For a while, the story of enterprise AI was mostly about the model. Get access to a stronger large language model, give it more context, connect it to your documents and systems, and the quality of the application should improve.

That story is now changing. More people have realised that a powerful model is not enough. AI also needs domain knowledge.

That is certainly true. But in financial services, it still does not quite explain the problem.

Modern large language models already know a surprising amount about finance. Ask one to explain bonds, swaps, options, corporate actions, settlement, collateral, market risk or accounting concepts, and the answer can be remarkably good. In breadth, an LLM can sometimes feel less like a textbook and more like a small financial library that you can talk to.

Yet many financial AI applications still struggle once they leave the demonstration environment and enter real work.

The problem is not simply that AI does not know enough finance.

The more interesting problem is that the financial knowledge AI possesses is often not the knowledge you need to solve the problem in front of you.

The impressive answer that does not quite help

This becomes obvious surprisingly quickly when you start using AI for real financial work.

Ask an AI assistant about a corporate action and it may give you a detailed explanation of dividends, stock splits, rights issues, mergers and tender offers. It might explain how each event affects shareholders, positions, settlement and valuation. The answer can be accurate, informative and much more detailed than you asked for.

Then ask:

Which of our trades are affected by this corporate action?

Suddenly, most of that knowledge is not enough.

To answer the real question, the AI needs to know which security the event applies to, how your organisation identifies that security, which trades are considered economically or operationally affected, what lifecycle state those trades are in, which dates matter, which system is authoritative and what your business process expects to happen next.

The AI may know a great deal about corporate actions while knowing almost nothing that allows it to answer this corporate-action question correctly.

That difference is easy to overlook because AI is exceptionally good at producing information. A response can be long, polished and full of correct financial terminology while containing very little of the information that actually determines the decision.

More information is not the same as more useful knowledge.

AI already knows the textbook

There is an important distinction here.

A language model can possess substantial general financial domain knowledge. It can know what an issuer is, what a swap is, what settlement means and how a bond normally behaves.

This matters. If the AI did not understand finance at all, almost nothing useful could be built on top of it.

But real financial work rarely asks only textbook questions.

A textbook asks:

What is an issuer?

Your organisation asks:

Which legal entity should be treated as the issuer of this security for this process?

A textbook asks:

What is a price?

Your valuation process asks:

Which price should we use, from which source, observed at what time, under which valuation rule?

A textbook can explain what a trade amendment is.

Your system needs to know whether an amended trade is considered the same trade, a new version of the trade, a new contractual state, or something else entirely.

These are not questions that require another chapter of general finance education. They require knowledge of how your organisation understands and operates its financial world.

That is a different kind of knowledge.

From domain knowledge to actionable knowledge

I find it useful to call this actionable knowledge.

General domain knowledge tells AI how finance usually works. Actionable knowledge tells it what something means here, how it relates to other things here, which facts should be trusted here, and what conclusion or action follows here.

Consider a simple question:

Who issued this security?

A general-purpose AI may understand perfectly well what an issuer is. But a real answer may depend on details that are specific to the organisation.

One system may store the direct legal issuer. Another may store the ultimate parent because that is what risk reporting needs. A legacy application may use a field called Issuer even though its actual meaning has changed over the years. A third-party data source may disagree with your internal security master.

The difficult part is no longer defining issuer.

The difficult part is understanding what your organisation means when it says issuer in this particular context.

This is why saying that enterprise AI “needs domain knowledge” is correct but incomplete.

AI often already has the domain knowledge.

What it frequently lacks is the organisation-specific knowledge that makes that domain knowledge actionable.

This is where AI becomes dangerously convincing

There is another reason this problem matters.

When AI does not know enough, it does not always behave like a traditional information system.

A database query that cannot find a row normally returns nothing. A deterministic rule engine that cannot satisfy a condition can fail.

A generative model has another option.

It can construct a plausible interpretation.

This is one of AI’s greatest strengths. It can infer, connect ideas, fill gaps and produce a coherent explanation from incomplete information.

It is also one of the reasons enterprise AI can be difficult to trust.

Suppose an AI understands the normal industry meaning of issuer, but your organisation uses that term differently in one internal dataset. If the internal meaning is not made explicit, the model has everything it needs to produce a sensible answer based on its general knowledge.

