How CrushBank turned legacy data into AI-ready systems with IBM Bob

Most enterprise AI projects eventually run into the same conundrum: the most useful data is often the hardest to reach.

For many companies, critical business knowledge sits inside legacy applications, on-premises databases and document stores that were never designed for AI. Some systems have no modern APIs. Some run inside a local data closet. Some hold years of operational context, but expose it only through old screens, stored procedures or manual exports.

CrushBank has worked on that problem in IT support for several years. The company helps support teams find answers faster, resolve tickets more consistently and improve customer satisfaction through AI-powered prioritization. Now CrushBank is applying that same data-first approach to broader enterprise use cases with help from IBM Bob™.

CrushBank already knew how to turn messy operational data into useful answers. However, IBM Bob gave the company’s developers the leverage to extend that pattern across legacy systems, data sources and industries. This approach offered an opportunity to expand its addressable market.

IBM Bob helps CrushBank developers analyze legacy systems, inspect schemas and generate ingestion code. They can create tests and build MCP servers that connect enterprise data to models such as Claude, ChatGPT, Llama and other AI systems. In this post, we’ll walk through how CrushBank uses IBM Bob to turn legacy data discovery into a repeatable engineering process for AI-ready data and applications.

Many organizations want to use AI, but they don’t see a clear path to business ROI. Their data might be valuable, but it’s scattered across old line-of-business applications, databases and documents.

A data lakehouse can help once the data gets there. It combines flexible storage with warehouse-style query capabilities, so teams can analyze structured data such as customer records alongside unstructured data such as documents and notes. But the bigger challenge is digging up legacy data, understanding what it means and reshaping it for AI.

That’s where IBM Bob changes the game. To surmount the challenge described earlier, it would take a forward-deployed team consisting of an engineer, account rep and data scientist. They would need weeks to complete a proof of concept (PoC), show value and win the deal.

Now, CrushBank can pair a developer with Bob and create the PoC in an afternoon. With IBM Bob, a full engagement that normally would take a four-person CrushBank team up to 6 weeks can be completed in half the time and one to two developers.

With Bob, a single developer can now traverse databases, reverse-engineer schemas, write ingestion code and build tests from scratch. The developer stays in control, while Bob accelerates the repetitive work required to inspect systems, propose schemas, generate code and document the implementation.

The time it takes to move from PoC to deployment has been reduced to around two weeks. Now, any post-PoC projects are tackled with one developer, a little help from a data specialist and IBM Bob.

The result: users can search across previously disconnected knowledge, ask an assistant about operational data, retrieve relevant documents or run analytics that combine multiple systems. For the customer organization, this method creates a shorter path from legacy data to AI value without replacing every application first.

When looking at the diagram, the process begins with the aggregation of legacy data. The developer and IBM Bob review the data and validate its structure to ensure AI-readiness. This approach allows MCP servers to query the data for use with AI models and applications. Here’s a walkthrough:

In the legacy enterprise data layer, CrushBank starts with applications, databases and documents that might not expose clean integration points. One real-world example involved an electrical engineering firm that used an on-premises SQL Server application to run its business.

The data was useful, but the application had no simple external access path. Before any AI application could use that data, the team had to understand the database, identify the right tables, define an extraction method, normalize the data and protect access.

The next part of the diagram is where IBM Bob comes into the picture. Bob helps developers with system analysis, data discovery, schema creation, ingestion pipeline development, Java-to-Python modernization and test generation.

Bob works inside the development workflow, not as a separate chat layer. Developers can use planning-oriented interactions to ask what the system needs, which questions the business wants answers for and which database structures appear relevant before moving into implementation.

That workflow keeps human judgment in the loop. In the SQL Server example, Bob didn’t directly query production data. Rather, it suggested sample queries and exports. A developer reviewed the requests, ran the queries and returned results.

Human-in-the-loop means a person reviews generated plans, code, queries or access steps before they become part of the working system. For legacy data work, that review protects against more than bugs—it helps prevent unsafe access patterns, incorrect assumptions and exposure of sensitive records.

