Vendors are usually forced to work with decrepit vendor management systems (VMSs) to access their own transaction data. "Decrepit" because the only acceptable mode of data retrieval is manual, i.e., a human operator must log in, click around, and download. No API, no MCP server. The large corporate customers running these VMSs are effectively saying to their vendors: Efficiency for me but not for thee; you must pay humans to do this work.
It makes sense due to the incentive structure in place. Vendors are not customers and the VMSs primarily exist to prevent these vendors from emailing or phoning the corporate billing department, pestering for transaction data. Keeping vendors at bay with a minimal web interface keeps corporate costs down. Again, the owner of the VMS is allowed to pursue efficiency, but the vendors themselves, how dare they try to run cost-efficient operations?
Before the advent of generative AI, the best approach for vendors seeking cost-efficiency in this hostile environment was to carefully maintain Robotic Process Automation (RPA) scripts, such as those that enabled UiPath to become a large, publicly traded company. However, RPA doesn't work very well; it's too fragile and costly. These facts are arguably reflected in UiPath's stock performance over the past few years and their pivot today toward Agentic AI, similar to many "formerly" SaaS companies.
Now, however, with the advent of generative AI, when a documented API is not available and a shadow API is not discoverable, instead of relying on fragile and expensive RPA maintenance, vendors can ask an AI agent to write the initial RPA script, but also be ready to rewrite it dynamically, any time it breaks. This "tools in a loop" capability, which enables scripts to bring together several applications to continue pursuing the goal, introduces a new level of durability to the core RPA concept. As VMSs update their user interfaces, these programs update themselves to account for the change and continue to function.
Are such programs interacting with an "API"? No, but in terms of how these self-healing programs can be applied in production, they do offer vendors a durable "application programming interface" lowercase, i.e., a programmatic way to reliably interface with remote data. Check out our CEO Mauricio Vergara discussing some of this in his recent interview with Unite.AI.