The Future of Enterprise Analytics Is AI That Knows What to Look For

01The enterprise has been teaching machines what things mean

Consider a simple business question:

“Why did customer retention decline last quarter?”

In a modern enterprise, answering this question can be considerably easier than it was a few years ago.

A semantic layer can tell an AI system what a “customer” is, how “retention” is defined, which dimensions can be used to segment it, and which data sources are authoritative.

A context layer can go further — connecting entities, relationships, business definitions, documentation, lineage, policies, and other relevant context.

The machine can now understand that:

Customer → purchased → Product → belongs to → Category

It can understand that an “active customer” has a specific business definition.

It can know which metric represents retention.

It can know where the underlying data lives.

And it can generate a query that is much more likely to be correct.

This is a major step forward.

But an analyst does not stop when the query returns a number.

02An analyst doesn’t just answer questions

Imagine the first result says:

Retention declined 4.2% quarter-over-quarter.

The analyst’s job has barely started.

They might ask:

• Which customer segments drove the decline?
• Is the decline concentrated in new or existing customers?
• Did a particular product or channel drive it?
• Did the change occur across all geographies?
• When did the decline actually begin?
• What changed around that time?
• Is the decline driven by acquisition quality, product usage, pricing, or service?
• Are there segments where retention actually improved?
• Is this a real behavioral change or a measurement artifact?

Each answer creates another question.

And the analyst keeps going until the pattern becomes sufficiently clear to explain.

This is the investigation loop:

Question → Query → Result → Hypothesis → Next Question → Query → Evidence → Insight

Semantic and context layers make this loop faster and more reliable.

But they don’t eliminate the loop.

The human still has to decide what to ask next.

03The bottleneck is moving

This is where the conversation around enterprise AI is changing.

The early challenge was:

Can AI understand my data?

Then:

Can AI answer questions about my data accurately?

Now the harder question is emerging:

Can AI independently investigate my data to find something worth knowing?

These are very different capabilities.

A system that can answer:

“What was revenue last quarter?”

is useful.

A system that can investigate:

“What changed in revenue, where did it change, what explains the change, and what should the business pay attention to?”

is doing something fundamentally different.

The first is question answering.

The second is investigation.

04From semantics to investigation

This is also where the evolution of semantic technology becomes interesting.

Traditional ontology focused on meaning:

What is a Customer?
What is a Product?
How are they related?

Knowledge graphs made those relationships traversable:

Customer → purchased → Product → Category

Business rules added operational logic:

IF customer is Gold
AND order value > $100
THEN provide free delivery.

AI agents introduce another requirement.

They need to understand not only:

What exists?

but also:

• What can I investigate?
• What evidence do I need?
• What question should I ask next?
• Which hypothesis is worth testing?
• When do I have enough evidence to make a claim?

That is a fundamentally different interaction with enterprise data.

05Context is necessary. It is not sufficient.

This does not make semantic or context layers less important.

Quite the opposite.

An investigating agent needs them even more.

An agent cannot reliably investigate an enterprise if it does not understand:

• what its entities represent
• how metrics are defined
• which data is authoritative
• how entities relate to one another
• what business rules apply
• what actions are permitted
• how results should be interpreted

Without this foundation, an agent may generate plausible but meaningless analysis.

But context answers:

“What does this data mean?”

Investigation requires another capability:

“Given what I know, what should I look at next?”

That is the missing layer.

06The analyst’s real advantage is not SQL

It is tempting to think that AI will replace analysts because it can write SQL.

But SQL has never been the hardest part of analytical work.

The difficult part is knowing which analysis to run.

An experienced analyst sees a result and immediately starts forming hypotheses.

Revenue is down.

Maybe it is a pricing issue.

Maybe customer mix changed.

Maybe one channel deteriorated.

Maybe a major customer cohort stopped purchasing.

Maybe nothing actually changed — the reporting logic did.

The analyst tests these possibilities, rejects some, follows others, and keeps narrowing the search.

This is why two analysts can have access to the same data and produce very different insights.

The scarce capability is not access to data.

It is the ability to know where to look next.

07AI agents change the economics of investigation

Humans are inherently constrained in how many questions they can investigate.

An analyst may have the time to investigate five, ten, perhaps twenty hypotheses around an important business problem.

But enterprise data contains millions of possible combinations:

Customers × Products × Channels × Geographies × Time × Behaviors × Transactions.

Most of those combinations will never be examined.

Not because they are irrelevant.

Simply because nobody had the time to ask.

Agents change this constraint.

An agent can take an initial business objective, generate candidate questions, query the relevant data, test hypotheses, discard weak explanations, pursue promising ones, and verify its findings against additional evidence.

The objective is no longer:

“Answer my question.”

It becomes:

“Investigate this business problem and tell me what matters.”

That is a much more powerful paradigm.

08From answering to discovering

This suggests a useful way to think about the evolution of enterprise intelligence:

Data → Semantics → Context → Questions → Investigation → Insight → Action

Each stage solves a different problem.

Data gives machines facts.

Semantics gives those facts meaning.

Context connects that meaning to the enterprise.

But questions and investigation turn that foundation into discovery.

And ultimately, insights become valuable when they change a decision or create an opportunity.

The next generation of enterprise AI therefore shouldn’t be judged only by how accurately it answers questions.

It should also be judged by:

What did it discover that nobody asked it to look for?

09The enterprise needs more than an AI that knows

For decades, we have invested in making enterprise data increasingly accessible to humans.

Then we started making it understandable to machines.

Now we are entering a world where machines can actively work with that data.

That changes the objective.

The goal is no longer simply to build AI that knows what the enterprise knows.

It is to build AI that can continuously investigate what the enterprise has not yet understood.

Because the biggest insights in an enterprise are often not hiding behind difficult queries.

They are hiding behind questions nobody had the time to ask.

AI doesn’t just need context. It needs curiosity.