BI, Finance, and IT needed answers from the same legacy ERP. The data lived in SQL Server, but useful analysis depended on knowing the ERP's obscure tables and fields. Business teams often needed IT to write a query before they could explore a question.
I built a Model Context Protocol (MCP) interface so people could use Claude, ChatGPT, and custom agents to work with current ERP data. Making the database understandable was as important as connecting it.
Give the tools enough context
The interface exposed more than 25 tools for schema discovery, SQL inspection, dependency analysis, data-dictionary lookup, and data access.
I wrote scripts to generate a data dictionary from the database, including descriptions of master tables and field definitions. Queries, skills, and operating instructions lived in Git so that knowledge could be maintained and reused.
Keep the production boundary clear
The tools used read-only database access, query validation, and Active Directory authentication. They could inspect and query data, but couldn't create, update, or delete production records.
That boundary let business users explore questions without giving a general-purpose model permission to change the ERP.
What changed for the teams
Finance and BI could do more of their analysis in familiar AI tools. IT and support staff could investigate ERP issues with less dependence on proprietary schema knowledge or a specialist writing each query.
I like finding ways to make an existing system more useful. Here, documenting what the data meant and exposing a bounded set of tools made knowledge that had been difficult to access available to more of the people who needed it.