AI world models raise the prospect of systems that can rehearse actions and anticipate outcomes. For boards, the opportunity lies in better operational decisions, supported by evidence about where those predictions can be trusted.
Much of the board conversation about artificial intelligence has centred on information: how quickly it can be produced, how effectively it can be analysed and how much effort can be removed from preparing it. Those remain useful questions. A further question deserves attention as AI research develops: what happens when a system can model an environment, consider alternative actions and predict their consequences before acting?
This is the ambition behind AI world models. For organisations with substantial physical operations, the potential implications extend into productivity, capital investment, workforce capability and operational assurance. My view is that boards should understand the development early, while requiring a clear connection between research results and the decisions their own organisations need to make.
The underlying idea is straightforward. A world model learns a representation of an environment and how it changes. That representation can help an AI system anticipate what might happen next, including how an action could affect the outcome. The relevance to a business depends on what the model captures, how accurately it predicts and whether those predictions improve performance under real operating conditions.
Three research examples help explain the direction. Google DeepMind’s Genie 3 demonstrated the generation of interactive environments from text prompts, with potential applications in training and evaluating AI agents. Its launch documentation also identified limitations, including constrained actions, difficulties modelling interactions between independent agents and a short duration of consistent interaction. The ability to generate an environment therefore needs to be assessed separately from its suitability for a particular operational purpose. Meta’s V-JEPA 2 describes a system trained on video to learn patterns in physical behaviour and support prediction and planning. Its laboratory demonstrations included robots reaching for objects, picking them up and placing them elsewhere. These are useful examples of a connection between learning from observation and taking physical action, within the conditions demonstrated. In the Dreamer research published in Nature, the researchers describe an agent that learns a model of its environment and improves its behaviour through imagined future scenarios. The paper reports performance across more than 150 tasks using a single configuration. It concerns the third generation of Dreamer, and provides experimental evidence for this approach to learning and control.
Taken together, these examples suggest a direction worth following. They do not establish that a system can understand the full complexity of an organisation, or that performance in a research environment will transfer directly into a commercial setting. The board implications that follow are my assessment of the research, rather than outcomes demonstrated by those studies. The commercial attraction is the possibility of reducing the cost of learning. Organisations already spend considerable time and money discovering how changes affect their operations. A revised production sequence, a different warehouse layout or a new maintenance approach can create consequences that become visible only after implementation. Where a model can represent the relevant conditions accurately, it may become possible to explore more alternatives before committing resources.
Consider a hypothetical distribution business evaluating changes to picking activity. Management might want to understand the effect on throughput, congestion, handling errors and employee workload. A useful modelling capability would need to capture the constraints that influence those outcomes, including equipment, space, demand variation and the interaction between people and machines. Its value would depend on whether it helped the business select and implement a better course of action. An impressive visual demonstration would provide only part of that evidence. This also means boards should ask how a proposed capability compares with existing methods. Simulation, optimisation and digital twins already offer ways to examine operational choices. A world model proposal should explain the additional value it expects to create: perhaps reducing the effort required to build a model, adapting to changing conditions or supporting a wider range of tasks. Management should establish that comparison before treating the technology as a new investment priority.
In my work on operating models and delivery performance, a recurring issue is the distance between local improvement and organisational benefit. An individual activity can become faster while the end-to-end outcome barely changes. Work simply reaches the next constraint sooner. The same discipline should apply here. A more capable robot or a better prediction engine creates value ONLY when the surrounding system can use that capability effectively. For example, increasing handling capacity may have little commercial effect if stock availability remains the binding constraint. Improving scheduling may deliver limited benefit where maintenance decisions or supplier commitments prevent the organisation from responding. Boards should therefore expect an investment proposal to show how the capability affects the wider flow of work, which dependencies need to change and who owns the resulting benefit.
That brings the discussion directly into operating model design. If AI is used to recommend a different course of action, somebody must have the authority to consider and implement it. If the system is permitted to act, its boundaries need to be explicit. Management should define the decisions it can make, the conditions requiring human review and the arrangements for stopping or reversing an action where that is possible. Responsibility needs to remain clear across operations, technology teams and suppliers. Assurance should focus particularly on the relationship between predicted and observed outcomes. A model can produce a convincing account of what will happen while missing a factor that matters in practice. Boards should expect management to explain how predictions will be tested, how errors will be recorded and what would cause confidence in the system to be withdrawn. Testing should include the situations in which the organisation is least comfortable being wrong.
The consequences of error should shape the deployment approach. A recommendation about a reversible scheduling adjustment allows a different level of experimentation from an instruction that moves machinery close to employees. An appropriate proposal would connect the permitted level of autonomy to evidence of performance, the severity of failure and the effectiveness of the available controls. Successful testing in a narrow setting should lead to a defined next stage of learning.
Workforce planning belongs in the same discussion. Where organisations explore these capabilities, they will need people who can explain how work actually happens, recognise when a prediction is implausible and respond when the system encounters unfamiliar conditions. Operational knowledge becomes part of the design and assurance process. Management should assess how responsibilities, supervision, training and maintenance requirements would change alongside any estimate of labour savings. This is why I would want workforce planning, technology investment and operational assurance to use a common evidence base. A business case becomes difficult to trust when the finance assumptions depend on one level of autonomy, the safety assessment assumes another and the workforce plan expects a third. Those assumptions need to be visible and consistent before the board is asked to approve expansion.
The practical starting point is a specific operational problem with a measurable baseline. Management should identify the decision it wants to improve, the cost of the current approach and the evidence needed to justify further investment. A bounded evaluation can then compare predictions with actual outcomes and assess the complete cost of using the capability, including integration, supervision, training, maintenance and recovery from failure. Progress should be judged against explicit criteria. These might include better decision quality, fewer errors, reduced downtime or improved throughput without unacceptable effects elsewhere. The evaluation should also establish whether a simpler approach could achieve a comparable result. Further funding can then follow evidence that the capability is useful, manageable and economically worthwhile in the intended setting.
Boards do not need to resolve whether world models will lead to human-level intelligence to make sensible decisions about them. They need to understand where the research could become relevant to their organisation, which claims remain unproven and what management would need to demonstrate. That creates room for purposeful experimentation while preserving a clear basis for investment.
For me, the significant prospect is an organisation becoming better at anticipating the consequences of its choices. Realising that prospect will require operational understanding, accountable leadership and disciplined learning alongside technical progress. The board’s role is to ensure those elements develop together, and that confidence grows from observed performance.