The misconception
“Agents” sound autonomous.
Something that understands your business.
Something that figures things out.
The assumption:
If we create a data agent, it will make our data usable.
That’s not what actually happens.
What a data agent actually is
A data agent is a structured layer between:
- your data
- and the interface querying it
It defines how that data should be interpreted.
A data agent is not intelligence.
It is structured context applied to data.
In practical terms, an agent includes:
- Data sources
The tables, views, and functions it can access - Context definition
How the system should interpret business concepts
(e.g., what “revenue” or “conversion” actually means) - Instructions
How queries should behave
(defaults, filters, grouping logic) - Metadata
Naming, definitions, and relationships between fields - Example queries (verified logic)
Predefined queries that guide correct outputs and enforce business rules
These are not optional enhancements. They are the mechanisms that prevent AI from guessing.
This allows systems like conversational analytics to operate without exposing underlying complexity.
But the defining point is this:
A data agent does not create meaning.
It applies structure to meaning that must already exist.
Why agents exist
Without agents, AI systems operate without context.
They:
- infer relationships
- guess definitions
- construct queries dynamically
This leads to:
- incorrect joins
- inconsistent outputs
- unreliable answers
Agents introduce constraints.
They define:
- what data is used
- how it is interpreted
- how outputs are structured
Without defined context, AI defaults to interpretation.
And interpretation is where inconsistency begins.
What people expect vs what actually happens
Expectation
- agents understand business logic
- agents fix messy data
- agents improve accuracy automatically
Reality
Agents only operate on what they’re given.
If your system is well-structured:
- outputs become more consistent
- queries behave predictably
- logic is applied uniformly
If it isn’t:
- inconsistencies are preserved
- incorrect logic is reinforced
- outputs remain unreliable
How agents map to your system
A data agent is not a standalone capability.
It reflects your system.
| Agent Component | System Layer |
|---|---|
| Data sources | Memory layer |
| Instructions | Processing + governance |
| Metadata | Semantic layer |
| Queries | Data modeling |
This is why agents are not a solution.
They expose how well your system is defined.
What agents actually depend on
Agents don’t require structure.
Your system does.
And agents make that dependency visible.
1. Modeled data
Agents rely on predictable tables.
If your data is still raw, the agent is forced to interpret it.
That interpretation is inconsistent.
This is where structure is created in Data modeling.
2. Defined meaning
Agents rely on definitions—not understanding.
If naming and relationships are unclear, outputs will vary.
Meaning must exist before interpretation.
This is enforced in the Semantic layer.
3. Stable system context
Agents don’t store or reconcile data.
They query it.
If your system is inconsistent, the agent reflects that inconsistency.
Where agents fail
Agents don’t fail independently.
They fail with the system.
1. Reinforcing incorrect logic
If a definition is wrong, the agent applies it consistently.
This creates the illusion of accuracy.
2. Masking structural issues
Outputs appear clean—even when underlying data is not.
This hides problems instead of resolving them.
3. Overconfidence in outputs
Structured responses feel trustworthy.
Even when they aren’t.
If this is happening:
AI doesn’t fix your data. It exposes it.
For a deeper breakdown, see Why AI analytics fails.
Where this fits in your system
Data agents sit between:
- structured data
- and the interface querying it
They are part of the execution layer.
But they do not:
- define your data
- create meaning
- resolve inconsistencies
They depend on:
- data modeling
- semantic definition
- system structure
If that system isn’t stable:
the agent cannot stabilize it
What this enables (when it works)
When the system is structured, agents provide control.
Not intelligence—control.
- consistent query behavior
- enforced logic
- predictable outputs
This allows you to define:
how your data is interpreted before it is queried
Connection to conversational analytics
Agents make conversational analytics usable.
Without them:
- queries are inferred
- results vary
With them:
- queries are guided
- outputs stabilize
But the dependency remains:
agents don’t fix the system
they depend on it
For how this appears at the interface, see Conversational analytics.
What to do next
If outputs are inconsistent, the issue isn’t the agent.
It’s the structure it depends on.
Agents rely on your system
See AI-ready data
Evaluate your system
See Evaluate
Final principle
A data agent doesn’t understand your data.
It enforces how your system defines it.
And if that system isn’t structured:
the outputs will still be consistent—just not correct.
Doug McCaffrey
Designs and maintains analytics systems that remain reliable over time.
Explore how this connects across your data estate: