Are We Building Institutional Capability Inside AI When We Don't Need To?

There is a direction emerging in enterprise AI that I think deserves more scrutiny.

As language models become more capable, we are giving them more of the institution.

We provide policies and procedures as context. We give agents access to systems and tools. We ask models to retrieve information, interpret rules, determine what should happen next, coordinate other agents, check their work and sometimes act on the result.

Each step can make sense in isolation.

Taken together, however, they raise a larger question:

Are we gradually rebuilding institutional capability inside AI?

And if so, do we actually need to?

Institutions already have capability

An organisation does not begin empty and wait for artificial intelligence to make it capable.

It already has people, systems, databases, policies, business rules, workflows, APIs, records, calculations, delegations, services and professional expertise.

Some of these capabilities are excellent. Some are fragmented, difficult to access or poorly documented. Some need modernisation.

But they exist.

The problem is often not the absence of capability.

It is that capability is difficult to discover and assemble around what someone is actually trying to achieve.

A worker may know the outcome they need but not which system contains the information.

They may know the situation but not which policy applies.

They may need an approval without knowing who holds the delegation.

They may need several organisational capabilities but have no simple way of bringing them together.

For decades, we dealt with this by teaching people how to navigate the institution.

Learn the applications. Find the forms. Search the intranet. Understand the process. Know who to call.

AI gives us an opportunity to change that relationship.

But I wonder whether we are taking the opportunity in the wrong direction.

What if AI makes capability discoverable instead?

Imagine someone in an organisation says:

We need to bring Acme in for a $75,000 piece of work. Can we get them approved?

A capable AI system could reason extensively about this request.

It could interpret the user's intent, search procurement policies, retrieve supplier information, analyse requirements, determine approval pathways, identify risks, propose an answer and perhaps ask another model to verify its reasoning.

That may look sophisticated.

But consider what the institution may already know.

Is Acme an existing supplier?

There may be an authoritative supplier register.

What procurement threshold applies?

There may be an established rule.

What evidence is required?

The procurement framework may already specify it.

Who can approve the expenditure?

The institution may already maintain delegations.

What process needs to occur?

There may already be a workflow or service capable of executing it.

These are not necessarily questions that need to be inferred.

They are things the institution may already be able to establish.

Complicated does not necessarily mean uncertain.

That distinction matters.

Establish what can be established

Language models are extraordinarily useful when language itself is part of the problem.

They can help interpret what someone means, resolve ambiguity, classify information, synthesise material, explain complexity and reason through genuinely uncertain situations.

Those are significant capabilities.

But not everything that appears in language is fundamentally a language problem.

A threshold that can be calculated does not need to be inferred.

An authoritative record does not need to be predicted.

A person's delegation does not become more authoritative because a model concludes that they probably possess it.

A business rule does not become more reliable because several AI agents agree on its likely meaning.

And an institutional action has not occurred merely because a model says that it has.

There is a simple principle here:

If something can be authoritatively established, establish it. Use inference for what genuinely needs to be inferred.

This is consistent with the broader direction of AI risk-management frameworks that emphasise governing, mapping, measuring and managing AI risks rather than treating model output as inherently trustworthy.[1]

The architectural choice underneath AI adoption

This creates two very different possible directions for enterprise AI.

In one, more and more of the institution moves into the AI layer.

Policies become model context.

Organisational knowledge becomes embeddings and memory.

Business rules become reasoning instructions.

Existing capabilities become agent tools.

Workflow becomes agent orchestration.

Authority becomes permissions assigned to artificial actors.

The model increasingly becomes the place where the institution is interpreted, reconstructed and operated.

There may be situations where elements of this architecture are appropriate.

But it should not become the default simply because models are increasingly capable of doing it.

There is another direction.

Keep institutional capability institutional.

Make it discoverable.

Let someone begin with what they are trying to achieve rather than requiring them to know which application, process or organisational unit can help them.

Use language technology to understand that intent and identify what is relevant.

Then engage the capabilities the institution already possesses.

The pattern becomes:

Human intent → interpretation → capability discovery → establishment → governed work → judgement where required → accountable outcome.

AI can participate throughout that sequence where it adds value.

It does not have to become the sequence itself.

This is not an argument for less AI

It is an argument for being more precise about what AI is for.

The distinction matters because otherwise we risk framing every organisational problem as an AI problem.

If a database already knows something, retrieve it.

If an API already provides a capability, invoke it.

If a rule can be deterministically executed, execute it.

If specialised software performs a calculation reliably, use it.

If an authorised person must exercise judgement, give that person the evidence and context needed to judge.

And where language, ambiguity, synthesis, perception, prediction or genuine uncertainty exists, use AI.

This does not diminish AI.

It allows us to use it where its capabilities are distinctive rather than asking it to imitate everything surrounding it.

The productivity question

This also changes how we should think about AI productivity.

Model activity is not itself productivity.

More tokens are not productivity.

More agents are not productivity.

More AI-generated work is not necessarily productivity.

The useful question may be:

What valuable uncertainty did we pay the model to resolve?

If the model helped understand an ambiguous request, discover relevant capabilities or reason through a situation that could not otherwise be mechanically established, the inference may have created substantial value.

If it spent thousands of tokens reconstructing a rule, authority or fact the organisation already possessed, we should ask whether that was intelligence or simply architectural duplication.

That matters because AI consumption is increasingly a variable operating cost. The FinOps Foundation argues that token economics should connect AI consumption to business outcomes, and notes that context, retries, orchestration and agent-to-agent communication can compound token consumption.[2]

The wider productivity case for generative AI remains significant, but the OECD notes that realising its long-term productivity potential depends on how the technology is implemented and diffused.[3]

At scale, architecture therefore matters.

There is also a question of institutional sovereignty

The issue goes beyond cost.

If an organisation gradually moves its policies, operating logic, context, memory and decision pathways into model-specific prompts, agents and orchestration, what happens when the model changes?

What happens when the provider changes?

What happens when the organisation needs to demonstrate exactly where a rule came from, who had authority, what evidence applied or why an action occurred?

An institution should be able to change its models without having to rediscover how the institution itself works.

Models should be replaceable.

Institutional capability should endure.

That means organisational knowledge, rules, authority, evidence and accountability need identities independent of whichever AI happens to interact with them.

AI may be the interface, not the institution

For most of the digital era, people had to learn the structure of technology before they could use organisational capability.

We learned applications, screens, menus, databases, forms and workflows.

Language models offer something genuinely different.

People may increasingly be able to begin with intent:

This is what I am trying to achieve.

AI can help the institution understand.

But understanding the request does not mean the model needs to become the institution that fulfils it.

Perhaps the larger opportunity is to make institutional capability discoverable from human intent.

People express what they need.

AI helps interpret what they mean.

The institution establishes what it already knows.

Existing capabilities do the work they are designed to do.

And people exercise judgement where judgement genuinely belongs.

That leads me to a different way of thinking about enterprise AI:

Do not rebuild institutional capability inside AI simply because AI has become capable of imitating it.

Use AI to make the institution's existing capability easier to discover, assemble and use.

The future institution may not be the one with the most AI inside it.

It may be the one in which people can reach what the institution already knows and can do, without needing to understand the machinery underneath.

AI should not have to become the institution. It should help people reach and use the institution's capabilities.

References

[1] National Institute of Standards and Technology (NIST), Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, 26 July 2024. https://doi.org/10.6028/NIST.AI.600-1

[2] J.R. Storment, FinOps Foundation, Token Economics: The Atomic Unit of AI Value, 10 May 2026. https://www.finops.org/insights/token-economics-the-atomic-unit-of-ai-value/

[3] Flavio Calvino, Daniel Haerle and Sarah Liu, OECD, Is generative AI a General Purpose Technology? Implications for productivity and policy, OECD Artificial Intelligence Papers No. 40, 27 June 2025. https://doi.org/10.1787/704e2d12-en