × Execution

The multiplicative form is intentional: a significant weakness in one factor can limit the value of the others. The best model, for example, may create little value if the knowledge base is confused or it is unclear who can authorise actions.

This is not a quantitatively validated equation, and the factors are not measured here on a common scale. It is a conceptual map: a reminder that innovation speed depends on complementary capabilities, and a way to frame questions for testing in organisational cases.

If the aim is to learn faster, measures should track the time needed to move from a problem to a useful decision. The first measure concerns context: if it takes weeks of meetings to reconstruct the relevant business picture, much of the potential gain is lost before an experiment even begins.

These measures are proposals, not established standards. An organisation would need to define their start and end points, the minimum quality of evidence, and the differences between use cases. Reducing time without monitoring errors, quality and consequences could reward the wrong kind of speed.

This paper proposes six levels to describe a possible maturity path. They are a discussion framework, not a validated diagnostic scale or a ranking of companies.

The highest level is not measured by the number of models, tools or agents. It is measured by the ability to learn continuously and reuse what has been learned. In this hypothesis, the move from level three to four depends largely on the quality of context and the knowledge base.

This perspective points to several areas of work:

These points are not a one-size-fits-all recipe. A useful starting point is a real problem, an identifiable owner and an experiment with explicit evaluation criteria.

The strategic question for business leaders may no longer be how much AI their organisation has adopted. A more useful question is: how quickly do we turn new knowledge into decisions, and decisions into experiments?

Companies that learn faster may eventually outperform those that merely automate faster. The next generation of projects may have a broader purpose: not only to deliver something, but to reduce uncertainty, produce evidence and generate new knowledge.

In this scenario, the knowledge base is no longer an archive at the end of a process. It is where the cycle begins again.

This paper distinguishes reported evidence from its proposed interpretation. The data on adoption, technological direction and published research come from the cited sources, each with a specific scope. The innovation-speed equation, six-level maturity model and “from projects to learning systems” framework are strategic models developed for this paper. They should be read as hypotheses, not as empirical findings already demonstrated.

The next stage of the work will examine five contrasting cases — luxury and retail, industry, technology, Asia and a large US company — to identify recurring patterns and enabling conditions, not to produce a ranking.