Digital Labor for Enterprise Automation

Digital Labor for Enterprise Automation

Building a unified Digital Labor platform with conversation, orchestration, and training capabilities.


This initiative is building a comprehensive Digital Labor platform that integrates conversational interfaces, efficient orchestration, and robust training modules. The platform utilizes advanced AI techniques, including foundation models, to significantly enhance the productivity of knowledge workers. It does this by facilitating dynamic, easy-to-use, swift, and dependable self-service automation, eliminating the need for technical expertise.

Digital Labor effectively coordinates a vast array of pre-designed modular skills, each an essential component of automation, defined via open API specifications. This setup empowers users to articulate their business objectives, develop and integrate new modular automation skills on the fly, guided by intuitive conversational interactions.

Mapping & Sequencing

Our platform assembles a large catalog of reusable skills — a skill being an enriched API that represents application capabilities. These APIs are stitched together to achieve a specific business goal. We experiment with AI Planning and FMs for generating executable sequences of APIs that accomplish the specified business goals.

Our state-of-the-art AI planner is capable of efficiently producing many valid sequences by searching through large search spaces in seconds. These search spaces are defined by a planning model, which is in turn acquired by establishing better connections between the APIs and their input and output signatures.

To enable better connections between the skills, it is required, where possible, that output fields be mapped to input fields, and those missing surfaced through the conversation to the business user when executing these composite skills. AI-driven semantic mapping helps by leveraging the API attributes and their documentation to propose candidate mappings to be included in the composition.

Maps also carry valuable information that can be leveraged by the AI Planner to prioritize some choices over others based on certain confidence conditions. As we introduce a human in the loop we can validate and override some of the choices, helping to learn preferences that relate to specific business goals.


FMs are being increasingly used to reason about complex goals and orchestrate and sequence a set of pluggable tools or APIs to accomplish those goals. This is especially relevant in enterprises, where such task-oriented digital systems could enhance productivity of users who interact with multiple applications to perform repetitive tasks.

The GenSeq (Grounded Generative Sequencing) framework offers a declarative pipeline specification to enable quick experimentation of the solution space, such as plugging in different FMs, applying constrained decoding, and choosing among outputs from variations in hyper-parameters. It also incorporates policy rules to ensure the generated API sequences adhere to corporate guidelines and regulations.

Helping LLMs Teach Themselves APIs

Many APIs come with specifications, but unfortunately, most of these specifications are not good enough to be reliably invoked by FMs. With Project OcAPI, we explore how such specifications can be enriched with synthetic data generation. In addition, we explore advances in low-code techniques, as useful API specifications must be co-created with humans.