AI in enterprise software

Enterprise software companies in this catalog build AI into platforms that already hold a company's records and workflows. Salesforce describes Agentforce as an AI agent platform whose agents use large language models and draw on business data such as Salesforce CRM data and data from Data 360. Glean's terms describe a platform that connects to the business applications and data sources a customer selects while keeping each application's identity, access controls, and sharing settings. ServiceNow says its AI Platform integrates with any cloud, model, and data source, and that its AI Control Tower is meant to discover, observe, govern, and secure AI across a company.123

Data platforms run models next to the data they store. Databricks' Foundation Model APIs let customers query open models from serving endpoints, and Databricks says it hosts those models itself. Snowflake says its Cortex AI Functions run models from several developers, including OpenAI, Anthropic, and Meta, and that the models it provides access to are deployed within the Snowflake service perimeter.46

The model behind an enterprise product often comes from another company, with its own terms. Glean's terms name OpenAI, Anthropic, and Google as examples of supported model providers and say that each model is governed by its own terms as well as Glean's agreement. Writer makes its own Palmyra X6 the default model across its platform, and administrators can also enable models from other providers or bring their own through AWS Bedrock, Microsoft Azure, and NVIDIA NIM.27

Some of these companies also publish models under terms separate from their hosted products. IBM says all Granite models are released under Apache 2.0. ServiceNow's Apriel-1.6-15B-Thinker card gives an MIT license and says the Apriel models are not intended for safety-critical uses without human oversight. Snowflake released base and instruct checkpoints of Arctic under Apache 2.0 in April 2024. Writer's Palmyra-mini card says it was fine-tuned from Qwen2.5-1.5B, published by the Qwen Team; a U.S. fine-tune does not make its base model American, and the catalog records that relationship as provenance.8910111213

Hosted models in these platforms are retired on the provider's schedule. Databricks says a deprecated model is given a retirement date 30 or 90 days ahead, and that once a model is retired, any workload using it no longer works. Its retirement table lists dates for Databricks' own DBRX models that have now passed.5