How to connect AI to SAP and SQL without migrating your data

Topic

Enterprise Intelligence

Kind

Perspective

Author

Dmitry Ermakov

Read time

4 min read

/

/

How to connect AI to SAP and SQL without migrating your data

/

/

How to connect AI to SAP and SQL without migrating your data

Most enterprises do not have a data problem.

They have an access problem.

The information employees need already exists. It sits inside SAP, SQL databases, CRM platforms, document repositories, ERP systems and dozens of specialist applications.

Getting an answer often requires knowing which system contains the information, understanding its structure, having the right permissions and navigating an interface designed around the system rather than the question.

Generative AI creates another possibility.

Instead of moving enterprise data into a new AI platform, we can bring intelligence to the systems where the data already lives.

The wrong starting point: move everything

A common enterprise AI architecture starts by centralising information.

Extract data.

Transform it.

Copy it into a data warehouse, vector database or new knowledge environment.

Synchronise it.

Then place an AI interface on top.

There are good reasons to create centralised data platforms. But doing so should not automatically become a prerequisite for enterprise AI.

Migration introduces duplication, synchronisation requirements, additional security boundaries and another copy of potentially sensitive information.

For many use cases, the AI does not need to own the data.

It needs controlled access to it.

Query the source

Imagine an employee asking:

Show me German customers whose revenue declined by more than 20% during the last two quarters.

The AI does not need every customer record copied into its context.

It needs to understand the request.

From there, an orchestration layer can determine which system contains the information, translate the intent into a controlled operation, execute that operation against the authorised source and return the result.

The architecture becomes:

User โ†’ intent โ†’ permissions โ†’ tool โ†’ source system โ†’ result โ†’ AI โ†’ interface

The enterprise system remains the system of record.

AI becomes the interaction and intelligence layer.

Natural language should not mean unrestricted SQL

Giving an LLM unrestricted database access is a bad shortcut.

A safer architecture separates interpretation from execution.

The model can determine that a user wants customer revenue data.

It should not automatically receive arbitrary database privileges.

Instead, the application can expose controlled tools such as:

  • get_customer_revenue()
  • compare_periods()
  • get_open_orders()
  • find_account_owner()

Each tool has defined parameters, permissions and outputs.

The AI selects the appropriate capability.

Normal software controls what actually happens.

Existing permissions should still matter

Natural language should never become a way around enterprise access controls.

If an employee cannot see payroll information in the underlying system, asking an AI assistant should not suddenly make that information available.

Identity therefore needs to travel through the complete execution chain.

The system should understand:

Who is asking?

Which organisation do they belong to?

What role do they have?

Which information can they access?

Which actions can they perform?

The answer can then be filtered before sensitive information reaches the model or interface.

SAP becomes a capability, not an interface

This becomes particularly interesting with large enterprise platforms.

SAP contains enormous amounts of operational information, but accessing it often requires knowledge of specific transactions, reports and data structures.

AI can create a different interaction model.

An employee might ask:

Which customers have overdue invoices above โ‚ฌ50,000?

Or:

Which products had the largest margin decline this quarter?

Or:

Prepare an overview of our ten biggest supply delays and identify the responsible suppliers.

The AI interprets the request.

SAP remains the source of truth.

The orchestration layer determines which authorised operations are required.

The employee receives an answer rather than navigating the underlying system manually.

SQL works the same way

SQL databases are especially suited to controlled natural-language interaction because the information is already structured.

But the safest pattern is rarely:

prompt โ†’ LLM-generated SQL โ†’ production database

A stronger pattern introduces validation and control:

prompt โ†’ intent โ†’ query plan โ†’ permissions โ†’ parameterised operation โ†’ database โ†’ validation โ†’ response

This gives engineering teams control over what can be queried and how.

For exploratory analytics, generated SQL can still be useful, but it should execute inside appropriate boundaries.

Zero migration changes the business case

Avoiding unnecessary migration can reduce one of the biggest barriers to enterprise AI adoption.

Instead of launching a multi-year programme to reorganise every system before AI becomes useful, organisations can start with selected systems and workflows.

Connect one database.

Expose five useful capabilities.

Give one team access.

Measure usage.

Add another system.

The architecture can grow incrementally.

This is how we think about ONE

ONE is being designed around a simple idea:

Your company should be queryable.

Not because every piece of enterprise information has been moved into another platform.

Because the intelligence layer understands how to interact with the systems the organisation already uses.

SAP.

SQL.

CRM.

Documents.

Internal APIs.

Operational software.

The interface can become much simpler while the systems underneath remain intact.

The enterprise interface is changing

For decades, enterprise software forced people to learn how databases and processes were structured.

Menus reflected tables.

Screens reflected workflows.

Employees learned where information lived.

AI reverses that relationship.

The employee expresses intent.

The system determines how to fulfil it.

That does not make SAP, SQL or ERP disappear.

It makes their interfaces less important.

And that may become one of the largest changes generative AI brings to enterprise software.

Avatar-1
Avatar-2
Avatar-4
Avatar-4
Avatar-3

Fresh insights, straight to your inbox.

Engineering breakdowns, sovereign-AI perspectives and practical guides from the people building enterprise AI. One email when there is something worth your time.