As access to artificial intelligence becomes easier, the ability to acquire AI tools is becoming less important than the institutional capacity to use them well.
AI adoption is accelerating.
Organizations are buying licences, introducing assistants, running pilots and giving employees access to increasingly capable models. Governments are developing strategies, training staff and experimenting with AI across administrative processes and public services.
These developments are easy to count.
They are also easy to mistake for capability.
An institution can have widespread access to AI while remaining poorly equipped to decide where it should be used, redesign processes around it, evaluate its outputs or remain accountable for the consequences.
The distinction matters because access to AI is becoming cheap.
Institutional capability is not.
Adoption is increasingly the easy part
The barriers to experimenting with AI have fallen remarkably quickly.
Many useful systems require little new infrastructure from the organization using them. Employees can access sophisticated models through ordinary software interfaces. Existing workplace applications increasingly incorporate AI by default.
This means that conventional indicators of digital progress can become misleading.
The number of licences purchased tells us something about access.
The number of employees using AI tells us something about diffusion.
The number of pilots tells us something about experimentation.
None necessarily tells us whether the organization has become better at achieving its objectives.
This gap is already visible in government.
The OECD’s Digital Government Outlook 2026 finds AI use in at least one area of government in 35 of 36 OECD countries surveyed. Yet the practical conditions needed to scale it remain uneven, including data governance, infrastructure, skills and organizational capacity. The OECD also identifies a proliferation of pilots whose impact is difficult to measure and whose potential to scale may be limited.
The interesting question is therefore shifting.
Not:
Has the institution adopted AI?
But:
What can the institution now do that it could not do before?
Capability sits around the technology
Consider a policy analyst using an AI system to produce a first draft of an analysis.
The model may summarize documents faster, generate alternatives and identify connections across large amounts of material.
That can produce an immediate productivity gain.
But whether it produces a better institution depends on everything surrounding the model.
Can the analyst recognize when the output is wrong?
Can the organization provide reliable data and context?
Are sensitive information and access rights handled appropriately?
Can assumptions and sources be traced?
Does someone remain responsible for the resulting recommendation?
Can the workflow distinguish between tasks that can be delegated and judgments that should remain human?
And does the time saved actually translate into better analysis, faster decisions or improved services?
These aren’t properties of the AI system.
They are properties of the institution using it.
The OECD makes a similar distinction in its work on an AI-ready public workforce. It argues that internal capability is necessary not only to use AI effectively, but also to maintain accountability, ensure compliance and align AI use with organizational objectives. Skills and internal capability remain among the most commonly cited barriers to public-sector adoption.
The pilot trap
AI makes experimentation unusually easy.
That is an advantage, but it can create an institutional trap.
A successful demonstration can show that a technology is capable of performing a task. It doesn’t necessarily show that an organization is capable of incorporating that technology into normal operations.
Scaling introduces different problems.
Data have to be maintained. Responsibilities have to be assigned. Procurement has to work. Staff need appropriate skills. Systems need monitoring. Errors need escalation routes. Benefits need to be measured against costs and risks.
A pilot can bypass many of these constraints.
An institution cannot.
This helps explain why an organization can simultaneously have dozens of AI experiments and very little transformation.
The relevant transition is not from no AI to AI.
It is from experimentation to institutionalization.
Capability is not just AI literacy
Training matters, but capability should not be reduced to teaching employees how to prompt a model.
The OECD reports that 32 of 36 countries surveyed have AI training programmes for government employees. Yet training for specific applications is considerably less common: only 13 reported training related specifically to AI use in public services, and the same number for policymaking.
General familiarity can help people use the technology.
Institutional capability requires more.
It includes the ability to redesign workflows, organize data, procure intelligently, evaluate performance, govern risks and understand when AI should not be used.
It also requires judgment.
As systems become easier to operate, knowing how to use AI may become less scarce than knowing where its use creates value and where it creates unacceptable uncertainty.
That is a different capability.
Efficiency and legitimacy
For public institutions, there is another constraint.
AI can improve efficiency without automatically increasing institutional legitimacy.
The OECD’s 2026 survey on trust in public institutions found that roughly four in ten respondents across participating countries were confident that government AI could improve services or reduce costs. Confidence was lower that governments would preserve safeguards such as fairness, transparency, privacy and effective human oversight.
This creates an important asymmetry.
An institution may become technically capable of automating a process before citizens consider that automation legitimate.
Capability in government therefore cannot mean maximizing automation.
It must also include the ability to determine where automation is appropriate, explain how systems are used, preserve accountability and recognize decisions whose legitimacy depends on meaningful human judgment.
In that sense, governance is not a constraint placed on AI capability.
It is part of the capability itself.
When adoption stops being a useful signal
During the early phase of a technology, adoption rates can reveal a great deal.
They tell us which organizations are experimenting, where investment is occurring and how quickly a technology is spreading.
But as access becomes widespread, adoption gradually loses some of its informational value.
If almost every organization has access to similar AI systems, possession of the technology ceases to differentiate them.
The difference moves elsewhere.
Some organizations will integrate AI into poorly designed processes and produce the same outcomes slightly faster.
Others will redesign work around new capabilities.
Some will accumulate pilots.
Others will build reusable infrastructure, governance and institutional knowledge.
Some will outsource understanding along with implementation.
Others will retain enough internal capability to challenge suppliers, evaluate systems and make informed choices.
The technology may increasingly be shared.
The capacity to organize around it will not be.
What would weaken this argument?
The distinction between adoption and capability would matter less if AI systems became sufficiently autonomous that organizational context, data quality, workflow design and human judgment ceased to be important constraints on performance.
It would also weaken if organizations with high AI adoption consistently achieved better outcomes regardless of their internal structures and capabilities.
Current evidence points in the opposite direction. The OECD’s assessment of government AI finds widespread adoption alongside uneven delivery capabilities and continuing weaknesses in data governance, procurement, skills, organizational capacity and impact measurement.
But the technology is changing quickly enough that this should remain an empirical question rather than an assumption.
Watch next
The most revealing indicators of institutional AI maturity may therefore become less technological.
Watch whether organizations move from counting tools, users and pilots toward measuring changed processes and outcomes.
Watch whether AI training moves from general literacy toward profession-specific capability.
Watch whether organizations develop enough internal expertise to evaluate vendors and systems rather than relying entirely on external providers.
Watch whether successful pilots become durable operating practices.
And watch what happens to accountability when AI becomes embedded deeply enough in everyday work that using it no longer feels like a distinct activity.
The next divide may not be between institutions that use AI and those that do not.
It may be between institutions that have AI and institutions that have learned how to change because of it.
Leave a Reply