In most data integration tools, the pipeline is the unit of development. You design, configure, and maintain them one at a time as independent objects. That works until you’re facing hundreds of tables across a migration, consolidation, or warehouse build, and the volume makes pipeline-level development unsustainable.
For these scenarios, Centerprise operates at a different level of abstraction. Models aren’t documentation artifacts that sit alongside your pipelines. They are executable metadata: the source of truth that drives pipeline generation, mapping logic, load sequencing, and schema deployment. Build the model right, and the pipelines follow.
Centerprise is also a full-featured pipeline builder. You can design individual pipelines visually or through the agentic AI for any integration use case. But when you’re dealing with dozens or hundreds of tables across a migration, consolidation, warehouse build, or synchronization, the model becomes the more efficient abstraction. That’s what this blog focuses on.
With the addition of Centerprise’s agentic AI stack, every step of that model-driven process, from discovery to deployment, can now be driven by conversation rather than clicks. Here’s how that works end to end.
Step 1: Reverse-Engineer Your Source Into a Model
Every pipeline project starts with the source. In Centerprise, the first step is connecting to your source database and reverse-engineering it into a visual data model. The platform pulls in tables, columns, data types, keys, and relationships, giving you a complete structural representation of the source system in seconds.
When the source schema is clean and well-documented, this is straightforward. You get a model, you understand the structure, and you move on to designing the target.
But source systems aren’t always clean. Foreign keys are sometimes implied but not enforced. Primary keys are missing. Relationships exist in application logic but not in the schema. Documentation is outdated or nonexistent. This is especially common with legacy systems and inherited databases.
This is where the agentic AI becomes valuable. The agent can use Centerprise’s profiling and discovery capabilities as tools to investigate an unclear schema on your behalf. It profiles the data, identifies candidates for missing keys and relationships, and surfaces what it finds. You review the results, confirm or adjust, and end up with a source model that accurately reflects the real structure of the database, not just what the schema definition says. From there, everything downstream, modeling, mapping, pipeline generation, starts on solid ground.
Step 2: Modeling - Build Your Target Structure Through Conversation
With the source understood, you need a target model. This is the blueprint that defines where data will land, whether that’s a Snowflake warehouse, a PostgreSQL database, or a dimensional star schema for analytics.
Centerprise’s data modeling module lets you create target models from scratch or by building on existing structures. You can forward-engineer those models into provider-specific DDL for SQL Server, Snowflake, PostgreSQL, Oracle, and others. For warehousing scenarios, the tool fully supports Kimball-style dimensional modeling and Data Vault 2.0.
The agentic AI makes this dramatically faster. Instead of dragging entities onto a canvas and configuring properties one by one, you can instruct the agent in plain language:
“Convert this OLTP structure into a star schema with Customer, Product, and Date dimensions.”
“Merge the Address and Location entities into a single table.”
The agent doesn’t just interpret these instructions. It executes them. Using built-in MCP-based tools, it directly creates entities, defines relationships, assigns keys, and modifies the model in real time. You see the changes immediately in the visual designer. This isn’t a chatbot generating suggestions you then implement manually. It’s an agent performing operations on your behalf.
Things get more complex when multiple source systems are involved. Different domains often have overlapping structures that aren’t immediately obvious. Your Sales system has a Customers table, your Marketing system has an Accounts table, and they represent the same business entity with different names. Without resolving these overlaps upfront, you end up with redundant tables in your target and duplicated pipelines downstream.
Centerprise’s model union feature handles this directly. It combines two or more models into a single unified structure, using AI to detect both exact and semantic overlaps across entities and attributes. The result is a clean, deduplicated target model ready for pipeline generation.
Step 3: Mapping - AI-Powered Semantic Matching at Scale
With source and target models in place, the next challenge is mapping: which source fields align to which target fields, and what transformations are needed in between? For a handful of tables, this is manageable. For hundreds, it’s a project in itself.
Centerprise’s mapping module tackles this at scale. It takes the source and target models and uses a combination of Astera’s internal mapping engine and AI-based semantic analysis to match entities and attributes automatically. The AI understands context: it recognizes that “ClientID” in one system maps to “CustomerNumber” in another, or that “cust_fname” and “cust_lname” should feed a concatenated “FullName” in the target. The agent also suggests calculated expressions, like deriving TotalAmount from UnitPrice and Quantity, or concatenating address components into a single field.
Engineers review, adjust, or override through conversation, reducing what would normally be weeks of field-level mapping work to a short review cycle.
Step 4: Bulk Pipeline Generation - From Mappings to Executable Flows
This is where the model-driven approach pays off decisively. With mappings confirmed, Centerprise auto-generates pipelines for every source-target table pair. Each pipeline comes pre-wired with the transformations, filters, lookups, and expressions defined during mapping. No one is building these one at a time.
The platform handles the operational complexity automatically. Table load order is determined based on referential integrity constraints, so parent tables load before children and foreign key relationships are preserved throughout the process. For incremental loads, Change Data Capture (CDC) is configured to track audit fields and move only new or updated records at each execution, keeping system load minimal and sync cycles efficient.
The behavior adapts to the use case:
For migrations and consolidations, keys are preserved or regenerated based on preference. Source system keys identify which system each record originates from, which is critical when consolidating similar datasets from multiple systems (e.g., Clients from System A, Customers from System B).
For warehousing, pipelines execute in ELT mode with auto-generated staging layers, automatic SCD handling for dimensions, and dimension lookups for fact tables. The model carries all the logic and the pipeline generator acts on it.
Step 5: Refinement - Adjust Models and Mappings Through Conversation
Models and mappings evolve as business rules get clarified, new source tables come in, or structural changes surface after reviewing the generated pipelines. With the agentic AI, these adjustments stay at the design layer:
“Add a Status column to the Customer dimension and set it as an SCD Type 1 attribute.”
“Split the Address entity into separate Billing and Shipping tables in the target model.”
“Update the TotalAmount mapping to use a tax-inclusive calculation.”
Because pipelines are generated from the models and mappings, they regenerate to reflect the updates. The engineer works at the design layer and the pipelines stay aligned automatically.
Step 6: Orchestration and Scheduling - Make It Self-Sustaining
With pipelines built and refined, the final step is orchestration. Centerprise’s workflow engine lets you execute pipelines in series or parallel, add conditional logic, trigger email notifications on success or failure, run custom SQL scripts, and chain multiple workflows together.
Workflows can be scheduled for time-based or event-based execution, monitored from a unified window, and managed with built-in alerting. The result is a self-sustaining data operation. Models define the structure, pipelines handle the movement and transformation, and workflows automate the execution. What started as a conversation with an AI agent becomes a production system that runs on its own.
Ready to stop building pipelines one at a time? Request a demo and see the model-driven approach in action.