Salesforce subsidiary Informatica has compressed data integration pipeline development from days or weeks into minutes through Copilot, a generative AI assistant that has produced roughly 10,000 customer-built pipelines and more than 25,000 expression-generation requests since its May 2025 mapping-generation launch. Adoption metrics signal daily workflow use rather than experimental curiosity, with thousands of customers returning to the feature repeatedly.
"Even straightforward pipelines could take days or weeks because users had to understand numerous transformations and configuration options. The complexity itself had become the bottleneck. Customers weren't struggling with integration capabilities. They were struggling with complexity," Nancy Chen, Vice President of Engineering at Informatica, told the Salesforce Engineering Energizers Q&A series.
Traditional Informatica mapping required engineers to inspect schemas across systems like Snowflake and Salesforce, manually define field mappings, configure transformations, and repeatedly validate outputs. As transformation logic grew, pipelines that should have taken hours stretched into days or weeks. Copilot flips the interaction model: users describe business intent conversationally and the system generates the pipeline, shifting focus from implementation details to outcomes, while complex cases such as Snowflake-to-Salesforce mapping — where schema, metadata, field relationships, and transformation logic all interact — become achievable without hand-coding each step.
The Copilot program began with a custom fine-tuned model, because early OpenAI APIs were not mature enough for the team's accuracy targets. As foundation models improved, the team migrated to OpenAI's hosted models, trading direct control over model behavior for higher accuracy and faster iteration. Mapping generation launched in May 2025, followed by mapping augmentation (inserting transformations into existing pipelines) which within two months of launch had already been used on more than 900 existing pipelines.
Copilot generates pipeline definitions from natural-language prompts, grounds them in metadata discovery and schema mapping for both source and target systems, and applies prompt engineering plus validation to filter low-confidence outputs. Because AI outputs are probabilistic — small prompt changes can yield different responses — the team invested heavily in broader test coverage than is typical for deterministic software, ensuring improvements in one workflow don't introduce regressions elsewhere. The architecture deliberately embraces external model evolution: instead of fighting upstream changes, the platform's testing, validation, and release practices continuously adapt as OpenAI models improve.
Roughly 10,000 customer pipelines have been generated via conversational prompts, with more than 25,000 expression-generation requests submitted and about 60% of generated expressions accepted without any manual edit. Mapping augmentation has been applied to over 900 existing pipelines, with approximately 80% of generated transformations accepted on first pass. Thousands of customers have adopted and returned to the experience, indicating the tool has crossed from novelty into daily workflow.
The team is shifting further toward prompt engineering, grounding, and validation techniques as foundation models continue to evolve. Going forward the focus is leveraging rapidly improving large language models rather than maintaining custom models, while continuing to expand coverage of transformation patterns and data quality handling — including varying data quality, formatting differences, missing values, and data cleansing requirements — across more enterprise data sources.