The third section is AI-ready data. Bob helped generate a database mapping and an Apache Iceberg schema. Bob also helped generate the Python code and Airflow Dags needed to move data from SQL Server into watsonx.data®. An Airflow Dag (directed acyclic graph) is a pipeline definition that controls which data ETL tasks run and in what order.

CrushBank’s data environment runs on IBM Cloud® and includes watsonx.data, Iceberg, the Milvus vector database (to support semantic search and retrieval), Presto and Spark. That architecture lets customers use an enterprise-grade data layer without standing up and operating the full stack themselves.

The fourth section in the diagram is the MCP server layer. CrushBank uses MCP servers to provide secure access, model connectivity and governance. Bob helps build this layer too. Once structured data is available in Iceberg and unstructured content is represented in Milvus, Bob can help plan the MCP server. It can generate access code and create instruction files that tell the AI system how to use the server.

CrushBank’s developers supply the domain judgment, deciding which business questions matter, which data is sensitive and which access patterns are acceptable. Bob handles the repetitive engineering around APIs, server code, instructions and tests.

Once the data is ingested and the MCP servers are ready to go, it is time to create the analytics and AI layer. The fifth stage is model choice. CrushBank’s analytics architecture is not tied to a single model. A customer can connect its data to Claude, ChatGPT, Llama or another model that supports the required integration pattern. That flexibility is important because many customers already have preferred AI assistants in place.

The final stage is the AI application. CrushBank can support search, assistants, knowledge retrieval and analytics. The same foundation can power CrushBank’s own interface, an assistant inside the customer’s preferred model environment or an agentic workflow built with IBM’s Langflow. For customers that need managed orchestration, CrushBank can also use IBM products such as watsonx Orchestrate® where orchestration fits the deployment.

The value of IBM Bob isn’t simply that it writes code. The bigger advantage is that it helps turn messy discovery work into artifacts engineers can review, test and reuse.

The delivery impact can be significant. Work that previously required a larger consulting-style team can shift toward a smaller team: often that’s one developer paired with Bob, supported by focused data expertise. Discovery that once took weeks can be compressed into days for some engagements.

In one analytics-heavy use case, CrushBank combined data from project systems, ERP data and CRM data to build a churn prediction model. The model used historical examples of churn to identify future customers at risk, giving the business a chance to intervene earlier. Bob helped identify patterns, build the ingestion path and create a simple Streamlit interface.

CrushBank achieved this result by training Bob with coding principles, standards and CrushBank’s documented best-practices. That context helps Bob generate work that fits the team’s conventions instead of producing generic code that needs heavy rework. Custom modes and rules within Bob separate planning, coding and review. Developers can reason through architecture, move into implementation and then inspect generated changes without leaving the workflow.

The AI doesn’t push code to production without extensive test harnesses and human peer review. CrushBank doesn’t auto-approve generated code. Code reviews, sensitive data scanning and human review keep Bob’s output inside the software development lifecycle. That discipline is especially important when the work involves legacy data, customer records or poorly documented systems.

CrushBank’s use of IBM Bob shows how AI-assisted development can expedite automation and maintain engineering discipline in the process. The architecture still depends on human review, secure access, data modeling and good deployment practices. Bob helps apply those practices faster across more customer environments.

That repeatability is the scaling advantage. Each customer might bring different legacy applications, databases and documents, but the pattern remains consistent: discover the system, map the schema, build ingestion, structure the data, expose it through governed MCP servers and connect it to the AI experience that fits the users’ needs.

IBM products support each layer of that pattern. IBM Bob accelerates planning, development, test generation and review—watsonx.data® provides the governed data layer. IBM Cloud supplies the infrastructure foundation. watsonx Orchestrate can support managed orchestration where customers need it.

For CrushBank, essentially any company looking to leverage their enterprise data for AI has become a prospect. By building with IBM Bob and IBM’s data and AI products, the company has expanded from an IT support AI business into a broader data enablement business.

Start your free IBM Bob trial