The answer may be financially sophisticated. It may be internally consistent. It may be written with absolute confidence.

And it may describe the wrong world.

This is more subtle than the classic idea of hallucination. The AI may not have invented something obviously absurd. It may have applied perfectly reasonable financial knowledge in a place where your organisation uses a different meaning.

The result is plausible but misaligned.

That kind of error is especially dangerous because it looks so much like expertise.

The organisation knows things the model does not

The good news is that the missing knowledge usually already exists somewhere.

Someone in the organisation knows what the Issuer field really means.

Someone knows why two security-master records disagree.

Someone understands why a particular price source is trusted for one process but not another.

Someone knows which lifecycle event changes a trade into a particular state.

Someone knows whether two identifiers refer to the same thing, different versions of a thing, or completely different things.

The organisation may therefore know far more about its own financial world than the AI does.

The problem is that this knowledge is rarely available as one coherent body of knowledge.

Some of it sits in people’s heads. Some is buried in documentation. Some exists only in source code or transformation logic. Some is encoded indirectly in database schemas. Some lives in spreadsheets, policy documents and operational procedures. Some survives mainly as institutional habit.

And some lives in that familiar enterprise knowledge-management system:

“Ask Sarah. She knows how this actually works.”

For decades, organisations could operate like this because humans filled in the semantic gaps.

AI cannot reliably depend on invisible organisational memory.

Giving AI more documents is not enough

The obvious response is to connect the AI to more information.

Give it access to the data warehouse.

Add retrieval over internal documentation.

Connect the data catalogue.

Index the policies.

Expose APIs.

Put more information into a vector database.

All of these things can be valuable. But access to more information does not automatically create aligned understanding.

Imagine that two internal documents use the word instrument differently. Retrieving both documents gives the AI more evidence, but it does not by itself explain that two meanings exist.

If two systems store different issuer relationships, exposing both datasets does not tell the AI why the relationships differ.

If two identifiers appear next to the same company name, retrieval does not tell the AI whether they identify the company, a security, a listing, an account or a record.

If two market-data sources report different values, access to both numbers does not tell the AI whether they conflict, were observed at different times, apply to different contexts, or represent different kinds of measurement.

The challenge is therefore not only:

Can AI find the information?

It is also:

Can AI understand what the information means?

That is a much harder problem.

The missing layer is organisational meaning

Once we look at the problem this way, the gap between impressive AI demonstrations and dependable enterprise applications becomes easier to understand.

General financial knowledge makes AI informative.

Organisation-specific financial knowledge makes it actionable.

For AI to solve real financial problems, it needs to understand things such as what the organisation considers to be the same thing or a different thing, what particular business terms mean, which roles entities are playing, how products and trades relate, what events change their state, which sources are authoritative, and when a fact or relationship is valid.

Those are not merely data-access questions.

They are questions about meaning.

And this is where financial ontology modelling enters the story.

Not because every organisation suddenly needs to become interested in philosophical ontology. And not because everything needs to be converted into RDF, OWL or a knowledge graph.

Ontology modelling starts much more simply.

What is the thing?

What merely identifies it?

Is this a kind of thing, a role it plays, or a state it is currently in?

How are these two things related?

When was that relationship true?

What happened to change it?

Who says this fact is true?

What evidence supports it?

These are ordinary business questions. Financial professionals answer them every day.

The problem is that we have not always made the answers explicit enough for machines to use them consistently.

AI needs your financial meaning

This leads to a slightly different way of thinking about financial AI.

The next challenge is not simply to teach AI more finance.

AI already knows a lot of finance.

The challenge is to connect that general knowledge to the financial meaning of the organisation itself.

That means making explicit what the organisation means by its concepts, how those concepts relate, which distinctions matter, how they change through time, and how facts can be trusted.

Only then can general financial knowledge become reliably actionable inside a particular institution.

So perhaps the real requirement for enterprise financial AI is not:

Give AI more knowledge.

It is:

Give AI the knowledge required to understand your financial world.

And before we can give that meaning to AI, we first need to make it clear to ourselves.

That is the starting point for financial ontology modelling.

Model the meaning before you model the data.

• *Financial Ontology Modelling